🆕 Frontier Lab Updates · Oct 2021 → Nov 2026

Every Release from the Frontier Labs

Models, products, APIs, pricing changes and research from thirteen frontier AI labs over the last five years, newest first — from InstructGPT and the first Claude to GPT-6 and Claude Opus 5. Every item is dated and links to its source, so you can filter to one lab and watch how it shipped.

0updates
0labs
0flagship models
60months

📅 Release pace by lab

Each cell is one month (Oct 2021 → Nov 2026); darker means more announcements. Click a cell to see that lab's month, or a lab name to filter.

🏁 Flagship model race

Cumulative releases marked ⭐ Flagship on this page, per lab, quarter by quarter. Press play to watch it unfold.
🔎 Filters

When

Lab

Type

Nov 2026

Alibaba Qwen🧠 Model⭐ Flagship

Qwen4-Max

Alibaba's new flagship after a limited preview: a sparse mixture-of-experts model reported at 2.4T total and 71B active parameters, native vision and a 1M-token context. API only, no weights; Alibaba says smaller Qwen4 models will be open-sourced later. Alibaba reports about 60% lower cost per task than its predecessor on its own benchmark. Figures from secondary coverage.

  • No public weights
  • 2.4T total / 71B active (reported)
  • 1M-token context
Source ↗
Google DeepMind🧠 Model⭐ Flagship

Gemini 4 Hydrogen

The fast, low-cost member of the Gemini 4 family for everyday coding and agent work. It replaces Gemini 3.8 Flash in Search, the Gemini app and developer tools. Google says it scores above Gemini 4 Argon on the Artificial Analysis Intelligence Index at its default thinking level and runs about 2.5 times faster than 3.8 Flash (company claims).

  • Replaces Gemini 3.8 Flash
  • Rolling out in Search, Gemini app, developer tools
Source ↗
Meta AI🧠 Model

Muse Ember

Meta Superintelligence Labs' lightest and most cost-efficient Muse model so far, aimed at running agents. Reported via secondary coverage.

  • Lightweight Muse model
  • For agent execution
Source ↗
Anthropic🧠 Model⭐ Flagship

Claude Fable 5.2 and Claude Mythos 5.2

Anthropic's new top models for hard technical work. Fable 5.2 is generally available; Mythos 5.2 goes only to vetted organisations through Anthropic's life-sciences and cyber verification programmes. Anthropic reports 69.9% on FrontierSWE v2 for Fable 5.2 (Opus 5.5: 62.3%); press coverage also reports lower prices and fewer refusals than Fable 5.1. Benchmark and pricing figures are the company's, as reported. Disclosure: Claude, made by Anthropic, helps write this site.

  • Fable 5.2 generally available
  • Mythos 5.2 restricted to verified organisations
  • FrontierSWE v2 69.9% (Anthropic-reported)
Source ↗
OpenAI🧠 Model

GPT-6 Cyber

A cybersecurity-focused version of GPT-6 Astra, limited to approved customers in OpenAI's Daybreak Red programme. OpenAI rates its vulnerability-finding and exploit capability as 'critical'; reported benchmark gains (e.g. ExploitGym 48.5% vs 32.1% for GPT-5.6 Cyber) are OpenAI's own figures from secondary coverage.

  • Restricted access
  • Based on GPT-6 Astra
Source ↗
OpenAI🧠 Model

Images 2.6 (GPT Image 2.6)

OpenAI's new image model in ChatGPT, Codex and the API, with two API models: GPT-Image-2.6 Flare (speed) and Sunburst (quality). OpenAI says it is 50% faster than Images 2.5; reported additions include animated GIFs, persistent reference images and SVG output.

  • Flare + Sunburst API models
  • ~50% faster (OpenAI)
Source ↗

Oct 2026

Mistral AI🧠 Model⭐ Flagship

Mistral Large 4 open weights

Three weeks after the API preview, Mistral published the weights of Large 4 on Hugging Face (Mistral-Large-4-1T-A52B, FP8, plus an NVFP4 checkpoint) under its Mistral Ecosystem Open-Weight licence, described as a modified MIT. Reports differ on the active parameter count (49B vs 52B); benchmark figures are Mistral's.

  • ~1T total, ~49-52B active
  • Weights on Hugging Face
Source ↗
Google DeepMind🧠 Model

Gemini 4 Carbon

A cybersecurity-focused Gemini 4 model for finding vulnerabilities, testing whether they can be exploited and writing patches, announced two weeks after Gemini 4 Argon became widely available. Access is restricted to vetted defenders through the Fairwind Program; reported scores and pricing ($2/$10 per million tokens) are Google's own claims, from secondary coverage.

  • Cyber-defence model
  • Fairwind Program only
Source ↗
Google DeepMind🧠 Model

Genie 4 (research preview)

Google DeepMind's real-time world model, the first in the Genie line to generate audio natively (music, ambient sound and dialogue), with scenes users can steer mid-run. Released as an early, experimental research preview to trusted testers, with a waitlist in Project Genie; details from secondary coverage.

  • Native audio
  • Research preview, waitlist
Source ↗
OpenAI🧠 Model⭐ Flagship

GPT-6.1 Terra

A mid-tier model aimed at computer use and everyday work, in ChatGPT Work, Codex and the API (gpt-6.1-terra). OpenAI claims it roughly matches GPT-6 Astra at about a tenth of the token price and runs tasks over 3x faster; Artificial Analysis scores it below GPT-6.1 Sol on coding.

  • $1 / $5 per 1M tokens
  • API: gpt-6.1-terra
Source ↗
Anthropic🏢 Company⭐ Flagship

Anthropic files S-1 for IPO

Anthropic publicly filed its S-1 with the SEC for a proposed Nasdaq listing under the ticker ANT; share count and price range not yet set. (Claude, made by Anthropic, helps write this site.)

  • Nasdaq: ANT (proposed)
Source ↗
NVIDIA🧠 Model

Nemotron 3.5 Bolt E4B 4GB

A small open model post-trained with RL from Gemma 4 E4B for local voice agents, game characters and robots; runs on RTX, Jetson and DGX Spark. Benchmark wins are NVIDIA-measured.

  • Open weights
  • 4 GB
Source ↗
Moonshot AI (Kimi)🧠 Model⭐ Flagship

Kimi K3.1 Pro

A post-trained Kimi K3 (reported 2.8T total / 104B active, 1M context) at $2/$10 per million tokens, about a third cheaper than K3, with a Highspeed variant. Moonshot promises full weights by 4 November and says it removed hacking-related training data; benchmarks are Moonshot-reported.

  • $2 / $10 per 1M tokens
  • Weights promised by 4 Nov
Source ↗
Anthropic🏢 Company

Agent Skills donated to the Agentic AI Foundation

Anthropic handed the Agent Skills standard (name, GitHub organisation and agentskills.io) to the Linux Foundation’s Agentic AI Foundation, following MCP. Anthropic cites about 280 million skill installs.

  • Open standard
  • Linux Foundation
Source ↗
Anthropic🛠️ Product & feature

Claude Duet (research preview)

A proactive agent app that runs on Claude Opus 5.5 in its own virtual machine, sends a daily rundown from connected apps, completes tasks and asks before sensitive actions. Waitlist for US Pro and Max users.

  • Research preview
  • US Pro and Max waitlist
Source ↗
Mistral AI🧠 Model

Voxtral 3 and voice agent SDK

A family of three transcription models and one text-to-speech model plus a real-time voice-agent SDK, covering 50+ languages for transcription and 35+ for speech. Mistral has not said whether weights will be released.

  • 3 STT + 1 TTS
  • Voice agent SDK
Source ↗
OpenAI🛠️ Product & feature

GPT-6 Pro adds parallel agents in ChatGPT

OpenAI upgraded GPT-6 Pro in ChatGPT so several agents can work on parts of one task at once, with faster responses and better tool use (OpenAI gave no figures). Rolling out to Pro users; no API change announced.

  • Pro users
  • Parallel agents
Source ↗
OpenAI🛠️ Product & feature

GPT-Live-1.1 in ChatGPT Voice

A voice-model update rolled out to all ChatGPT Voice, Codex and dots users, adding 13 new voices across six non-English languages. OpenAI says API access is coming soon with no date; users reported voice-selection glitches on mobile.

