Anthropic shipped Claude Fable 5.1 and Claude Mythos on September 1, 2026, and the number that matters most is not on the benchmark chart — it is the price of a cache read, now $0.25 per million tokens, a 75% cut that takes up to 45% off the cost of running agentic work with Fable. Fable 5.1 is also, in the middle of an AI wave defined by images and video, a model that makes neither.
Claude Fable 5.1 is Anthropic's new flagship for coding and knowledge work, generally available today with the API id
claude-fable-5-1. Claude Mythos is the identical model with looser safeguards, restricted to vetted cyberdefenders and life scientists. Headline Fable results: 73.4% on CursorBench 3.2.0, 1853 on GDPval-AA v2, 52.6% on Terminal-Bench-Science. Input and output pricing is unchanged at $10/$50 per million tokens; cache reads dropped 75%. Neither model generates images or video.
Claude Fable 5.1 and Mythos are one model with two sets of safeguards
Claude Fable 5.1 and Claude Mythos are the same underlying model. The difference is entirely in the guardrails wrapped around it.
- Fable 5.1 — generally available to everyone, on the API and every major cloud.
- Mythos — vetted access only, through the Cyber Verification Program for defensive security work and the Life Sciences Verification Program, the latter built with the US government. Currently limited to a set of US organisations.
You can see the split in the benchmark table below. On Terminal-Bench 4.0, Fable 5.1 scores 55.8% and Mythos scores 60.9%. Anthropic attributes that five-point gap to cyber safeguards firing on tasks the unrestricted model completes, and says it expects the difference to shrink now that those safeguards are more precise.
Claude Fable 5.1 benchmark scores
Anthropic published Fable 5.1 against Fable 5, Opus 5, and GPT-5.6 Sol. Every score below is from the official announcement, read September 2, 2026.
| Benchmark | Fable 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Terminal-Bench-Science 0.1 | 52.6% | 24.7% | 29.0% | 22.4% |
| Terminal-Bench 4.0 | 55.8% | 42.0% | 52.3% | 37.3% |
| GDPval-AA v2 (knowledge work) | 1853 | 1723 | 1824 | 1711 |
| CursorBench 3.2.0 (coding) | 73.4% | 70.5% | 70.0% | 67.2% |
| AutomationBench | 31.4% | 17.1% | 26.9% | 19.6% |
| OSWorld 2.0 (partial) | 77.9% | 72.9% | 75.4% | — |
| OSWorld 2.0 (strict) | 41.7% | 36.1% | 39.6% | — |
| Humanity's Last Exam (no tools) | 60.9% | 57.8% | 56.6% | — |
| Humanity's Last Exam (with tools) | 65.0% | 63.8% | 63.6% | — |
| Input price / M tokens | $10 | $10 | — | — |
| Output price / M tokens | $50 | $50 | — | — |
| Cache read / M tokens | $0.25 | $1.00 | — | — |
Anthropic's own benchmark table, from the Claude Fable 5.1 announcement. Note the Mythos 5.1 line under Terminal-Bench 4.0 — that is the safeguard gap, not a different model.
The one Fable 5.1 score worth staring at
Terminal-Bench-Science more than doubled, 24.7% to 52.6% — a 27.9-point gain, against a next-largest of 14.3 points on AutomationBench. Anthropic flags a standard error of ±3.5–4.5 points per model, so read the small gaps loosely. A 27.9-point jump is not noise.
The coding gains are real and modest
CursorBench went 70.5% to 73.4%. Fable 5 was already strong here; 5.1 is better, not transformed. The striking reports come from partners, not the scoreboard. Millennium says Fable 5.1 found the cause of a one-in-a-million crash that no engineer and no other model had explained in four to five years — it disassembled a vendor library and matched it against the core dump.
The Claude Fable 5.1 price cut is the actual release
Anthropic left input and output pricing alone at $10 and $50 per million tokens. What changed is cache reads, which now cost $0.25 per million — 75% less than before.
That sounds like a footnote until you notice that in context-heavy, tool-heavy agent work, cache reads are most of the bill.
What 25% and 45% actually mean
Anthropic measured four weeks of real August 2026 usage and reports two figures for Fable 5.1 against Fable 5:
- Typical workloads — around 25% cheaper, covering usage across Claude Enterprise, Claude Code, and the API.
- Highly agentic workloads — up to about 45% cheaper, where context-heavy, tool-heavy work makes cache reads the dominant cost.
Cognition moved its Opus 5 traffic in Devin to Fable 5.1 on launch day, saying the new cache pricing finally makes a Fable-class model economical for workloads it had kept on the cheaper tier. Dan Shipper at Every put it as "Fable-level intelligence, Opus-level price, Sonnet-speed," measuring roughly twice the speed of Opus 5 at half the tokens.
What Claude Fable 5.1 does not do
The launch language — "the world's most advanced models" — invites the wrong conclusion, and the answer is not in the announcement at all. It is one line in Anthropic's model documentation, which almost no coverage quotes:
All current models support text and image input, text output, multilingual capabilities, vision, and tool use.
Text output. That is the whole answer. Fable 5.1 can read the screenshot you paste in and tell you what is wrong with your CSS. It cannot draw you a product shot, and nothing in the release says it ever will.
Fable 5.1 reads images. It does not create them. Anthropic makes no image or video generation claim for Fable 5.1 or Mythos anywhere in the release.
