Claude Fable 5.1 leads the Intelligence Index at 66, five points clear of every non-Anthropic model
Anthropic released Claude Fable 5.1 on September 1, 2026, and by the end of the day Artificial Analysis had it at the top of its leaderboard. According to Artificial Analysis, Claude Fable 5.1 at maximum effort scores 66 on the Intelligence Index (v4.1.1) as of September 1, 2026, ahead of every other model the site has evaluated. The nearest model from another lab sits at 60.9, a gap of five points.
The numbers below are Artificial Analysis's, an independent benchmarking firm that runs the same evaluations against every model at its own expense. Where we quote a price, it is Anthropic's published list price or the rate on our own inference platform, Princeps Sator, which serves 25 non-Claude models and publishes the same Artificial Analysis scores beside each one.
Key takeaways
- Claude Fable 5.1 scores 66 on the Artificial Analysis Intelligence Index (v4.1.1), first of every model evaluated as of September 1, 2026.
- The lead over other labs is 5.1 points. GPT-5.6 Sol and Grok 4.6 tie for the best non-Anthropic score at 60.9.
- It is Anthropic's most expensive model, at $10 per million input tokens and $50 per million output tokens, twice the price of Claude Opus 5 for a three point gain.
- Benchmarking it cost $8,523. Artificial Analysis spent 140 million output tokens running the index once.
- The cheapest way to get within nine points is about 1% of the price. GLM-5.3-Flash scores 57.5 and costs $0.15 and $0.50 per million tokens on Princeps Sator.
Fable 5.1 holds the top two spots of the index by itself
Artificial Analysis evaluates each model at several effort settings and lists them separately. Claude Fable 5.1 at maximum effort scores 66, at extra-high effort 65, and at high effort 62. The first two are the two highest scores on the board, and the third lands level with Claude Fable 5, the model it replaces.
| Model (effort setting) | Intelligence Index v4.1.1 | Lab |
|---|---|---|
| Claude Fable 5.1 (max) | 66 | Anthropic |
| Claude Fable 5.1 (xhigh) | 65 | Anthropic |
| Claude Opus 5 (max) | 63.1 | Anthropic |
| Claude Fable 5 (max) | 62.1 | Anthropic |
| Claude Fable 5.1 (high) | 62 | Anthropic |
| GPT-5.6 Sol (max) | 60.9 | OpenAI |
| Grok 4.6 (high) | 60.9 | SpaceXAI |
| Kimi K3 (max) | 59.7 | Kimi |
| GLM-5.3 (max) | 59.5 | Z AI |
| Qwen3.8 Max | 58.1 | Alibaba |
| GLM-5.3-Flash | 57.5 | Z AI |
| GPT-5.5 (xhigh) | 56.3 | OpenAI |
| Gemini 3.7 Flash (high) | 56.0 | |
| DeepSeek V4 Pro (max) | 53.2 | DeepSeek |
Artificial Analysis publishes the Fable 5.1 scores to the nearest point for now; the other rows carry the one decimal its data feed reports. Source: Artificial Analysis, Claude Fable 5.1, Intelligence Index v4.1.1, retrieved September 1, 2026.
The index is a composite of nine evaluations: GDPval-AA (real-world office work), τ³-Banking (tool use), Terminal-Bench 2.1 (agentic coding), SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR (long-context reasoning). Artificial Analysis describes the current version in its v4.1.1 launch note. It has not yet published the per-evaluation breakdown for Fable 5.1, so we cannot say which of the nine drove the gain over Fable 5; when Fable 5 launched in June, it led on five of the ten evaluations then in the index, and set the record on Humanity's Last Exam at 53%.
The price did not move, except the part that matters for agents
Claude Fable 5.1 costs exactly what Claude Fable 5 did: $10 per million input tokens and $50 per million output tokens, with a 1 million token context window and 128,000 tokens of output per request. The one change is cache reads. Re-reading a cached prompt prefix now costs $0.25 per million tokens, a quarter of the Fable 5 rate and 2.5% of the base input price, where every other Claude model charges 10%.
| Claude Fable 5.1, USD per million tokens | Rate |
|---|---|
| Input | $10.00 |
| Output | $50.00 |
| Cache read | $0.25 |
| Cache write (5 minute) | $12.50 |
| Cache write (1 hour) | $20.00 |
| Batch input | $5.00 |
| Batch output | $25.00 |
Source: Anthropic, What's new in Claude Fable 5.1.
That cache-read cut is aimed at one workload: a coding or research agent that runs for hours and resends the same growing transcript on every tool call. In that loop, most of the billed tokens are cache reads, and a fourfold cut there does more to the invoice than any change to the headline price would. Anthropic's own guidance is blunt about where the model fits: start with Claude Opus 5 for most workloads, and reach for Fable 5.1 when Opus 5 at higher effort still falls short.
Running the benchmark once cost $8,523
Artificial Analysis reports what it spends to evaluate each model, and Fable 5.1 is among the most expensive runs it has published: 140 million output tokens and $8,523.16 to complete the Intelligence Index one time. When Fable 5 launched, the Humanity's Last Exam portion alone cost about $2,200, then a record.
The reason is that Fable 5.1 thinks before every answer and cannot be told not to. Adaptive thinking is always on, the thinking tokens are billed as output at $50 per million, and the only lever is the effort setting. That is also why the three effort settings land at different places on the leaderboard: max effort buys three more points than high effort, and pays for them in tokens.
What changed for developers, and what broke
For a team already on Claude Fable 5, the upgrade is a one-line model ID change with three breaking edges, all documented by Anthropic:
- Forced tool use is gone. Asking the model to call a specific tool, or any tool, now returns an error. Anthropic's reasoning is that a forced call skips thinking and pushes the model's working-out into the tool arguments.