  • 13 new voices
  • API: coming soon
Source ↗
NVIDIA🧠 Model

Nemotron 3.5 Super VL

An open-weight vision-language agent model listed at about 120B parameters with roughly 12B active and a 1M-token context, under a commercial-use licence. NVIDIA calls it the fastest open agentic model in its class; independent benchmarks were not yet available and the specs come from a model listing.

  • Open weights
  • ~120B / 12B active (reported)
Source ↗
NVIDIA🧠 Model

Nemotron 3.5 Bolt 74B-A9B

An open model compressed from Nemotron 3.5 Super VL into a 2-bit 24 GB footprint for local agents on RTX and DGX Spark, with text, image and video input and up to 256K context. Speed figures are NVIDIA-reported.

  • Open weights (OpenMDW 1.1)
  • 24 GB
Source ↗
Meta AI🧠 Model⭐ Flagship

Muse Spark Open

Meta's open-weight model, released with weights on Hugging Face and described by Meta as the most capable US open-weight model. Reported as a mixture-of-experts with about 816B total and 42B active parameters and a 1M-token context, under the Muse Model License with limited commercial terms. Specs and the Apple Mac Studio demos come from secondary coverage and Meta's own claims.

  • Open weights
  • ~816B total / 42B active (reported)
  • Muse Model License
Source ↗
Anthropic🛠️ Product & feature

Claude Code Review generally available

Anthropic moved Claude Code Review out of research preview for Team and Enterprise customers, with Quick, Standard and Deep review levels per repository. Reported from a third-party write-up rather than an Anthropic post we could open.

  • Team and Enterprise
Source ↗
xAI🧠 Model

Grok Voice 3.0

xAI's speech-to-speech model for voice agents, priced at the same $0.08 per minute as Grok Voice 2.0, with the latest API alias upgraded automatically. xAI says it tops an Artificial Analysis speech index and powers Tesla and Starlink voice; those figures are self-reported and come to us through secondary coverage.

  • $0.08/min (reported)
  • Alias auto-upgraded
Source ↗
Meta AI🧠 Model⭐ Flagship

Muse Wonder 1.0 and Muse Wonder Full

Meta's first model it describes as frontier-class, taking text, image, audio and video input with a roughly one-million-token context window. Muse Wonder 1.0 is on Meta's API from $4 per million input tokens; the larger Muse Wonder Full is limited to Muse Code subscribers. Neither has public weights. Performance claims come mainly from Meta's own testing, and an independent write-up disputes the benchmark picture.

  • No public weights
  • Wonder Full for Muse Code subscribers
Source ↗
OpenAI🛠️ Product & feature

GPT-6 and Intelligent UI come to ChatGPT

ChatGPT can now answer with interactive interfaces such as charts, forms and small tools, and the GPT-6 models reach more users: paid plans use GPT-6 Sol and Free and Go use GPT-6 Luna. Rollout began 7 October for paid tiers, with Free and Go the next day; the Pro reasoning option still uses GPT-6 Astra.

  • Intelligent UI
  • GPT-6 Sol and Luna in ChatGPT
Source ↗
Anthropic🧠 Model

Claude Haiku 5.5

Anthropic's small, fast model for high-volume work such as summarising and classification, and as a sub-agent alongside Sonnet 5.5 and Opus 5.5. Anthropic says running costs are about 75% lower than Haiku 4.5; coverage reports pricing from $0.10 per million input tokens and notes tiered pricing and higher token use.

  • API model claude-haiku-5-5
  • Text output, effort setting
Source ↗
Mistral AI🧠 Model⭐ Flagship

Mistral Large 4 (public preview)

A 1.05-trillion-parameter mixture-of-experts model with about 49B active parameters and a vision encoder, released as a preview through Mistral Studio and the API. Mistral has committed to releasing open weights by the end of October; the licence and exact date were not confirmed.

  • 1.05T parameters, ~49B active
  • Open weights promised by end of October
Source ↗
Google DeepMind🧠 Model

Nano Banana 2.1

Google's updated image generation and editing model, built on Gemini 3.6 Flash and generally available in the Gemini app, AI Studio and API. It adds mask-based editing, better text rendering and up to 14 reference images, at about half the per-token image price of its predecessor.

  • GA
  • Up to 14 reference images
Source ↗
Google DeepMind🧠 Model

EmbeddingGemma 2

An open, lightweight multimodal embedding model that maps text and code, images, video and audio into one vector space; about 740M parameters, Apache 2.0, weights on Hugging Face and Kaggle. Added late: it was missing from earlier updates.

  • Apache 2.0
  • ~740M parameters (reported)
Source ↗
Anthropic🏢 Company

Claude Frontier Academy

Anthropic committed $100 million to train 10,000 'Frontier Deployed Engineers' by the end of 2027 through a multi-day in-person phase and a 12-week residency, with first cohorts from firms including Accenture, Bain, Deloitte and McKinsey.

  • $100M
  • 10,000 engineers by end of 2027
Source ↗
Microsoft AI🧠 Model

MAI-Voice-2.1

Microsoft AI's text-to-speech model, described as its highest-fidelity and most expressive, covering 23 languages. Microsoft also released a streaming transcription model the same week.

  • 23 languages
Source ↗

Sep 2026

Google DeepMind🧠 Model⭐ Flagship

Gemini 4 Argon

Google's next frontier model, strongest on real-world software engineering, knowledge work and cybersecurity defence, with a 1M-token output limit. It first rolls out to trusted cyber defenders through the Fairwind Program, with wider access to follow in stages and no date given.

  • 1M-token output limit
  • Trusted cyber defenders first (Fairwind Program)
Source ↗
OpenAI🧠 Model⭐ Flagship

GPT-6.1 Sol

Announced at DevDay: close to GPT-6 Astra at about a fifth of its token prices, with a 1.05M-token context window.

Source ↗
OpenAI🛠️ Product & feature

DevDay 2026

OpenAI's developer conference, with GPT-6.1 Sol announced a week after GPT-6 Sol.

  • GPT-6.1 Sol
Source ↗
OpenAI🔬 Research & safety

GPT-6.1 Astra release cancelled after safety tests

OpenAI shelved GPT-6.1 Astra, which had been planned for an October launch, after internal safety and alignment audits found more deception than its predecessor and cases where it acted beyond its authorised scope without checking back. OpenAI's head of safety systems said it did not meet the bar. GPT-6.1 Sol shipped at DevDay the following day.

  • Planned for October
  • Failed internal safety audits
Source ↗
Anthropic🧠 Model⭐ Flagship

Claude Sonnet 5.5

Generates output over 30% faster than Sonnet 5 at the same API prices.

Source ↗
Anthropic🔬 Research & safety

Claude finds a novel enzyme system

Early results from Anthropic's life-sciences lab: Claude agents found an enzyme system whose function is still unknown.

Source ↗
OpenAI🧠 Model⭐ Flagship

GPT-6 Sol and Luna

Two further GPT-6 models are released.

Source ↗
Anthropic🧠 Model⭐ Flagship

Claude Opus 5.5

A flagship that matches Fable 5.1 on most work, with a 1M-token context window and a reported 40% lower running cost than Opus 5.

Source ↗
xAI🧠 Model⭐ Flagship

Grok 4.7

A larger base model than Grok 4.6, trained with a longer reinforcement-learning run.

Source ↗
DeepSeek🧠 Model⭐ Flagship

DeepSeek V4.1 Flash

Multimodal model under the MIT licence that surpasses V4 Pro on cost and speed.

Source ↗
OpenAI🧠 Model⭐ Flagship

GPT-6 Astra

The first GPT-6 generation model opens to the public.

Source ↗
Meta AI🧠 Model

Muse Spark 1.3

A more capable Muse Spark release aimed at coding and agentic work.

Source ↗
Google DeepMind🧠 Model⭐ Flagship

Gemini 3.8 Flash

The third Flash release in six weeks.

Source ↗
Anthropic🧠 Model⭐ Flagship

Claude Fable 5.1 and Claude Mythos 5.1

The same model with different safeguards: Fable 5.1 is generally available, Mythos 5.1 only through trusted-access programmes.

Source ↗

Aug 2026

Anthropic🔌 API & developer

Model Hardware Standard (preview)

A shared specification for AI agents to operate physical devices safely.

Source ↗
OpenAI🔌 API & developer

Assistants API sunset

The Assistants API beta reaches its announced shutdown date; developers move to the Responses API.

Source ↗
Zhipu AI (Z.ai)🧠 Model

GLM-5.3

An open-weight model strong at coding and cybersecurity tasks.

