The Frontier Model Market Landscape: Who Builds, Trains, Serves and Sells the Models
Behind every AI product is a stack of companies: labs that build models, open-weight challengers, the software that trains and serves them, the hubs that distribute them, and the chips and clouds they run on. This visual deep dive maps each layer as of September 2026 โ with verified model line-ups and prices, an interactive cost calculator, a build-vs-buy break-even chart and a runnable cost model.
01 The Market Stack
From a model's creation to its use, five layers each have their own companies, business models and competitive dynamics. Click a layer:
02 Frontier Model Providers (Closed / Proprietary)
Closed providers sell access, not weights: through an API, their own apps, and the big clouds' model catalogs. Their line-ups have converged on the same shape โ a flagship, a balanced workhorse and a fast, cheap tier, most with million-token context and adjustable reasoning effort.
- GPT-6 Astra โ new flagship; approved users 3 Sep 2026, general availability the next day; ~1.05M context; $10 / $50
- GPT-5.6 family (July 2026) in three tiers: Sol (frontier), Terra (balanced), Luna (fast, low cost)
- Via API, ChatGPT and Amazon Web Services
- Claude Fable 5.1 (1 Sep 2026) โ $10 / $50
- Claude Opus 5.5 (22 Sep 2026) โ $4 / $20; 1M context
- Claude Sonnet 5 โ $2 / $10 ยท Claude Haiku 4.5 โ $1 / $5
- Claude Mythos 5.1 โ restricted to trusted-access programmes
- Via API, Claude apps, AWS Bedrock, Google Vertex AI, Microsoft Foundry
- Gemini 3.1 Pro โ current Pro model; 1M context; natively multimodal
- Gemini 3.5 Flash (19 May 2026) โ $1.50 / $9; default in the Gemini app
- Gemini Omni โ any-input-to-video generation model (I/O 2026)
- Gemini 3.5 Pro โ announced, still unreleased as of late August 2026
- Muse Spark (April 2026) โ Meta's first proprietary frontier model, from Meta Superintelligence Labs; API $1.25 / $4.25
- Also open weights โ see section 03
Prices are list prices in US dollars per million input / output tokens as of September 2026. Caching and batch discounts can cut effective cost substantially.
03 The Open-Weight Ecosystem
Open-weight models publish their parameters so anyone can download, fine-tune and host them โ the basis of self-hosting, on-device AI and most academic research. In 2026 the largest open models are giant sparse mixtures-of-experts, and China's labs release many of the strongest.
- Muse Glimmer 30B (August 2026) โ dense, Apache 2.0, runs on one GPU, tuned for agents and tool use
- Llama 4 Scout and Maverick (April 2025) โ MoE, Llama 4 Community License
- Meta's 2026 strategy is hybrid: closed Muse Spark plus open Glimmer
- DeepSeek-V4 (April 2026) โ V4-Pro 1.6T total / 49B active; 1M context
- DeepSeek-V4.1-Flash (September 2026) โ 552B total, ~8B / 16B active
- Strong reasoning and coding; detailed technical reports
- Qwen3.5 (February 2026) โ open Qwen3.5-397B-A17B; proprietary Qwen3.5-Plus
- Qwen3.8 (August 2026) โ open Qwen3.8-2.4T-A95B, the first open Max-class Qwen, and dense Qwen3.8-27B (Apache 2.0)
- Mistral 3 (December 2025) โ Mistral Large 3, a 675B-total / 41B-active MoE, plus dense 14B, 8B and 3B models
- All Apache 2.0
04 The Training Software Stack
Almost every model โ closed or open โ is trained with a small set of frameworks that implement the parallelism and memory techniques from Part 2:
05 Inference and Serving Engines
Serving engines implement continuous batching, paged KV cache, speculative decoding and quantization (Part 2). The open-source field has consolidated around a few engines. Pick what you need:
| Engine | Strengths | Notes |
|---|---|---|
| vLLM | High-throughput serving, PagedAttention, continuous batching, OpenAI-compatible API, broad hardware support | The most widely used open engine |
| SGLang | Fast serving for complex programs; RadixAttention prefix caching; strong on agents and structured output | Popular for large MoE deployments |
| TensorRT-LLM | NVIDIA-optimized kernels, quantization, in-flight batching | Often run under NVIDIA Dynamo or Triton |
| llama.cpp | Lightweight local inference on CPUs, Apple silicon and consumer GPUs; GGUF quantization | Behind many desktop apps (e.g. Ollama) |
| Hugging Face TGI | Production serving pioneer | In maintenance mode since December 2025; Hugging Face now recommends vLLM or SGLang |
| Commercial platforms | Managed inference, autoscaling, multi-model serving, monitoring | Together AI, Fireworks, Baseten and the clouds' own services |
06 Model Hubs and Distribution
07 Cloud and Accelerator Infrastructure
The whole market runs on a handful of chip families. NVIDIA dominates, but every hyperscaler now also designs its own accelerator, and frontier labs spread workloads across several to secure capacity and price.
