⚡ Key takeaways
- Every frontier lab now sells agents and agent platforms, not just models.
- All of them add cheaper tiers each quarter (Hydrogen, Sol, Haiku-class models).
- Restricted-access tiers for cyber and bio work are spreading (Mythos 5.2, GPT-6 Cyber, Gemini 4 Carbon).
- Self-improvement is starting to be rated: GPT-6.2 Astra is reported as rated High for AI self-improvement.
- Distribution differs more than models do: consumer apps, clouds, Microsoft 365, Search.
🗺️ The playbooks
Thesis 1: Labs differ most in distribution, not in models.high confidence
| Lab | Builds itself | Pricing move | Distribution | Latest (Nov 2026) |
|---|---|---|---|---|
| OpenAI | Models, Codex, agent platform, apps | Two tiers (Astra/Sol), lower API prices | ChatGPT (reported 1B+ weekly users ⚠️), API, Microsoft | GPT-6.2 Astra and Sol, 17 Nov |
| TPUs, cloud, models, harness, apps | Fast cheap tier (Hydrogen) | Search, Android, Workspace, Cloud | Gemini 4 Hydrogen, 12 Nov | |
| Anthropic | Models, Claude Code, Agent SDK, MCP | Price cuts with Fable 5.2 (reported) | API, AWS and Google clouds, enterprise | Claude Fable 5.2 / Mythos 5.2, 9 Nov |
| Meta | Models, apps | API from $4 per million input tokens | Facebook, Instagram, WhatsApp | Muse Ember (Nov) |
| Microsoft | Platforms, Copilot, some models | Credits and seats | Microsoft 365, Azure | Foundry hosted agents GA (Jul) |
| Chinese labs | Models, clouds | Low list prices | Alibaba Cloud, open weights | Qwen4-Max, 16 Nov (closed) |
So what: When models converge, distribution decides who wins customers; companies without their own distribution must partner with someone who has it.
📊 Three shared patterns
Thesis 2: Up the stack, restricted tiers, rated self-improvement.medium confidence
- Up the stack.
- Every lab ships a coding agent and an agent platform; harness features are copied within months.
- Restricted tiers.
- Cyber and bio-capable models go to vetted organisations; the labs decide who qualifies.
- Rated self-improvement.
- System cards begin to rate how much a model can improve AI systems; independent checks of those ratings are rare.
So what: Each pattern creates demand for something the labs do not sell: independent checks of their gates and ratings.
🧭 A year of announcements, counted
🔧 Reproduce this: release_cadence.py
release_cadence.py · reads data/lab-updates.json
"""release_cadence.py — how each frontier lab spent 2026, from the site's own data.
Reads data/lab-updates.json (Frontier Lab Updates) and counts, per lab, the 2026 items
by type (model, product, API, pricing, research, company) and flagship releases.
Standard library only.
"""
import json, os
d = json.load(open(os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "..", "lab-updates.json")))
T = [t[0] for t in d["types"]]
by = {}
for x in d["items"]:
if not str(x["date"]).startswith("2026"): continue
r = by.setdefault(x["lab"], {t: 0 for t in T}); r[x["type"]] += 1; r["flag"] = r.get("flag", 0) + (1 if x.get("flagship") else 0)
print(f"{'lab':10} {'total':>5} " + " ".join(f"{t[:7]:>7}" for t in T) + " flagship")
for lab, r in sorted(by.items(), key=lambda kv: -sum(kv[1][t] for t in T)):
tot = sum(r[t] for t in T)
print(f"{lab:10} {tot:5d} " + " ".join(f"{r[t]:7d}" for t in T) + f" {r.get('flag', 0):8d}")
tot = sum(sum(r[t] for t in T) for r in by.values()); prod = sum(r["product"] for r in by.values())
print(f"2026 items: {tot}; products/features {prod / tot:.0%}. Coverage is uneven by lab (see article), so compare patterns, not totals")$ python3 release_cadence.py
lab total model product api pricing researc company flagship
anthropic 38 13 9 1 1 6 8 11
openai 19 12 5 1 0 1 0 7
google 15 15 0 0 0 0 0 5
qwen 8 8 0 0 0 0 0 3
meta 7 7 0 0 0 0 0 3
xai 7 5 0 1 0 0 1 1
nvidia 7 6 0 0 0 1 0 1
mistral 5 5 0 0 0 0 0 2
deepseek 4 4 0 0 0 0 0 2
zhipu 3 3 0 0 0 0 0 1
moonshot 3 3 0 0 0 0 0 2
microsoft 2 2 0 0 0 0 0 0
amazon 1 0 1 0 0 0 0 0
2026 items: 119; products/features 13%. Coverage is uneven by lab (see article), so compare patterns, not totals
Data caveat. The lab tracker logs Anthropic far more densely than other labs (including product and company items). Totals are not a measure of activity; use the table for the kinds of things each lab announces.
📄 Method and sources
- Profiles built from the site's Frontier Lab Updates (sourced, hand-curated).
- User numbers and some prices are company statements reported by press ⚠️.
- Pattern labels are our reading.
Facts are sourced; figures marked ⚠️ come from secondary reports or company statements. Scores, weights and predictions are our judgement and are labelled. Disclosure: Claude, made by Anthropic, helps write this site. Anthropic is one of the companies analysed here, under the same rules as the others.
Frontier Lab Updates (site data)
🧭 Foundations: the AI market, Nov 2026
Live data: AI Market Map & Investments · AI Investment Tracker.
Read
How the labs compete
You are here
Read
Read