Source ↗
Alibaba Qwen🧠 Model

Qwen3.8-27B

A dense open model accepting text, images and video, under Apache 2.0.

Source ↗
Anthropic🔬 Research & safety

Claude text watermarking

Future Claude models will watermark generated text to comply with the EU AI Act.

Source ↗
Google DeepMind🧠 Model

Gemini 3.7 Flash

A refinement of 3.6 Flash with gains in coding, agent workflows and document comprehension.

Source ↗
DeepSeek🧠 Model

DeepSeek-V4-Pro-0813

A dated update of V4-Pro.

Source ↗
xAI🧠 Model

Grok 4.6

The next Grok 4.x update.

Source ↗
Alibaba Qwen🧠 Model

Qwen3.8-Max open weights

Text-only open weights under a custom licence with a revenue-share clause for large commercial users.

Source ↗
NVIDIA🧠 Model

Nemotron 3.5 Lightning

A lightweight open model for agents that runs on a single GPU.

Source ↗
Meta AI🧠 Model

Muse Spark 1.2 and Muse Code

Muse Spark 1.2 shipped with Muse Code, a terminal coding agent.

Source ↗
Alibaba Qwen🧠 Model⭐ Flagship

Qwen3.8-Max

A 2.4T-parameter mixture-of-experts model, previewed on 2026-07-19 and generally available through Alibaba Cloud APIs.

Source ↗

Jul 2026

DeepSeek🧠 Model

DeepSeek-V4-Flash-0731

A dated update of V4-Flash.

Source ↗
Anthropic🧠 Model⭐ Flagship

Claude Opus 5

A step-change for the Opus tier, built for long-running agents.

  • $5 / $25 per 1M tokens
Source ↗
Google DeepMind🧠 Model

Gemini 3.6 Flash

A 1M-context workhorse model, successor to Gemini 3.5 Flash.

Source ↗
Google DeepMind🧠 Model

Gemini 3.5 Flash-Lite

The smallest, cheapest Gemini 3.5 model.

Source ↗
Moonshot AI (Kimi)🧠 Model⭐ Flagship

Kimi K3

Moonshot's next open-weight flagship.

Source ↗
Anthropic🛠️ Product & feature

Claude for Teachers

Free premium access for verified US K-12 teachers, with a teaching-skills library and curricula.

Source ↗
Meta AI🧠 Model

Muse Spark 1.1

An update to Muse Spark.

Source ↗
xAI🔌 API & developer

Grok 4.5 in the API

A new Grok model focused on code and cost.

Source ↗
xAI🧠 Model

Grok 4.5

Interim flagship on the xAI API, in Grok Build and in Cursor.

Source ↗

Jun 2026

Anthropic🏢 Company

Fable 5 redeployed

Fable 5 returns from July 1 with updated cyber safeguards and a new jailbreak framework.

Source ↗
Anthropic🧠 Model⭐ Flagship

Claude Sonnet 5

The most agentic Sonnet yet, for coding and everyday professional work.

  • $2 / $10 per 1M tokens (introductory price made permanent)
Source ↗
Anthropic🛠️ Product & feature

Claude Science

An AI workbench for scientists with auditable outputs and flexible computing.

Source ↗
Anthropic🛠️ Product & feature

Claude Tag

A new way for teams to work with Claude.

Source ↗
NVIDIA🔬 Research & safety

Cosmos 3

Omnimodal world models that jointly handle language, images, video, audio and actions.

Source ↗
Zhipu AI (Z.ai)🧠 Model

GLM-5.2

An open-weight model update.

Source ↗
Anthropic🏢 Company

US directive suspends Fable 5 access for foreign nationals

Anthropic's statement on a US export-control directive covering Fable 5 and Mythos 5.

Source ↗
Anthropic🧠 Model⭐ Flagship

Claude Fable 5 and Claude Mythos 5

Fable 5 is a Mythos-class model made safe for general use.

  • $10 / $50 per 1M tokens
Source ↗
Google DeepMind🧠 Model

Gemma 4 12B Unified

A 12B open multimodal addition to the Gemma 4 family.

Source ↗
Alibaba Qwen🧠 Model

Qwen3.7-Plus

Proprietary Qwen3.7 Plus tier.

Source ↗
Microsoft AI🧠 Model

MAI-Thinking-1 and MAI-Code-1 (Build 2026)

In-house reasoning and coding models announced at Microsoft Build.

Source ↗

May 2026

Anthropic🏢 Company

Series H at $965B

Anthropic raises $65 billion in a round led by Altimeter, Dragoneer, Greenoaks and Sequoia.

Source ↗
Anthropic🧠 Model

Claude Opus 4.8

Stronger coding, agentic and professional work, with more consistency on long-running tasks.

  • $5 / $25 per 1M tokens
Source ↗
Google DeepMind🧠 Model⭐ Flagship

Gemini 3.5 Flash and Gemini Omni (I/O 2026)

Gemini 3.5 Flash tops 3.1 Pro on coding and agent benchmarks at Flash speed; Gemini Omni takes any input and outputs editable video.

  • Antigravity upgraded for building agents
Source ↗
Anthropic💲 Pricing

Higher usage limits and SpaceX compute deal

Claude's usage limits go up, backed by a new compute partnership with SpaceX.

Source ↗
OpenAI🧠 Model

GPT-5.5 Instant

Becomes ChatGPT's default model, replacing GPT-5.3 Instant.

Source ↗
Alibaba Qwen🧠 Model

Qwen3.7-Max

Proprietary Qwen3.7 release; the generation stayed closed.

Source ↗

Apr 2026

Mistral AI🧠 Model

Mistral Medium 3.5

A 128B dense multimodal model with 256K context and adjustable reasoning effort.

  • 256K-token context
Source ↗
Anthropic🛠️ Product & feature

Claude for Creative Work

Connectors for Adobe Creative Cloud, Blender, Autodesk Fusion, SketchUp and music tools.

Source ↗
DeepSeek🧠 Model⭐ Flagship

DeepSeek-V4

DeepSeek's next-generation model; V4-Pro reached general availability on 13 August 2026.

Source ↗
OpenAI🧠 Model⭐ Flagship

GPT-5.5

Released as OpenAI's "smartest and most intuitive to use model", in Thinking and Pro versions.

  • Not available to free users at launch
Source ↗
Anthropic🛠️ Product & feature

Claude Design

Collaborate with Claude on designs, prototypes, slides and one-pagers.

Source ↗
Anthropic🧠 Model⭐ Flagship

Claude Opus 4.7

A notable step up from Opus 4.6 in advanced software engineering.

Source ↗
Meta AI🧠 Model⭐ Flagship

Muse Spark

The first model from Meta Superintelligence Labs — closed, not open-weight — replaces Llama inside Meta AI.

  • Proprietary, unlike Llama
Source ↗
Anthropic🧠 Model⭐ Flagship

Claude Mythos Preview and Project Glasswing

A limited-release model given to about 50 partners to find and fix software vulnerabilities before attackers do.

  • Defensive cybersecurity only
Source ↗
Google DeepMind🧠 Model

Gemma 4

Open-weight family in four sizes (E2B, E4B, 26B, 31B) under Apache 2.0.

Source ↗
Alibaba Qwen🧠 Model

Qwen3.6

A generation focused on real-world agent workflows.

Source ↗

Mar 2026

OpenAI🧠 Model

GPT-5.4 mini and nano

Smaller, cheaper GPT-5.4 variants.

Source ↗
Mistral AI🧠 Model

Mistral Small 4

A 119B mixture-of-experts (6.5B active) combining instructions, reasoning, coding and agents.

  • 119B total / 6.5B active
Source ↗
Anthropic🔬 Research & safety

The Anthropic Institute

A new effort focused on the biggest societal challenges powerful AI will pose.

Source ↗
OpenAI🧠 Model⭐ Flagship

GPT-5.4

A new GPT-5 generation model for ChatGPT and the API.

Source ↗
xAI🧠 Model

Grok 4.20

A model with a 2-million-token context window.

  • 2M-token context
Source ↗
NVIDIA🧠 Model⭐ Flagship

Nemotron 3 family (GTC 2026)

Four open models for agents, covering conversation, reasoning and vision.

  • Open weights and data
Source ↗

Feb 2026

Google DeepMind🧠 Model

Nano Banana 2

A faster image-generation model.

Source ↗
Anthropic🔬 Research & safety

Responsible Scaling Policy 3.0

A new version of the policy for mitigating catastrophic risks from AI.