- AWS โ broadest AI services, Bedrock
- Microsoft Azure โ Foundry, OpenAI models
- Google Cloud โ Vertex AI, TPUs
- Oracle Cloud (OCI)
- CoreWeave, Lambda, Nebius, Nscale
- Together AI (GPU clusters + inference)
- High-performance clusters, on-demand GPUs
- Compute is the largest cost of training and serving
- Capacity is often reserved years ahead
- Software stacks (CUDA, ROCm, XLA, Neuron) shape portability
08 Build vs Buy
Every organization chooses a point on a spectrum from "call an API" to "train our own model". Each step down gives more control and data isolation, and costs more engineering:
| Option | Examples | Control | Cost | Speed to use | Ops burden | Data governance | Typical use |
|---|---|---|---|---|---|---|---|
| Closed API | OpenAI, Anthropic, Google Gemini | Low | Pay per token | Fast | Low | Provider handles (contracts, regions, zero-retention options) | Most applications, rapid deployment |
| Self-hosted open weights | DeepSeek-V4, Qwen3.8, Mistral 3, Muse Glimmer | High | Infrastructure | Medium | High | Your infrastructure | Custom deployment, full control, regulated data |
| Fine-tuned, hosted open model | Qwen or Mistral on a cloud platform | Medium | Per token or per hour | Medium | Medium | Your data in your cloud account | Domain adaptation |
| Train from scratch | Custom model | Very high | Very high | Slow | Very high | Your infrastructure | Frontier research, unique capabilities |
The money question โ when does self-hosting beat an API? โ depends on volume. A GPU cluster costs the same whether it is busy or idle; an API costs per token. Move the sliders:
The same model as a script, using September 2026 list prices and the default assumptions above:
"""What does a workload cost on each frontier API โ and when does self-hosting an open model win?
API list prices (USD per 1M tokens, input / output) as published in September 2026. Self-hosting uses
stated assumptions you should replace with your own quotes and benchmarks. Standard library only.
"""
import math
API = { # model: (input $/1M, output $/1M)
"GPT-6 Astra": (10.00, 50.00),
"Claude Fable 5.1": (10.00, 50.00),
"Claude Opus 5.5": ( 4.00, 20.00),
"Claude Sonnet 5": ( 2.00, 10.00),
"Gemini 3.5 Flash": ( 1.50, 9.00),
"Muse Spark (Meta)": ( 1.25, 4.25),
"Claude Haiku 4.5": ( 1.00, 5.00),
}
WORKLOADS = { # name: (requests per month, input tokens, output tokens per request)
"support chatbot": (2_000_000, 1_500, 300),
"RAG over documents": ( 500_000, 8_000, 500),
"coding agent": ( 50_000, 60_000, 4_000),
}
def monthly(model, reqs, tin, tout):
pin, pout = API[model]
return reqs * (tin * pin + tout * pout) / 1e6
print("Monthly API cost (list prices, no caching or batch discounts)\n")
print(f" {'model':<20}" + "".join(f"{w:>22}" for w in WORKLOADS))
for m in API:
print(f" {m:<20}" + "".join(f"{'$' + format(round(monthly(m, *w)), ','):>22}" for w in WORKLOADS.values()))
# Self-hosting an open-weight model: assumptions, not quotes
GPU_HOUR, GPUS, TOK_PER_S, UTIL = 2.50, 8, 12_000, 0.5 # $/GPU-hour, GPUs per replica, total tokens/s per replica, average utilization
per_replica = GPU_HOUR * GPUS * 730 # 730 hours per month
capacity = TOK_PER_S * UTIL * 3600 * 730 # tokens a replica serves per month
print(f"\nSelf-hosted open model: {GPUS} GPUs x ${GPU_HOUR}/h = ${per_replica:,.0f}/month per replica, "
f"~{capacity / 1e9:.1f}B tokens/month at {UTIL:.0%} utilization (assumptions)")
for w, (reqs, tin, tout) in WORKLOADS.items():
toks = reqs * (tin + tout)
replicas = max(1, math.ceil(toks / capacity))
