Claude Fable 5.1
Claude Fable 5.1 is the most capable model Anthropic offers to all customers, released on Sep 1, 2026. It costs US$ 10 per million input tokens and US$ 50 per million output, the same as Fable 5, but cache reads fell 75%, to US$ 0.25. In the launch table it wins all nine rows; in the system card it loses some tests even to its predecessor. Three weeks later, Opus 5.5 moved ahead of it on Anthropic's own tests, at US$ 4 and US$ 20, and became the recommended starting point.
Claude Fable 5.1 in official numbers
Data published by the maker in its documentation and announcement. The source of each item is at the end of the page.
- Maker
- Anthropic
- API identifier
claude-fable-5-1- Release
- Sep 1, 2026
- Input
- US$ 10 per million tokens
- Output
- US$ 50 per million tokens
- Cache read
- US$ 0.25 per million (2.5% of input; it was US$ 1 on Fable 5)
- Cache write
- US$ 12.50 (5 minutes) or US$ 20 (1 hour)
- Batch API
- US$ 5 input and US$ 25 output
- Fast mode
- not available for Fable 5.1
- Context window
- 1 million tokens, no surcharge
- Max output
- 128 thousand tokens
- Reliable knowledge
- through June 2026 (January on Fable 5)
- Reasoning
- adaptive, always on; five effort levels, default high on the API
- Where it runs
- Anthropic API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS; in the apps, Pro, Max, Team and Enterprise plans
- Data retention
- 30 days; zero retention only with Anthropic's authorization
- Retirement
- not before Sep 1, 2027
Where Claude Fable 5.1 does well, where it falls behind and who it is for
Where it does well
- Science in the terminal: 52.6% on Terminal-Bench-Science 0.1 in the launch table, against 29.0% for Opus 5 and 24.7% for Fable 5.
- Cache bill: cache reads fell from US$ 1 to US$ 0.25 per million. On real usage measured by Anthropic, the bill came out about 25% lower on typical workloads and 45% lower on highly agentic ones.
- Coding for less: 73.4% on CursorBench 3.2.0 at max effort, for US$ 9.64 per task, against 70.5% for Fable 5 at US$ 17.32, as measured by Cursor.
- Office work: 1,853 Elo on GDPval-AA v2, against 1,824 for Opus 5 and 1,723 for Fable 5.
- Long engineering work: 0.57 on FrontierSWE v2, ahead of Opus 5 (0.52), Fable 5 (0.48) and GPT-5.6 Sol (0.32), leading 22 of 33 tasks.
Where it falls behind
- Behind Opus 5.5: in the table published with Opus 5.5, it loses on terminal work (55.8% against 66.4%) and office work (1,735 against 1,846), while costing more than twice as much per token.
- Does not always beat its predecessor: in the system card, it loses to Fable 5 on ARC-AGI-1 (97.5% against 98.5%) and on GDP.pdf with tools (85.1% against 87.1%).
- Savings only in the cache: in Anthropic's indexed cost, the share of all other tokens went up (from 60 to 64 on typical workloads). Without caching, the bill does not drop.
- Speed: Anthropic rates its latency as the slowest in the current lineup, and there is no fast mode.
- Safeguards: dual-use cyber and life sciences R&D requests are rerouted to Opus models, and the safeguards still trigger more often than Opus 5's.
Who it makes sense for
- Agent sessions that run for many hours, such as multi-step research that ends in a document, spreadsheet or deck.
- Teams that already measured Opus 5.5 at high effort on their own tasks and it fell short: that is the documentation's rule.
- Apps that reread a lot of cached context, such as a large repository or a contract archive.
- Pro, Max, Team and Enterprise subscribers who want the most capable model in the Claude apps for long, asynchronous tasks.
When to pick another model
- As a default starting point: Anthropic says to start with Opus 5.5, which costs less than half per token.
- Live chat or quick answers: it is the slowest in the lineup and has no fast mode.
- Penetration testing, exploit generation and life sciences R&D: the safeguards reroute those requests to Opus models.
- Integrations that rely on forced tool use, prefill or non-default
temperature: they return a 400 error. - Operations that require zero data retention: the default is 30 days, and the exception depends on Anthropic's authorization.
What Anthropic published at launch
Fable 5.1 shipped on Sep 1, 2026 together with Mythos 5.1. Both have the same weights and the same price; only the level of safeguards changes. Mythos, more permissive on cyber and life sciences, is limited to a set of US organizations. Fable 5.1 is the one any customer can use.