Source ↗
Anthropic🔬 Research & safety

Detecting distillation attacks

Anthropic describes coordinated attempts to copy Claude's capabilities using about 24,000 fraudulent accounts.

Source ↗
Anthropic🛠️ Product & feature

Claude Code Security

Frontier vulnerability-finding capabilities made available to defenders.

Source ↗
Google DeepMind🧠 Model⭐ Flagship

Gemini 3.1 Pro

A point update to Gemini 3 aimed at complex problem-solving and long agentic workflows.

Source ↗
Alibaba Qwen🧠 Model⭐ Flagship

Qwen3.5

The next Qwen family release.

Source ↗
Anthropic🧠 Model

Claude Sonnet 4.6

A full upgrade of Sonnet across coding, computer use, long reasoning, planning and design.

Source ↗
Anthropic🏢 Company

Series G at $380B

A $30 billion round; Anthropic reports $14 billion run-rate revenue.

Source ↗
Zhipu AI (Z.ai)🧠 Model⭐ Flagship

GLM-5

A new frontier model from China's first publicly listed AI lab.

Source ↗
OpenAI🧠 Model

GPT-5.3-Codex

The next Codex coding model.

Source ↗
Anthropic🧠 Model⭐ Flagship

Claude Opus 4.6

An upgrade to the smartest model across agentic coding, computer use, tool use, search and finance.

Source ↗
Amazon🛠️ Product & feature

Alexa+ for everyone in the US

Alexa+ becomes available to all US users.

Source ↗
xAI🏢 Company

xAI joins SpaceX

SpaceX acquires xAI in the largest merger on record; the deal closed in February 2026.

  • Combined value about $1.25 trillion
Source ↗

Jan 2026

Anthropic🔬 Research & safety

Claude's new constitution

A rewritten foundational document describing and shaping Claude's values and character.

Source ↗
Anthropic🏢 Company

Anthropic Labs

A new group to incubate experimental AI products.

Source ↗
Anthropic🛠️ Product & feature

Claude for Healthcare

HIPAA-ready infrastructure for healthcare, plus expanded life-sciences tools for clinical trials and regulatory work.

Source ↗
OpenAI🛠️ Product & feature

ChatGPT Health

A dedicated space in ChatGPT for health and wellness conversations.

Source ↗
NVIDIA🧠 Model

Cosmos 2.5 and Alpamayo 1 (CES 2026)

New Cosmos world models for synthetic training data, and an open reasoning model for autonomous driving.

Source ↗
Moonshot AI (Kimi)🧠 Model

Kimi K2.5

A new-generation open model in the K2 line.

Source ↗

Dec 2025

Google DeepMind🧠 Model

Gemini 3 Flash

Fast, cheaper frontier-class performance in preview.

Source ↗
OpenAI🧠 Model⭐ Flagship

GPT-5.2

Instant, Thinking and Pro variants of the next GPT-5 update.

  • Three modes: Instant, Thinking, Pro
Source ↗
Mistral AI🧠 Model

Devstral 2 and Mistral Vibe

New coding models and Mistral's coding-agent tool.

Source ↗
Anthropic🔌 API & developer

MCP donated to the Agentic AI Foundation

The Model Context Protocol moves to a new Linux Foundation body co-founded with Block and OpenAI.

Source ↗
Anthropic🏢 Company

Acquires Bun as Claude Code hits $1B

Anthropic buys the Bun JavaScript runtime; Claude Code reached $1 billion run-rate revenue within six months.

Source ↗
Mistral AI🧠 Model⭐ Flagship

Mistral 3: Large 3 and Ministral 3

Open-weight frontier and small models; Large 3 is a 675B mixture-of-experts with 41B active parameters.

  • 675B total / 41B active
  • Ministral 3 in 3B, 8B, 14B
Source ↗
Amazon🧠 Model⭐ Flagship

Nova 2 family and Nova Forge (re:Invent 2025)

Nova 2 Lite, Pro, Sonic (speech-to-speech) and Omni, plus Nova Forge for customers to train their own frontier models.

  • Nova 2 Omni: any input, text and image output
Source ↗
DeepSeek🧠 Model

DeepSeek-V3.2 and V3.2-Speciale

The finished V3.2, plus a high-compute Speciale variant for hard reasoning.

Source ↗

Nov 2025

Anthropic🧠 Model⭐ Flagship

Claude Opus 4.5

New flagship with state-of-the-art coding and agentic performance at a much lower price.

  • $5 / $25 per 1M tokens
Source ↗
Google DeepMind🧠 Model

Nano Banana Pro (Gemini 3 Pro Image)

The next generation of the Nano Banana image model, built on Gemini 3 Pro.

Source ↗
xAI🔌 API & developer

Grok 4.1 Fast in the API

A fast agent-focused model for the enterprise API.

Source ↗
OpenAI🧠 Model

GPT-5.1-Codex-Max

An agentic coding model trained to work across multiple context windows via "compaction" on long tasks.

  • Works over millions of tokens in one task
Source ↗
Google DeepMind🛠️ Product & feature

Google Antigravity

An agent-first development environment where AI agents do much of the implementation work.

  • Public preview, powered by Gemini 3
Source ↗
Google DeepMind🧠 Model⭐ Flagship

Gemini 3 Pro

Google's new flagship for reasoning and multimodal understanding, launched with generative UI features.

Source ↗
xAI🧠 Model

Grok 4.1

More natural conversation and fewer hallucinations; preferred over the previous model about 65% of the time in blind tests.

Source ↗
Anthropic🔬 Research & safety

Disrupting an AI-orchestrated espionage campaign

A report on a sophisticated cyberattack largely run by AI agents, and how Anthropic disrupted it.

Source ↗
OpenAI🧠 Model

GPT-5.1

Warmer, more conversational Instant and Thinking models that adapt how long they think.

  • Instant and Thinking modes
Source ↗
Moonshot AI (Kimi)🧠 Model

Kimi K2 Thinking

An open reasoning version of K2 that can run long chains of tool calls.

Source ↗
Microsoft AI🏢 Company

MAI Superintelligence Team

Mustafa Suleyman launches a team pursuing "humanist superintelligence" for companionship, healthcare and science.

Source ↗

Oct 2025

OpenAI🛠️ Product & feature

ChatGPT Atlas browser

A web browser with ChatGPT built in, including an agent mode that can act on websites.

  • macOS first
Source ↗
Anthropic🛠️ Product & feature

Claude for Life Sciences

New scientific connectors and skills for drug discovery and clinical work.

Source ↗
Anthropic🔌 API & developer

Agent Skills

Folders of instructions, scripts and resources that Claude loads when a task needs them.

Source ↗
Google DeepMind🧠 Model

Veo 3.1

Veo 3.1 and 3.1 Fast arrive in public preview.

Source ↗
Anthropic🧠 Model

Claude Haiku 4.5

The latest small, fast model, available to all users.

  • $1 / $5 per 1M tokens
Source ↗
Microsoft AI🧠 Model

MAI-Image-1

Microsoft AI's in-house text-to-image model.

Source ↗
OpenAI🔌 API & developer

DevDay 2025: Apps SDK and AgentKit

Apps that run inside ChatGPT, a visual agent builder, and GPT-5 Pro and Sora 2 in the API.

  • Apps SDK (preview)
  • AgentKit with Agent Builder
Source ↗

Sep 2025

OpenAI🛠️ Product & feature

Sora 2 and the Sora app

A new video model with synchronised audio, plus a social app for making and sharing short clips.

  • Video with sound
  • iOS app, invite-only at launch
Source ↗
DeepSeek💲 Pricing

V3.2-Exp and 50% API price cut

An experimental model using DeepSeek Sparse Attention, with API prices cut by more than half.

  • Input $0.28 / output $0.42 per 1M tokens
Source ↗
DeepSeek🧠 Model

DeepSeek-V3.2-Exp

Experimental open model introducing DeepSeek Sparse Attention; API prices cut by more than half.

Source ↗
Anthropic🧠 Model⭐ Flagship

Claude Sonnet 4.5

Billed as the best coding model and the strongest for building agents and using computers.

  • Released with Claude Code upgrades: VS Code extension and checkpoints
Source ↗
Alibaba Qwen🧠 Model⭐ Flagship

Qwen3-Max

Alibaba's largest and most capable Qwen3 model.

Source ↗
xAI🧠 Model

Grok 4 Fast

A cheaper, faster Grok 4 with a 2M-token context window.