That is not a knock on the model — it is the documented shape of it. The same docs page carries three Fable 5.1 specs the announcement never mentions, and they are the ones that actually matter when you wire it into something:
| Spec | Claude Fable 5.1 |
|---|---|
| Context window | 1M tokens |
| Max output | 128K tokens |
| Reliable knowledge cutoff | June 2026 |
| Default effort on the API | high |
| Retirement no sooner than | September 1, 2027 |
Where the pixels actually come from
If what you need is an image or a video, the relevant frontier is elsewhere. On Kavel that means Seedance 2.5 or Veo 3.1 for video, and GPT Image 2 or Gemini 3 Pro Image for stills — all runnable in the browser on a free tier. The full image tools and video tools are there too, and the pricing page shows what each generation costs in credits.
The practical division of labour, if you are building anything: a text-output model writes the code and drafts the prompt; a generative media model renders the frame. Fable 5.1 is very good at the first job and cannot do the second at all.
What Fable 5.1 and Mythos did in the lab
| Result | Model | What it did |
|---|---|---|
| Protein binders | Mythos | Affinities 10x higher than the best entries in Adaptyv Bio's competitions on three targets; hit rate near 50% across 12 targets, against a typical 10–15%. Designs went to two external organisations for lab validation. |
| Venus elevation map | Fable 5.1 | Trained a network on 30-year-old NASA Magellan radar data to map a third of the planet at 2–3 km resolution instead of 10–20 km, with heights up to 25% more accurate. Being released under a Creative Commons licence ahead of NASA VERITAS and ESA EnVision. |
| GPU kernels | Mythos | Custom kernels for seven open-source deep learning models, up to 2.5x faster with identical outputs, cutting estimated GPU costs on genome-wide analyses by 30–60%. Anthropic says this normally takes performance engineers weeks. |
None of this is a Fable 5.1 product feature you can buy. It is a capability demonstration, and the protein and kernel work ran on Mythos, not the model you get.
Fable 5.1 safeguards got looser where they were wrong
Two changes matter to anyone who has been annoyed by a refusal.
Biology safeguards on Fable 5.1 and Fable 5 now fire 85% less often for benign requests about elementary biology and medical questions. Life-sciences R&D queries still route to the Opus models.
Cybersecurity safeguards were retuned so Claude Code users should see roughly 60% fewer interventions per session than under Fable 5's rules. Fable 5.1 is now permitted to identify software vulnerabilities — defensive work. Penetration testing, exploit generation, and binary-based vulnerability scanning still get redirected to Opus.
One Fable 5.1 change can break existing tooling. New API accounts created from September 1, 2026 onward can no longer manually edit Claude's prior context in a multi-turn conversation while preserving the transcript of Claude's prior thinking. This is an anti-distillation measure. Existing accounts are unaffected for now, but it will apply to everyone on future model releases.
Separately, models released after August 2, 2026 carry an invisible text watermark under the EU AI Act Code of Practice, with a detection API in private preview for regulators, researchers, and obligated enterprises. It encodes nothing about the user or their conversations.
Claude Fable 5.1 cost and availability
Fable 5.1 is available today across Amazon Web Services, Google Cloud, and Microsoft Azure, and on the Claude API as claude-fable-5-1. Enterprise Frontier Safeguards, which allow zero data retention with customer-controlled infrastructure, roll out in phases starting this fall; eligible customers can run Fable 5.1 with zero data retention in the meantime.
Mythos stays behind the CVP and LSVP, US organisations only, with Anthropic saying it is coordinating with the US government to widen access.
Claude Fable 5.1 FAQ
Is Claude Fable 5.1 available to everyone?
Yes. Fable 5.1 is generally available on the Claude API, Claude Code, Claude Enterprise, Claude.ai, and the major clouds. Mythos is not — it requires vetting through the Cyber Verification Program or the Life Sciences Verification Program, and is currently limited to US organisations.
Can Claude Fable 5.1 generate images or video?
No. The Fable 5.1 announcement claims coding, knowledge work, agentic research, computer use, and vision — where vision means interpreting images you supply. No image or video generation capability is announced. For generation, use a dedicated media model such as GPT Image 2 or Seedance 2.5.
How much does Claude Fable 5.1 cost?
$10 per million input tokens and $50 per million output tokens, unchanged from Fable 5. Cache reads dropped 75% to $0.25 per million, which Anthropic measures as roughly 25% off typical workloads and up to 45% off highly agentic ones.
What is the difference between Claude Fable 5.1 and Mythos?
They are the same model. Mythos carries more permissive safeguards for cybersecurity and life sciences work and ships only through trusted access programs. On Terminal-Bench 4.0 that safeguard gap shows up as 55.8% for Fable 5.1 versus 60.9% for Mythos.
Is Claude Fable 5.1 better than Opus 5?
On every published benchmark, yes — 73.4% versus 70.0% on CursorBench 3.2.0, 1853 versus 1824 on GDPval-AA v2, 52.6% versus 29.0% on Terminal-Bench-Science. Partners also report Fable 5.1 running about twice as fast as Opus 5 on half the tokens.
Does Claude Fable 5.1 watermark its output?
Yes, invisibly. Under the EU AI Act Code of Practice, models released after August 2, 2026 carry a statistical text watermark. It is undetectable without the detection API, has no effect on output quality, and encodes nothing about the user.
Sources
- Introducing Claude Fable 5.1 and Claude Mythos 5.1 — Anthropic, September 2026 (all benchmark, pricing, safeguard, and research figures above)
- Claude Fable 5 and Claude Mythos 5 — Anthropic, June 2026 (the prior release these numbers are measured against)
- Claude models overview — Anthropic platform docs, read September 2, 2026 (the modality line, context window, max output, knowledge cutoff, and retirement date)
- Anthropic newsroom — the September 1, 2026 publication date
- Adaptyv Bio protein design competitions — the external baseline for the binder affinity comparison
Last updated: September 2026.