In the announcement table, it wins all nine rows against Fable 5, Opus 5 and GPT-5.6 Sol. The system card says the default evaluation setup is max effort, averaged over five trials, and the announcement notes that tasks where the safeguards intervened scored zero on OSWorld 2.0.
| Fable 5.1 Anthropic | Fable 5 Anthropic | Opus 5 Anthropic | GPT-5.6 Sol OpenAI | |
|---|---|---|---|---|
| Terminal-Bench-Science 0.1 science in the terminal |
52.6% | 24.7% | 29.0% | 22.4% |
| Terminal-Bench 4.0 coding in the terminal |
55.8% | 42.0% | 52.3% | 37.3% |
| GDPval-AA v2 office work, Elo |
1,853 | 1,723 | 1,824 | 1,711 |
| OSWorld 2.0, partial computer use |
77.9% | 72.9% | 75.4% | not published |
| OSWorld 2.0, strict computer use |
41.7% | 36.1% | 39.6% | not published |
| Humanity's Last Exam without tools |
60.9% | 57.8% | 56.6% | not published |
| Humanity's Last Exam with tools |
65.0% | 63.8% | 63.6% | not published |
| AutomationBench business workflows |
31.4% | 17.1% | 26.9% | 19.6% |
| CursorBench 3.2.0 coding |
73.4% | 70.5% | 70.0% | 67.2% |
Figures from the Claude Fable 5.1 announcement; the system card says the default evaluation setup is max effort. The dot marks the best in each row. On Terminal-Bench-Science, Anthropic reports a standard error of 3.5 to 4.5 points per model.
Where it does not come first
The 212-page system card has the full table, and there Fable 5.1 does not lead everything: it loses tests to Opus 5 and even to Fable 5. On FrontierCode 1.1 Main, it scored 50.9% at medium, 2.6 points behind Fable 5 at xhigh.
Two reading traps. The HealthBench Professional score in the summary table (62.1%) is the length-adjusted value; the raw score is 74.2%, ahead of Opus 5. And the score changes with the test version: 73.4% on CursorBench 3.2.0, 51.8% on 4.0. It is a different test, not a worse model.
Competitors also publish figures for it, and they do not match. xAI gives Fable 5.1 57.9% on Terminal-Bench 4.0 and 70.0% on DeepSWE v1.1; Anthropic published 55.8% and 67.4%, the same figures OpenAI used. On Harvey, a legal work test, xAI shows 6.7% and Anthropic 19.09%.
| Fable 5.1 | Fable 5 | Opus 5 | |
|---|---|---|---|
| SWE-bench Multilingual coding |
89.1 | 86.6 | 89.5 |
| SWE-bench Multimodal coding with images |
54.7 | 54.1 | 59.4 |
| ARC-AGI-1 abstract reasoning |
97.5% | 98.5% | 97.5% |
| ARC-AGI-2 abstract reasoning |
90.0% | 89.2% | 90.42% |
| HealthBench Professional length-adjusted |
62.1% | 63.3% | 59.8% |
| GDP.pdf with tools |
85.1% | 87.1% | 85.5% |
| Toolathlon Verified Pass@1 |
77.8 | not published | 80.6 |
Claude Fable 5.1 and Mythos 5.1 system card, table 8.1.A and the section for each test. The dot marks the best in each row. On ARC-AGI-2, GPT-5.6 Sol appears with 92.5% in the same table.
The savings come from the cache
Anthropic announced Fable 5.1 as about 25% cheaper than Fable 5 on typical workloads and up to 45% cheaper on highly agentic work. The chart behind that claim, built on real usage from August 2026, splits the bill into two parts, and all of the savings sit in cache reads. The share of all other tokens went up slightly. On usage with little caching, the announced savings do not show up.
| Cache reads | All other tokens | Total | |
|---|---|---|---|
| Fable 5 typical workload |
40 | 60 | 100 |
| Fable 5.1 typical workload |
11 | 64 | 75 |
| Fable 5 highly agentic workload |
65 | 35 | 100 |
| Fable 5.1 highly agentic workload |
17 | 38 | 55 |
Indexed cost chart from the Claude Fable 5.1 announcement: real usage in August 2026, default effort, four weeks.
Fable 5.1, Opus 5.5 or GPT-6 Astra
On Sep 22, Anthropic launched Opus 5.5 saying it performs at the level of Fable 5.1 on most work. On CursorBench 4.0, Opus 5.5 at medium scores 52.5%, against 51.8% for Fable 5.1 at max, and in an internal test it rewrote HAProxy from C to Rust in 9.5 hours, against 12, at 51% lower cost. Anthropic cautions that the real gap is narrower than these scores suggest.