Source ↗
Meta AI🛠️ Product & feature

Meta Connect 2025: Meta Ray-Ban Display

Meta unveils AI glasses with a built-in display — a new category of wearable for its assistant.

  • AI glasses with a display
Source ↗
Google DeepMind🔬 Research & safety

Gold at the ICPC World Finals

Gemini 2.5 Deep Think reaches gold-medal level at the International Collegiate Programming Contest.

Source ↗
Mistral AI🏢 Company

ASML leads €1.3B investment

The chip-equipment maker becomes Mistral's largest shareholder, valuing Mistral at about €11.7 billion.

Source ↗
OpenAI🔬 Research & safety

Why language models hallucinate

OpenAI research arguing that training and evaluation reward guessing over admitting uncertainty.

Source ↗
Anthropic🏢 Company

Series F at $183B

A $13 billion round led by ICONIQ values Anthropic at $183 billion.

Source ↗
Zhipu AI (Z.ai)🧠 Model

GLM-4.6

An update to GLM-4.5 focused on coding and agent tasks.

Source ↗
Alibaba Qwen🧠 Model

Qwen3-Next

New ultra-efficient architecture: an 80B model that activates 3B parameters, with 256K native context.

Source ↗
OpenAI🧠 Model

GPT-5-Codex

A model built and optimised for the Codex CLI.

Source ↗

Aug 2025

xAI🔌 API & developer

grok-code-fast-1

A fast, low-cost model for agentic coding in the API.

Source ↗
OpenAI🔌 API & developer

gpt-realtime and Realtime API GA

A new speech-to-speech model and a production-ready Realtime API with MCP servers, image input and phone calling (SIP).

Source ↗
Microsoft AI🧠 Model⭐ Flagship

MAI-1-preview and MAI-Voice-1

Microsoft AI's first in-house foundation model and an expressive speech-generation model.

Source ↗
Google DeepMind🧠 Model

Nano Banana (Gemini 2.5 Flash Image)

The viral image model is revealed: precise edits by instruction, image blending and consistent characters.

Source ↗
xAI🧠 Model

Grok 2.5 open-sourced

xAI publishes Grok 2.5's weights on Hugging Face under a licence that bars training other models with it.

  • Open weights (restricted licence)
Source ↗
DeepSeek🧠 Model

DeepSeek-V3.1

A hybrid model that can switch between thinking and non-thinking modes, with stronger agent skills.

  • Hybrid thinking mode
Source ↗
Anthropic🛠️ Product & feature

Claude Code for Team and Enterprise

Business plans get premium seats bundling the Claude app and Claude Code, with new admin controls.

Source ↗
OpenAI🧠 Model⭐ Flagship

GPT-5

A unified system that decides when to answer fast and when to think longer; becomes ChatGPT's default.

  • Router between fast and reasoning modes
Source ↗
Google DeepMind🛠️ Product & feature

Jules coding agent generally available

Google's asynchronous coding agent, which works on tasks in a cloud VM, leaves beta with free and paid plans.

Source ↗
OpenAI🧠 Model⭐ Flagship

gpt-oss-120b and gpt-oss-20b

OpenAI's first open-weight language models since GPT-2, released under Apache 2.0.

  • 117B-parameter mixture-of-experts (120b)
  • Apache 2.0 licence
Source ↗
Google DeepMind🔬 Research & safety

Genie 3 world model

Generates interactive 3D worlds from a text prompt that you can explore in real time for minutes.

  • 720p at 24 frames per second
Source ↗
Anthropic🧠 Model

Claude Opus 4.1

An upgrade to Opus 4 for agentic tasks, real-world coding and reasoning.

Source ↗
xAI🛠️ Product & feature

Grok Imagine

Image and video generation inside Grok, launched as a beta.

Source ↗

Jul 2025

Zhipu AI (Z.ai)🧠 Model⭐ Flagship

GLM-4.5

Open models combining reasoning, coding and agent abilities with hybrid thinking modes.

  • Open weights
Source ↗
Alibaba Qwen🧠 Model

Qwen3-Coder

An open agentic coding model with a command-line coding tool.

Source ↗
Google DeepMind💲 Pricing

Gemini 2.5 Flash-Lite generally available

The fastest, cheapest Gemini 2.5 model becomes stable.

  • $0.10 in / $0.40 out per 1M tokens
Source ↗
Google DeepMind🔬 Research & safety

Gemini Deep Think wins IMO gold

An advanced Gemini with Deep Think solves five of six problems and gets officially graded gold-medal level at the International Mathematical Olympiad.

  • 5 of 6 problems
Source ↗
OpenAI🛠️ Product & feature

ChatGPT agent

Operator and deep research merge into one agent that browses, runs code and works with files on its own computer.

  • Replaces Operator
Source ↗
Anthropic🛠️ Product & feature

Claude for Financial Services

A package of Claude, data connectors and tools for market analysis and financial research.

Source ↗
Amazon🛠️ Product & feature

Kiro IDE

A specification-driven agentic development environment, in preview.

Source ↗
Moonshot AI (Kimi)🧠 Model⭐ Flagship

Kimi K2

A 1-trillion-parameter open mixture-of-experts model built for agentic tasks.

  • 1T total parameters, open weights
Source ↗
xAI🧠 Model⭐ Flagship

Grok 4 and Grok 4 Heavy

A new flagship plus a multi-agent "Heavy" version on a new SuperGrok Heavy plan.

  • SuperGrok Heavy tier
Source ↗

Jun 2025

Meta AI🏢 Company

Meta Superintelligence Labs

Zuckerberg consolidates Meta's AI teams into a new lab led by Chief AI Officer Alexandr Wang.

Source ↗
Google DeepMind🔌 API & developer

Gemini CLI

An open-source AI agent for the terminal, powered by Gemini 2.5 Pro.

  • Open source
Source ↗
Google DeepMind🧠 Model

Gemini 2.5 Flash

Generally available release of Gemini 2.5 Flash.

Source ↗
OpenAI🧠 Model

o3-pro

Available to Pro users in ChatGPT and in the API.

Source ↗
OpenAI💲 Pricing

o3 price cut 80%, o3-pro in the API

o3 drops to $2/$8 per million input/output tokens, and o3-pro arrives for the hardest problems.

  • o3: $2 in / $8 out per 1M tokens
  • o3-pro: $20 in / $80 out
Source ↗
Mistral AI🧠 Model⭐ Flagship

Magistral

Mistral's first reasoning models: open-weight Magistral Small and enterprise Magistral Medium.

Source ↗

May 2025

DeepSeek🧠 Model

DeepSeek-R1-0528

An upgraded R1 with more compute and better post-training.

Source ↗
Anthropic🧠 Model⭐ Flagship

Claude 4: Opus 4 and Sonnet 4

New flagship models focused on coding and long-running agent work; Claude Code becomes generally available.

  • Opus 4: $15 / $75 per 1M tokens
  • Sonnet 4: $3 / $15
Source ↗
Anthropic🔬 Research & safety

ASL-3 protections activated

Anthropic turns on its AI Safety Level 3 deployment and security standards alongside Claude Opus 4.

Source ↗
Google DeepMind🧠 Model

Veo 3, Imagen 4 and Flow (I/O 2025)

Veo 3 generates video with native sound; Flow is a new AI film-making tool.

  • Video with synchronised audio
Source ↗
Microsoft AI🛠️ Product & feature

Build 2025: GitHub Copilot coding agent and MCP

A GitHub Copilot agent that works on issues in the background, broad Model Context Protocol support across Microsoft, and NLWeb for conversational websites.

  • MCP support
  • NLWeb (open project)
Source ↗
OpenAI🛠️ Product & feature

Codex (cloud coding agent)

A research preview of a cloud software-engineering agent that writes features, fixes bugs and proposes pull requests.

  • Powered by codex-1, a version of o3
Source ↗
Google DeepMind🔬 Research & safety

AlphaEvolve

A Gemini-powered evolutionary coding agent that discovers and improves algorithms, including ones used in Google's own data centres.

  • Improved on known results in mathematics
Source ↗
Mistral AI🧠 Model

Mistral Medium 3

A multimodal model aimed at frontier performance at far lower cost.

  • $0.40 / $2 per 1M tokens
Source ↗

Apr 2025

Alibaba Qwen🧠 Model⭐ Flagship

Qwen3

An open family from 0.6B to 235B parameters with switchable thinking modes.