self_cost = replicas * per_replica
cheapest = min(API, key=lambda m: monthly(m, reqs, tin, tout))
print(f" {w:<20} {toks / 1e9:>6.1f}B tokens -> {replicas} replica(s) ${self_cost:>9,.0f} vs cheapest API here "
f"({cheapest}) ${monthly(cheapest, reqs, tin, tout):>9,.0f}")
print(f"\nAt full replicas, self-hosting costs ${per_replica / capacity * 1e6:.2f} per 1M tokens. Blended API price for the chatbot mix,")
print("and the monthly volume where one replica breaks even with it:")
reqs, tin, tout = WORKLOADS["support chatbot"]
for m, (pin, pout) in API.items():
blended = (tin * pin + tout * pout) / (tin + tout)
print(f" {m:<20} ${blended:>6.2f} per 1M break-even at {per_replica / blended * 1e6 / 1e9:>5.1f}B tokens/month")
print("\nSelf-hosting also needs engineers, on-call, evaluation and a model good enough for the task.")Output:
Monthly API cost (list prices, no caching or batch discounts) model support chatbot RAG over documents coding agent GPT-6 Astra $60,000 $52,500 $40,000 Claude Fable 5.1 $60,000 $52,500 $40,000 Claude Opus 5.5 $24,000 $21,000 $16,000 Claude Sonnet 5 $12,000 $10,500 $8,000 Gemini 3.5 Flash $9,900 $8,250 $6,300 Muse Spark (Meta) $6,300 $6,062 $4,600 Claude Haiku 4.5 $6,000 $5,250 $4,000 Self-hosted open model: 8 GPUs x $2.5/h = $14,600/month per replica, ~15.8B tokens/month at 50% utilization (assumptions) support chatbot 3.6B tokens -> 1 replica(s) $ 14,600 vs cheapest API here (Claude Haiku 4.5) $ 6,000 RAG over documents 4.2B tokens -> 1 replica(s) $ 14,600 vs cheapest API here (Claude Haiku 4.5) $ 5,250 coding agent 3.2B tokens -> 1 replica(s) $ 14,600 vs cheapest API here (Claude Haiku 4.5) $ 4,000 At full replicas, self-hosting costs $0.93 per 1M tokens. Blended API price for the chatbot mix, and the monthly volume where one replica breaks even with it: GPT-6 Astra $ 16.67 per 1M break-even at 0.9B tokens/month Claude Fable 5.1 $ 16.67 per 1M break-even at 0.9B tokens/month Claude Opus 5.5 $ 6.67 per 1M break-even at 2.2B tokens/month Claude Sonnet 5 $ 3.33 per 1M break-even at 4.4B tokens/month Gemini 3.5 Flash $ 2.75 per 1M break-even at 5.3B tokens/month Muse Spark (Meta) $ 1.75 per 1M break-even at 8.3B tokens/month Claude Haiku 4.5 $ 1.67 per 1M break-even at 8.8B tokens/month Self-hosting also needs engineers, on-call, evaluation and a model good enough for the task.
With these assumptions, all three example workloads are cheaper on an API. Self-hosting pays off above a few billion tokens a month โ sooner against flagship prices, later against small models โ or when control, data residency or customization matter more than price.
09 Market Structure
Two approaches compete, with a growing ecosystem between them:
10 Buyer Personas
- Uses models via API or cloud platforms
- Focuses on building products
- Buys: model APIs, managed serving
- Fine-tunes or self-hosts open models
- Needs security and compliance
- Buys: cloud infrastructure, model platforms, tooling
- Trains or fine-tunes models
- Uses distributed training and large compute
- Buys: GPUs, cloud clusters, training stack
- Experiments with new architectures
- Uses open models and tools
- Buys: compute, open-source tools, datasets
11 Key Market Trends, 2024 โ 2026
12 Where Differentiation Happens
13 Commoditizing vs Strategic Layers
An interpretive framework, not a prediction: some layers are becoming interchangeable, while others remain hard to copy and command a premium.
14 How the Pieces Connect
Next: Frontier Model Case Studies โ DeepSeek-V3 โ R1, Llama 4 and Qwen3 taken apart.