Two model classes, one flat rate each.
A FlatClaw tenant pays for a node held warm, not for tokens. The flagship class turns one H100-class card into output at about 66 cents per million tokens. The frontier class runs a frontier-grade open model on eight GPUs inside the tenancy at about a dollar per million, a twentieth of frontier-lab list rates. Here is the math, refreshed October 2026.
Each class against the hosted models it competes with.
October 2026 list prices per million output tokens, one table per class, next to published head-to-head results. The flagship class is priced with the cheapest hosted models and outscores the mid-tier ones; the frontier class is priced with the cheapest hosted models and plays in the frontier tier.
Price, per 1M output tokens
The hosted models a flagship tenant would otherwise rent, from the entry tier up to Sonnet 5.5, including the open models the cheapest hosts serve.
| Claude Haiku 5.5tiered by request size | $0.50 – $2.50 |
| Llama 4 Maverick, hostedopen weights, cheapest hosts | $0.60 |
| FlatClaw flagship classGemma 4 31B at FP8 on one H100 | $0.66 |
| GPT-5.6 Luna | $1.20 |
| Gemini 3.5 Flash-Lite | $2.50 |
| Gemini 3.8 Flash | $3.75 |
| GPT-5.4 mini | $4.50 |
| Claude Haiku 4.5 | $5 |
| Mistral Medium 3.5, hosted | $7.50 |
| Claude Sonnet 5.5 | $10 |
Ability: Gemma 4 31B against the mid-tier hosted models
Published results against the hosted models on the price list. Gemma 4 31B outscores Claude Haiku 4.5, Llama 4 Maverick and Mistral Medium 3.5 on knowledge and reasoning, edges GPT-5.4 mini and Gemini 3.5 Flash-Lite, and leads Haiku 4.5 by 200 points of human preference.
Bars are drawn relative to the best score in each row. Arena Elo as of June 2026; benchmark figures as published by the labs and the Artificial Analysis, Vals and benchlm leaderboards; see sources.
Price, per 1M output tokens
The frontier-tier hosted models, and the same GLM-5.2 weights rented from a hosted API.
| FlatClaw frontier class, fast nodeGLM-5.2 on eight B200s, NVFP4 with multi-token prediction | $0.54 |
| FlatClaw frontier classGLM-5.2 on eight H200s, FP8 with multi-token prediction | $1.04 |
| GLM-5.2, rented from Z.aithe same open weights, hosted | $4.40 |
| Gemini 3.1 Pro$18 above 200K-token prompts | $12 |
| GPT-5.6 Terra | $12 |
| Claude Opus 5.5 | $20 |
| GPT-5.6 Solpromotional, $30 list | $20 |
| GPT-5.5 | $30 |
| Claude Fable 5.1 | $50 |
Ability: GLM-5.2 against the frontier labs
Published head-to-head results. GLM-5.2 leads GPT-5.5 on real-world bug fixing and Humanity's Last Exam, ties Opus-class models on agentic tool use, and sits within a few points everywhere else.
Bars are drawn relative to the best score in each row. Figures as published by the labs and the LM Council and Morph leaderboards; see sources.
| Claude Sonnet 5.5 at $10 | $33,000 |
| Claude Opus 5.5 at $20 | $66,000 |
| GPT-5.5 at $30 | $99,000 |
| GLM-5.2 rented at $4.40 | $79,000 |
| Claude Opus 5.5 at $20 | $360,000 |
| Claude Fable 5.1 at $50 | $900,000 |
| GLM-5.2 rented at $4.40 | $282,000 |
| Claude Opus 5.5 at $20 | $1.28 million |
| Claude Fable 5.1 at $50 | $3.2 million |
What a token costs on each node.
All-in figures: the GPU node, the control plane and the weights volume, at indicative list prices on our reference lane, the same classes on Azure, AWS and Google Cloud landing in the same band on reserved terms.
| Flagship class | Frontier class | |
|---|---|---|
| Model | Gemma 4 31B at FP8, 256K context | GLM-5.2, 744B mixture of experts, 1M context |
| Cloud node | 1 × H100 80 GB | 8 × H200 at FP8, or 8 × B200 at NVFP4 (fast) |
| Flat rate, held warm | about $2,200 a month | about $18,700 (H200) or $34,700 (B200) a month |
| Aggregate output, measured | about 1,260 tokens/s at concurrency 128 | about 6,800 tokens/s (H200) or 24,000 tokens/s (B200) |
| Output tokens per month | about 3.3 billion | about 18 billion (H200) or 64 billion (B200) |
| Per 1M output tokens | about $0.66 | about $1.04 (H200) or $0.54 (B200) |
Throughput references: Gemma 4 31B at FP8 on one H100 with SGLang, about 1,260 output tokens per second aggregate at concurrency 128. GLM-5 with multi-token prediction as measured by SemiAnalysis InferenceX, about 850 output tokens per second per H200 at FP8 and about 3,000 per B200 at NVFP4, eight of each per node. It is the batched aggregate that pays the bill, which is what naive comparisons miss.