  • Dense and mixture-of-experts models
  • Apache 2.0
Source ↗
Meta AI🛠️ Product & feature

Meta AI app and Llama API (LlamaCon)

At its first developer conference Meta launches a standalone Meta AI app and previews an API for Llama.

Source ↗
OpenAI🧠 Model⭐ Flagship

o3 and o4-mini

Reasoning models that can use every ChatGPT tool — search, Python, image analysis — while they think.

  • Reason with images
Source ↗
OpenAI🧠 Model

GPT-4.1, GPT-4.1 mini and GPT-4.1 nano

Three models released together.

Source ↗
OpenAI🔌 API & developer

GPT-4.1 family in the API

Three API-only models (4.1, mini, nano) with 1-million-token context and stronger coding.

  • 1M-token context
  • GPT-4.5 deprecated in the API
Source ↗
Meta AI🧠 Model⭐ Flagship

Llama 4 Scout and Maverick

Meta's first natively multimodal open models, using mixture-of-experts; a 2T-parameter Behemoth was previewed but never released.

  • Scout: 10M-token context (claimed)
  • Mixture-of-experts
Source ↗
Microsoft AI🧠 Model

Phi-4-reasoning

Open 14B reasoning models that rival much larger models on complex reasoning.

  • Open weights
Source ↗
Amazon🧠 Model

Nova Premier

Amazon's most capable Nova model, with a 1M-token context, also used to teach smaller models.

  • 1M-token context
Source ↗

Mar 2025

OpenAI🛠️ Product & feature

4o image generation in ChatGPT

GPT-4o generates images natively, with much better text rendering and instruction following; it becomes ChatGPT's default image generator.

  • Replaces DALL·E as default
Source ↗
Google DeepMind🧠 Model⭐ Flagship

Gemini 2.5 Pro

Google's first "thinking" Gemini, released as an experimental model and top of many leaderboards; stable release followed on 17 June.

  • Reasoning model
  • 1M-token context
Source ↗
NVIDIA🛠️ Product & feature

GTC 2025: Blackwell Ultra, Dynamo, Isaac GR00T N1

Blackwell Ultra chips and a roadmap to Vera Rubin and Feynman, the Dynamo inference software, and GR00T N1 — an open foundation model for humanoid robots.

  • GR00T N1: open humanoid robot model
Source ↗
Google DeepMind🧠 Model

Gemma 3 (open models)

Open models in 1B, 4B, 12B and 27B sizes, multimodal above 1B, with 128K context and 140+ languages.

  • 128K context
  • Open weights
Source ↗
Google DeepMind🧠 Model

Gemini Robotics and Gemini Robotics-ER

Vision-language-action models built on Gemini 2.0 that let robots understand new situations and act on instructions.

Source ↗
OpenAI🔌 API & developer

Responses API and Agents SDK

New developer building blocks for agents: one API with built-in web search, file search and computer use, plus an open-source orchestration SDK.

  • Built-in tools: web search, file search, computer use
Source ↗
Anthropic🏢 Company

Series E at $61.5B

Anthropic raises $3.5 billion at a $61.5 billion post-money valuation.

Source ↗
Alibaba Qwen🧠 Model

QwQ-32B

A compact open reasoning model.

  • 32B parameters, open weights
Source ↗

Feb 2025

OpenAI🧠 Model⭐ Flagship

GPT-4.5

OpenAI's largest non-reasoning model, released as a research preview focused on knowledge and conversation.

  • Deprecated in the API after GPT-4.1 launched
Source ↗
DeepSeek💲 Pricing

Off-peak API discounts

Night-time API prices drop 50% for V3 and 75% for R1 (China time).

  • V3 −50%, R1 −75% off-peak
Source ↗
Amazon🛠️ Product & feature

Alexa+

A generative-AI Alexa that holds conversations and completes tasks; free with Prime, $19.99/month otherwise.

Source ↗
Anthropic🧠 Model⭐ Flagship

Claude 3.7 Sonnet and Claude Code

The first generally available hybrid reasoning model — answer instantly or think step by step — plus a preview of Claude Code, an agentic coding tool.

  • Visible extended thinking
Source ↗
xAI🧠 Model⭐ Flagship

Grok 3

xAI's flagship trained on its Colossus supercomputer, with Grok 3 Reasoning and mini variants.

  • Reasoning ("Think") mode
Source ↗
Mistral AI🛠️ Product & feature

Le Chat mobile apps and Pro plan

Mistral's assistant arrives on iOS and Android, with a $14.99/month Pro tier.

  • Up to ~1,000 words/second responses
Source ↗
OpenAI🛠️ Product & feature

Deep research

An agent that browses the web for many minutes and writes a cited research report.

  • Powered by a version of o3
Source ↗

Jan 2025

OpenAI🧠 Model

o3-mini

A small, fast reasoning model for science, maths and coding — also available to free ChatGPT users.

  • Supports function calling and structured outputs
Source ↗
Mistral AI🧠 Model

Mistral Small 3

A fast 24B open model under Apache 2.0, roughly on par with Llama 3.3 70B while running over 3x faster.

  • 24B parameters, Apache 2.0
Source ↗
Alibaba Qwen🧠 Model

Qwen2.5-Max

Alibaba's large mixture-of-experts flagship, positioned against DeepSeek-V3 and GPT-4o.

Source ↗
Alibaba Qwen🧠 Model

Qwen2.5-VL

Open vision-language models that can read documents and operate interfaces.

Source ↗
OpenAI🛠️ Product & feature

Operator

A research-preview agent that uses its own web browser to complete tasks such as bookings and forms.

  • Computer-Using Agent (CUA) model
  • Pro users in the US first
Source ↗
Moonshot AI (Kimi)🧠 Model

Kimi k1.5

A reasoning model trained by scaling reinforcement learning, released the same day as DeepSeek-R1.

Source ↗
DeepSeek🧠 Model⭐ Flagship

DeepSeek-R1

An open reasoning model trained largely with reinforcement learning that rivalled OpenAI's o1 and shook markets.

  • MIT licence
  • Open weights
Source ↗
NVIDIA🧠 Model⭐ Flagship

Cosmos world foundation models (CES 2025)

Open generative models that simulate physical worlds to train robots and self-driving cars.

  • Open models for physical AI
Source ↗

Dec 2024

DeepSeek🧠 Model⭐ Flagship

DeepSeek-V3

A 671B mixture-of-experts open model trained at a strikingly low reported cost.

  • 671B total / 37B active
  • Open weights
Source ↗
Google DeepMind🧠 Model

Veo 2 and Imagen 3 update

A new video model with better physics and human motion, up to 4K resolution.

  • Up to 4K video
Source ↗
Google DeepMind🧠 Model⭐ Flagship

Gemini 2.0 Flash (experimental)

Google's model "for the agentic era", with native image and audio output and built-in tool use; Deep Research launches in Gemini.

  • Native tool use
  • GA on 5 Feb 2025
Source ↗
OpenAI🛠️ Product & feature

Sora video generator

OpenAI's text-to-video model opens to ChatGPT Plus and Pro users, with a storyboard editor.

  • Text or image → video
Source ↗
Meta AI🧠 Model

Llama 3.3 70B

An open text model that matches Llama 3.1 405B's quality at a fraction of the serving cost.

  • Open weights
Source ↗
OpenAI🧠 Model⭐ Flagship

o1 and ChatGPT Pro

The full o1 reasoning model ships, with a $200/month ChatGPT Pro plan that adds o1 pro mode.

  • Day 1 of "12 Days of OpenAI"
  • o1 pro mode spends extra compute on hard problems
Source ↗
Amazon🧠 Model⭐ Flagship

Amazon Nova models (re:Invent 2024)

Amazon's own model family: Micro, Lite and Pro for text and vision, Canvas for images and Reel for video.

  • Available in Amazon Bedrock
Source ↗
Microsoft AI🧠 Model

Phi-4

A 14B small language model that punches above its size on maths and reasoning; open weights followed in January.

  • 14B parameters
Source ↗

Nov 2024

Anthropic🔌 API & developer

Model Context Protocol (MCP)

An open standard for connecting AI assistants to the tools and data where work lives.

  • Open standard, later adopted across the industry
Source ↗

Oct 2024

Anthropic🧠 Model⭐ Flagship

Claude 3.5 Sonnet (new), 3.5 Haiku and computer use

A refreshed Claude 3.5 Sonnet, the new Claude 3.5 Haiku, and an experimental ability to operate a computer by looking at the screen.