- Thinking blocks are tied to the model. Fable 5.1 can read earlier models' reasoning, but no earlier model can read Fable 5.1's. A conversation that moves back to Opus 5 loses its reasoning for those turns.
- Editing conversation history invalidates reasoning. Change the system prompt, the tool list, or an earlier turn, and every later thinking block is rejected. Accounts created on or after August 31, 2026 get the error by default. Agent frameworks have to be append-only from now on.
Anthropic also lists behaviour shifts that arrive without any code change: fewer parallel tool calls per turn, fewer progress updates during long runs, denser prose, and a tendency to rewrite a whole file where a small edit would do. Every piece of text Fable 5.1 produces also carries Anthropic's statistical watermark on every platform. None of these show up in a benchmark score; all of them show up in a production bill.
Claude Mythos 5.1, released the same day, is the same underlying model without Fable's safety classifiers, offered only to approved organizations in Anthropic's Project Glasswing program for cybersecurity and life sciences work.
Where the 25 models on Princeps Sator land, and what they cost
Fable 5.1 is the best model money can buy today. The question for most workloads is whether the last five to eight points are worth eight to a hundred times the price, and that is a question our own platform exists to answer. Princeps Sator is a prepaid, pay-per-token inference API that serves 25 models over the same three wire formats the big labs use (OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages), so Claude Code, Codex, Cline, and Roo work against it by changing one base URL. It does not serve Claude. Its models page publishes the same Artificial Analysis scores used above beside every rate, which makes the comparison below straightforward.
| Model | Intelligence Index v4.1.1 | Points behind Fable 5.1 | Sator input, $/M | Sator output, $/M | Output price as % of Fable 5.1 |
|---|---|---|---|---|---|
| Claude Fable 5.1 (max), Anthropic list price | 66 | 0 | $10.00 | $50.00 | 100% |
| Grok 4.6 | 60.9 | 5.1 | $2.00 | $6.00 | 12% |
| Kimi K3 | 59.7 | 6.3 | $3.00 | $15.00 | 30% |
| GLM-5.3 | 59.5 | 6.5 | $1.40 | $4.40 | 8.8% |
| Qwen3.8 Max | 58.1 | 7.9 | $2.00 | $6.00 | 12% |
| GLM-5.3-Flash | 57.5 | 8.5 | $0.15 | $0.50 | 1.0% |
| DeepSeek V4 Pro | 53.2 | 12.8 | $1.32 | $3.96 | 7.9% |
Sator rates are the published per-token prices on sator.princep.org/pricing as of September 1, 2026; Artificial Analysis scores are for the highest effort setting it evaluated.
Two rows stand out. GLM-5.3 is 6.5 points behind Fable 5.1 on the index and costs $1.40 and $4.40 per million tokens on Sator, under 9% of Fable 5.1's output price. Its cache-read rate on Sator, $0.26 per million, is almost exactly what Anthropic just cut Fable 5.1's cache reads to. GLM-5.3-Flash scores 57.5, within 8.5 points of the leader, at 1% of the output price. For a batch classification job, a first-pass document summariser, or a coding agent's cheap subagents, that trade is usually the right one. For the hardest step in the chain, the one where Opus 5 at high effort still gets it wrong, Fable 5.1 now has the evidence to justify itself.
Sator also shows the speed figures Artificial Analysis has measured for these models, which is one place they beat the frontier outright. GLM-5.3 streams at 73 tokens per second with a 1.6 second time to first token; Claude Fable 5 at max effort streamed at 58 tokens per second and took 55 seconds before its first token, because it was thinking. Artificial Analysis has not yet published speed for Fable 5.1.
Frequently asked questions
Is Claude Fable 5.1 the best AI model in September 2026? On the Artificial Analysis Intelligence Index (v4.1.1), yes. It scores 66, ahead of Claude Opus 5 at 63.1 and the best non-Anthropic models, GPT-5.6 Sol and Grok 4.6, at 60.9.
How much does Claude Fable 5.1 cost? $10 per million input tokens and $50 per million output tokens, the same as Claude Fable 5. Cache reads dropped to $0.25 per million. Batch requests are half price at $5 and $25.
What is the difference between Claude Fable 5.1 and Claude Mythos 5.1? The same model. Fable 5.1 carries safety classifiers and is available to everyone on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Mythos 5.1 has no classifiers and is limited to approved Project Glasswing participants.
How does Claude Fable 5.1 compare to Claude Opus 5? Fable 5.1 scores 66 to Opus 5's 63.1 at maximum effort, at twice the price per token. Anthropic itself recommends starting with Opus 5 and moving up only when evaluations on Opus 5 at higher effort still fall short.
Can I use Claude Fable 5.1 through Princeps Sator? No. Sator serves 25 non-Claude models, the highest scoring of which is Grok 4.6 at 60.9. Claude Code users can point the tool at Sator with a single base URL setting and choose any of those models; see the Sator docs.
What to watch next
Artificial Analysis has scored Fable 5.1 but not yet published its speed, latency, or per-evaluation results; those usually follow within days and will show whether the gain came from the agentic evaluations or the knowledge ones. Anthropic has said nothing about bringing the 2.5% cache-read rate to Opus 5 or Sonnet 5, and whether it does will decide how much of the agentic-coding market Fable 5.1 actually takes. On our side, the Sator models page refreshes its benchmark snapshot as Artificial Analysis publishes, so the comparison table above will update as GLM, Kimi, Qwen and DeepSeek ship their answers.