Flagship class, in one line
About $0.66 per million output tokens. Against Sonnet 5.5 at $10 that is a fifteenth of the price for the same tier of work; against Opus 5.5, GPT-5.5 and Fable 5.1 it is 30–76× cheaper. It undercuts renting Gemma 4 31B from the cheapest host, and the node is private.
Frontier class, in one line
About $1.04 per million output tokens on the H200 node, $0.54 on the B200 node. Against Opus 5.5 at $20, GPT-5.5 at $30 and Fable 5.1 at $50 that is 19–93× cheaper for a model in the same tier, running inside the tenancy, and a quarter to an eighth of what renting the same weights from Z.ai costs. The fast node is both the cheaper token and the faster stream.
What the frontier class buys that a rate card cannot show.
A frontier-grade model inside the tenancy is a different product from a frontier-grade API, even at the same price per token.
A hosted frontier model decides what it will help with, and the vendor's policy sits between your people and the model. Claude Fable 5.1 ships with additional safety measures on dual-use capabilities, and the ungated variant is sold only to approved organizations. GLM-5.2 is open weights running in your tenancy: security research, red-team tooling, vulnerability analysis, chemistry, legal, whatever your business is, runs under your rule books, your approvals and your audit trail, with no provider-side classifier in the loop and no account to lose.
Weights, prompts, documents and case records stay on the node the customer owns, on the cloud they already run or on their own rack. No retention policy to negotiate, no training clause to read, no sub-processor list to review. The data-locality test is mechanical: tcpdump on the egress shows no inference traffic leaving.
The node is the tenant's. There is no per-minute cap at the moment a workflow fans out, no priority tier to buy, and the model version is pinned until the tenant chooses to move. A hosted model is retired, repriced or re-tuned on the vendor's calendar; this one changes on the customer's.
The GPU line is the same in a busy month and a quiet one, so the tenant can let agents run without a meter ticking: overnight batch jobs, long agentic sessions, a storefront chat widget on a sale day. Per-token pricing taxes exactly the usage that creates value.
Pick the node for the work.
The choice is not model brand or vendor preference. It is whether the hardest work is allowed to leave the building.
The default tenant. One H100-class card, a flat rate, and the cheapest private token on the market for the work most teams actually do all day.
For tenants whose hardest work must run on a frontier-grade model inside the tenancy: deep reasoning, long agentic runs, research the hosted models will not touch. Eight GPUs, one flat rate, no meter.
The flagship class as the floor and the frontier node for the hardest work, behind the same endpoint. Same platform, same agents, same front door.
The numbers move as open weights improve, as serving stacks improve, as providers reprice, and as GPU rates change. We re-run the math quarterly and ship whichever shape is cheapest per resolved task. Rate card refreshed October 8, 2026; hardware prices are live on the builds page.
- Anthropic API pricing
- OpenAI API pricing
- Gemini API pricing
- Z.ai GLM-5.2 pricing
- Gemma 4 31B hosted providers (Artificial Analysis)
- GLM-5 on B200 and H200 with multi-token prediction (SemiAnalysis InferenceX)
- GLM-5.2 serving recipe (SGLang docs)
- GLM-5.2 benchmarks: SWE-bench Pro, MCP-Atlas, HLE, GPQA, AIME (apidog)
- SWE-bench Verified and coding arena leaderboards (LM Council)
- SWE-bench Pro leaderboard (Morph)
- Gemma 4 31B vs Claude Haiku 4.5 (Artificial Analysis)
- Gemma 4 benchmarks and arena Elo
- Llama 4 Maverick (Artificial Analysis)
- GPT-5.4 mini benchmarks and pricing (benchlm)
- Gemini 3.5 Flash-Lite (Artificial Analysis)
- Mistral Medium 3.5 (Vals AI)