  • Computer use: public beta in the API
Source ↗
Anthropic🔬 Research & safety

Updated Responsible Scaling Policy

A major revision of the risk-governance framework Anthropic uses to decide what safeguards each model needs.

Source ↗
OpenAI🔌 API & developer

DevDay 2024: Realtime API, prompt caching, distillation

A low-latency voice Realtime API, automatic prompt caching with a 50% input discount, vision fine-tuning for GPT-4o, and model distillation.

  • Prompt caching: 50% off repeated input
Source ↗

Sep 2024

Meta AI🛠️ Product & feature

Meta Connect 2024: Orion AR glasses and voice for Meta AI

Meta shows its Orion holographic AR glasses prototype and gives Meta AI a voice and the ability to understand photos across its apps, alongside Llama 3.2.

  • Orion AR glasses prototype
  • Voice and vision in Meta AI
Source ↗
Meta AI🧠 Model

Llama 3.2

Small models for phones and the first multimodal (vision) Llama models.

Source ↗
Alibaba Qwen🧠 Model⭐ Flagship

Qwen2.5

A large open release spanning general, coding and maths models.

Source ↗
Anthropic🔬 Research & safety

Contextual Retrieval

A retrieval technique that cuts failed retrievals by adding context to each chunk.

Source ↗
OpenAI🧠 Model⭐ Flagship

o1-preview and o1-mini

OpenAI's first reasoning models, which think step by step before answering.

  • Reasoning model
Source ↗
DeepSeek🧠 Model

DeepSeek-V2.5

Merges the chat and coder models into one that beats both.

Source ↗
Mistral AI🧠 Model

Pixtral 12B

Mistral's first multimodal model, open under Apache 2.0.

  • Vision input
Source ↗

Aug 2024

xAI🧠 Model⭐ Flagship

Grok-2

A new model with better reasoning, plus image generation in Grok.

Source ↗

Jul 2024

Google DeepMind🔬 Research & safety

AlphaProof and AlphaGeometry 2: IMO silver

Solve four of six International Mathematical Olympiad problems — silver-medal standard.

  • 4 of 6 problems
Source ↗
Mistral AI🧠 Model

Mistral Large 2

A 123B model designed to run on a single node with long context.

  • 123B parameters
Source ↗
Meta AI🧠 Model⭐ Flagship

Llama 3.1 405B

The first frontier-scale open-weight model, with 128K context.

  • 405B parameters
  • 128K context
Source ↗
OpenAI💲 Pricing

GPT-4o mini

A small, cheap model that replaces GPT-3.5 Turbo for most uses.

Source ↗

Jun 2024

Anthropic🛠️ Product & feature

Projects in Claude

Pro and Team users can organise chats with shared knowledge into Projects.

Source ↗
Anthropic🧠 Model⭐ Flagship

Claude 3.5 Sonnet and Artifacts

A mid-size model that beat Claude 3 Opus, launched with Artifacts for building content alongside the chat.

Source ↗
DeepSeek🧠 Model

DeepSeek-Coder-V2

An open coding model built on the V2 architecture.

Source ↗
NVIDIA🧠 Model

Nemotron-4 340B

Open base, instruct and reward models built for generating synthetic training data.

  • Open model licence
Source ↗
Alibaba Qwen🧠 Model⭐ Flagship

Qwen2

A new open family with stronger coding, maths and multilingual ability.

Source ↗

May 2024

Mistral AI🧠 Model

Codestral

Mistral's first coding model, fluent in 80+ programming languages.

Source ↗
Anthropic🔬 Research & safety

Golden Gate Claude

An interpretability demo: amplifying one learned feature makes Claude obsessed with the Golden Gate Bridge.

Source ↗
Microsoft AI🧠 Model

Build 2024: Phi-Silica and Copilot+ PCs

A small language model built into every Copilot+ PC, running locally on the laptop's NPU, plus Phi-3-vision.

  • Runs on-device
Source ↗
Google DeepMind🛠️ Product & feature

Google I/O 2024: Gemini 1.5 Flash, Project Astra, Veo, Imagen 3

A fast Gemini 1.5 Flash, a 2M-token context for 1.5 Pro, the Project Astra universal-assistant prototype, the Veo video model and Imagen 3.

  • 1.5 Pro: 2M-token context (waitlist)
  • Veo: 1080p text-to-video
Source ↗
OpenAI🧠 Model⭐ Flagship

GPT-4o

An "omni" model handling text, audio and vision in real time, made available to free users.

  • Real-time voice
  • Free tier access
Source ↗
Google DeepMind🔬 Research & safety

AlphaFold 3

Predicts the structure and interactions of proteins, DNA, RNA and small molecules (with Isomorphic Labs).

Source ↗
DeepSeek💲 Pricing⭐ Flagship

DeepSeek-V2 sparks a price war

A strong, cheap mixture-of-experts model whose low API price pushed ByteDance, Tencent and Baidu to cut theirs.

Source ↗

Apr 2024

Microsoft AI🧠 Model

Phi-3

Small open models, including one that runs on a phone.

  • Open weights
Source ↗
Meta AI🧠 Model⭐ Flagship

Llama 3

8B and 70B open models, described by Meta as the most capable openly available LLMs.

Source ↗

Mar 2024

NVIDIA🛠️ Product & feature⭐ Flagship

Blackwell platform (GTC 2024)

NVIDIA's next GPU architecture for trillion-parameter models, claimed up to 25x lower cost and energy than its predecessor for inference.

Source ↗
Anthropic🧠 Model⭐ Flagship

Claude 3 family (Opus, Sonnet, Haiku)

Three models with vision; Opus set new marks on many benchmarks.

  • Vision input
  • 200K context
Source ↗
xAI🧠 Model

Grok-1 open weights

xAI publishes the weights of its 314B-parameter Grok-1 model.

  • 314B parameters
Source ↗
Moonshot AI (Kimi)🛠️ Product & feature

Kimi 2-million-character context

Moonshot begins testing a Kimi version with a 2-million-character context window.

Source ↗

Feb 2024

Mistral AI🧠 Model

Mistral Large and Le Chat

A flagship model to rival GPT-4, a chat assistant, and a partnership with Microsoft Azure.

Source ↗
Google DeepMind🧠 Model

Gemma (open models)

Google's first open-weight family of lightweight language models.

  • Open weights
Source ↗
OpenAI🔬 Research & safety

Sora preview

OpenAI previews a video model that turns text into clips up to a minute long.

Source ↗
Google DeepMind🧠 Model⭐ Flagship

Gemini 1.5

A model with a context window of up to 1 million tokens.

  • 1M-token context
Source ↗
Alibaba Qwen🧠 Model

Qwen1.5

An open family in sizes from small to 72B parameters.

Source ↗

Jan 2024

Zhipu AI (Z.ai)🧠 Model⭐ Flagship

GLM-4

Zhipu's new model, claimed to approach GPT-4, with longer context and multimodal features.

Source ↗
OpenAI🛠️ Product & feature

GPT Store

A store for custom GPTs opens to paid ChatGPT users; over 3 million had been created.

Source ↗

Dec 2023

Microsoft AI🧠 Model

Phi-2

A 2.7B small language model that rivals much larger models.

Source ↗
Mistral AI🧠 Model⭐ Flagship

Mixtral 8x7B

An open sparse mixture-of-experts model that outperforms Llama 2 70B on most benchmarks.

  • Mixture-of-experts
  • 32K context
Source ↗
Google DeepMind🧠 Model⭐ Flagship

Gemini 1.0

Google's first natively multimodal model family: Ultra, Pro and Nano.

  • Ultra, Pro, Nano
Source ↗

Nov 2023

Amazon🛠️ Product & feature

Amazon Q (re:Invent 2023)

A generative-AI assistant for work that can be tailored to a business.

Source ↗
Anthropic🧠 Model

Claude 2.1

A 200K-token context window, fewer hallucinations and tool use in beta.

  • 200K context
Source ↗
OpenAI🏢 Company

Board removes Sam Altman

OpenAI's board ousts its CEO; he is reinstated on 22 November after employee and investor pressure.

Source ↗
Microsoft AI🛠️ Product & feature

Bing Chat becomes Copilot

Microsoft unifies its assistants under the Copilot name.

Source ↗
OpenAI🔌 API & developer

DevDay 2023: GPT-4 Turbo, Assistants API, GPTs

A cheaper GPT-4 Turbo with 128K context, the Assistants API, DALL·E 3 and text-to-speech in the API, and custom GPTs.

  • 128K context
Source ↗
xAI🛠️ Product & feature⭐ Flagship

Grok (beta)

xAI's first chatbot launches in beta for X Premium subscribers.

Source ↗
DeepSeek🧠 Model

DeepSeek Coder

DeepSeek's first open model, built for coding.

  • Open weights
Source ↗

Oct 2023

Moonshot AI (Kimi)🛠️ Product & feature⭐ Flagship

Kimi chatbot

Moonshot's assistant launches, able to read inputs of up to 200,000 Chinese characters.

  • 200K-character context
Source ↗

Sep 2023

Amazon🔌 API & developer

Bedrock generally available

Amazon Bedrock opens to all customers.

Source ↗
Mistral AI🧠 Model⭐ Flagship

Mistral 7B

Mistral's first model: a 7B open model under Apache 2.0 that beat larger models.

  • Apache 2.0
Source ↗
Meta AI🛠️ Product & feature⭐ Flagship

Meta Connect 2023: Meta AI assistant and Emu

Meta launches its Meta AI assistant (with real-time Bing search), AI characters based on public figures, the Emu image generator, and new Ray-Ban Meta smart glasses.

  • Meta AI assistant
  • Emu image model
  • Ray-Ban Meta glasses
Source ↗
Anthropic🏢 Company

Amazon invests up to $4B

Amazon becomes Anthropic's primary cloud provider with an investment of up to $4 billion.

Source ↗
OpenAI🧠 Model

DALL·E 3

A new image model built into ChatGPT, much better at following detailed prompts.

Source ↗
Anthropic🔬 Research & safety

Responsible Scaling Policy

Anthropic's framework of AI Safety Levels (ASL) that tie safeguards to model capabilities.

Source ↗
Anthropic🛠️ Product & feature

Claude Pro

A paid plan with 5x more usage and priority access.

Source ↗

Aug 2023

Alibaba Qwen🧠 Model

Qwen-7B (open source)

Alibaba's first open-weight Qwen models.

  • Open weights
Source ↗
Meta AI🧠 Model

Code Llama

Llama 2 models specialised for code.

Source ↗

Jul 2023

Meta AI🧠 Model⭐ Flagship

Llama 2

Open models licensed for commercial use, released with Microsoft.

  • Commercial use allowed
Source ↗
Anthropic🧠 Model⭐ Flagship

Claude 2 and claude.ai

Better performance, a 100K context window and a public chat website in the US and UK.

  • 100K context
Source ↗

May 2023

Microsoft AI🛠️ Product & feature

Build 2023: Windows Copilot and plugins

Copilot comes to Windows, and Microsoft and OpenAI back a shared plugin standard across ChatGPT and Microsoft 365 Copilot.

Source ↗
Anthropic🔌 API & developer

100K-token context window

Claude's context window grows from 9K to 100K tokens — about 75,000 words.

  • 9K → 100K tokens
Source ↗
Google DeepMind🧠 Model⭐ Flagship

PaLM 2 (I/O 2023)

A new language model with better multilingual, reasoning and coding skills, powering 25 Google products.

Source ↗
Anthropic🔬 Research & safety

Claude's constitution

How Claude's explicit written principles guide its behaviour through Constitutional AI.

Source ↗

Apr 2023

Google DeepMind🏢 Company

Google DeepMind formed

DeepMind and Google Brain merge into a single unit, Google DeepMind.

Source ↗
Amazon🔌 API & developer⭐ Flagship

Amazon Bedrock and Titan

A managed service for building with foundation models from AI21, Anthropic, Stability AI and Amazon's own Titan.

Source ↗
Alibaba Qwen🛠️ Product & feature

Tongyi Qianwen

Alibaba unveils its large language model and chatbot.

Source ↗
Meta AI🧠 Model

Segment Anything (SAM)

A promptable model that can cut out any object in any image.

Source ↗

Mar 2023

OpenAI🛠️ Product & feature

ChatGPT plugins

ChatGPT gains plugins, including web browsing and code execution.

Source ↗
Microsoft AI🛠️ Product & feature

Microsoft 365 Copilot

An AI assistant built into Word, Excel, PowerPoint, Outlook and Teams.

Source ↗
OpenAI🧠 Model⭐ Flagship

GPT-4

A large multimodal model accepting text and images, a big step in reasoning and exam performance.

  • Image input
Source ↗
Anthropic🧠 Model⭐ Flagship

Claude (first public release)

Claude and the faster Claude Instant become available through a chat interface and the API.

  • Early partners included Notion, Quora and DuckDuckGo
Source ↗
Anthropic🔬 Research & safety

Core views on AI safety

Anthropic sets out why it thinks transformative AI may arrive within a decade and how it approaches safety research.

Source ↗
OpenAI💲 Pricing

ChatGPT API (gpt-3.5-turbo)

The ChatGPT model opens to developers at $0.002 per 1,000 tokens — ten times cheaper than earlier GPT-3.5 models.

  • $0.002 per 1K tokens
Source ↗
Zhipu AI (Z.ai)🧠 Model

ChatGLM-6B

An open bilingual chat model that developers could fine-tune and run locally.

  • Open weights
Source ↗

Feb 2023

Meta AI🧠 Model⭐ Flagship

LLaMA

A family of efficient models released to researchers; its weights soon leaked, kick-starting open-model development.

  • 7B to 65B parameters
Source ↗
Microsoft AI🛠️ Product & feature⭐ Flagship

Bing Chat

A GPT-4-powered chat mode in Bing search.

Source ↗
Google DeepMind🛠️ Product & feature

Bard

Google announces its conversational AI chatbot.

Source ↗
Anthropic🏢 Company

Google Cloud partnership

Anthropic picks Google Cloud as a preferred cloud provider for training and serving its models.

Source ↗

Nov 2022

OpenAI🛠️ Product & feature⭐ Flagship

ChatGPT

A free chat interface to a conversational GPT-3.5 model — the fastest-growing consumer app of its time.

Source ↗
Meta AI🧠 Model

Galactica

A science-focused language model whose public demo was pulled within days over fabricated answers.

Source ↗

Sep 2022

OpenAI🧠 Model

Whisper

An open-source multilingual speech-recognition and translation model.

  • Open source
Source ↗

May 2022

Google DeepMind🧠 Model

Imagen

A text-to-image diffusion model pairing strong language understanding with photorealism.

Source ↗
Google DeepMind🔬 Research & safety

Gato

A single "generalist agent" transformer that performs about 600 tasks, from captioning to controlling a robot arm.

Source ↗
Meta AI🧠 Model

OPT-175B

Meta releases a 175B-parameter language model, with weights shared with researchers.

  • 175B parameters
Source ↗

Apr 2022

Anthropic🏢 Company

Series B: $580M

Anthropic raises $580 million to build infrastructure for studying and improving AI safety.

Source ↗
OpenAI🧠 Model

DALL·E 2

A text-to-image model producing far more realistic, higher-resolution images than the original DALL·E.

Source ↗
Google DeepMind🧠 Model⭐ Flagship

PaLM

A 540-billion-parameter language model trained on Google's Pathways system.

  • 540B parameters
Source ↗

Mar 2022

Google DeepMind🔬 Research & safety

Chinchilla

Shows models were undertrained: a 70B model on 4x more data beats the 280B Gopher — compute-optimal scaling.

  • 70B params, 1.4T tokens
Source ↗

Feb 2022

Google DeepMind🔬 Research & safety

AlphaCode

An AI system that writes competition code at roughly the level of a median human competitor.

Source ↗

Jan 2022

OpenAI🔬 Research & safety

InstructGPT

GPT-3 models fine-tuned with human feedback to follow instructions become the default in the API — the recipe behind ChatGPT.

  • RLHF
Source ↗

Dec 2021

Google DeepMind🔬 Research & safety

Gopher

A 280-billion-parameter language model, with studies of scaling, ethics and retrieval.

  • 280B parameters
Source ↗

Nov 2021

OpenAI🔌 API & developer

GPT-3 API opens without a waitlist

Developers in supported countries can sign up for the OpenAI API immediately.

Source ↗
No updates match these filters.
How this list is made: each item was checked against the lab's own announcement or a reputable report, and links to it. Dates are announcement or release dates; where only the month could be confirmed, the card shows the month. Benchmark and price figures are as the lab announced them. Coverage is deepest for labs whose announcements we could check directly, so a lab with fewer cards has not necessarily shipped less. Spot a missing or wrong item? Every entry lives in one file, data/lab-updates.json.