That is why the documentation says to start with Opus 5.5 and move up to Fable 5.1 only when Opus at high effort falls short. Outside Anthropic, the price peer is GPT-6 Astra, from OpenAI, at the same US$ 10 and US$ 50, but with cache reads four times as expensive and a surcharge above 272 thousand tokens.
| Fable 5.1 Anthropic | Opus 5.5 Anthropic | GPT-6 Astra OpenAI | |
|---|---|---|---|
| Input | US$ 10 | US$ 4 | US$ 10 |
| Output | US$ 50 | US$ 20 | US$ 50 |
| Cache read | US$ 0.25 | US$ 0.20 | US$ 1 |
| Cache write 5 minutes at Anthropic |
US$ 12.50 | US$ 5 | US$ 12.50 |
| Context window | 1 million | 1 million | 1.05 million |
| Above 272 thousand tokens | same price | same price | 2x input and cache, 1.5x output |
Anthropic and OpenAI pricing pages, checked on Oct 1, 2026.
Examples of using Claude Fable 5.1
An agent reads dozens of reports, cross-checks the numbers in a spreadsheet and builds the deck, in a session of hours that spans several apps. It is the use Anthropic describes for Fable 5.1, and GDPval-AA measures this kind of deliverable.
The whole codebase goes into the context and stays cached. Every new question rereads the repository for US$ 0.25 per million tokens, a quarter of what it cost on Fable 5. This is the kind of usage where the announced savings show up.
The system runs the task on Opus 5.5 and only moves up to Fable 5.1 when verification fails. It is the rule in Anthropic's documentation, and it avoids paying US$ 50 per million output tokens for what Opus solves at US$ 20.
Where Claude Fable 5.1 stands out and where it falls behind, by the numbers
Claude Fable 5.1 has published results on 34 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Dec 5, 2025 and Nov 30, 2026 that took the same test. Against the whole archive, which includes 2023 models, a recent release lands near the top of almost everything, and that says nothing.
Where it stands out
Tests where it sits in the top third of its own generation.
Where it falls behind
Tests where it sits in the bottom half of its own generation.
See all 34 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| AIME 2024 and 2025 (OTIS) | 100.0% | #1 / 78 (tied with 11) | #1 / 168 | 100.0% GPT-5.5 |
| Dtbench | 97.6% | #3 / 66 (tied with 1) | #3 / 148 | 98.9% Claude Opus 5.5 |
| Lmca | 65.46 | #2 / 66 | #2 / 121 | 68.23 Claude Opus 5.5 |
| CritPt | 31.1% | #6 / 77 | #6 / 112 | 32.3% GPT-5.6 Sol |
| WeirdML | 92.9% | #2 / 50 | #2 / 112 | 93.6% GPT-6 Astra |
| SciCode | 63.1% | #2 / 75 | #2 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 90.2% | #4 / 67 | #4 / 86 | 93.7% GPT-6 Astra |
| ARC-AGI 2 | 90.0% | #6 / 46 | #6 / 83 | 95.0% GPT-6 Astra |
| ALE-Bench | 2,143 | #6 / 60 | #6 / 82 | 2,951 GPT-6 Astra |
| WebDev Arena | 1,758 | #5 / 70 | #5 / 82 | 1,820 Claude Opus 5.5 |
| ARC-AGI | 97.5% | #7 / 46 (tied with 2) | #7 / 81 | 98.5% Claude Fable 5 |
| SimpleQA Verified | 70.8% | #6 / 62 | #6 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 48.0% | #17 / 62 (tied with 3) | #17 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 58.0% | #7 / 60 (tied with 1) | #7 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 100.0% | #1 / 67 (tied with 2) | #1 / 74 | 100.0% Claude Fable 5.1 |
| Frontiermath Tier 4 V2 | 87.8% | #6 / 57 | #6 / 67 | 100.0% GPT-6.1 Sol |
| Vending Bench 2 | 5,422 | #25 / 46 | #26 / 57 | 15,515 GPT-6 Astra |
| Apex Agents | 68.6% | #4 / 48 | #4 / 50 | 82.4% Gemini 4 Argon |
| FrontierCode | 50.9% | #6 / 43 | #6 / 43 | 54.6% Claude Opus 5.5 |
| Hle | 46.5% | #2 / 14 | #2 / 39 | 54.8% GPT-6 Astra |
| Gdp Pdf | 29.6% | #8 / 37 | #8 / 37 | 35.0% GPT-6.1 Sol |
| Furniture Assembly | 70.0% | #6 / 33 | #6 / 34 | 83.3% Claude Opus 5.5 |
| Blueprint Bench 2 | 41.9% | #4 / 30 | #4 / 31 | 54.4% Gemini 4 Argon |
| Gso | 88.2% | #1 / 13 | #1 / 28 | 88.2% Claude Fable 5.1 |
| Ebr Bench | 57.1% | #4 / 23 | #4 / 27 | 76.2% GPT-6 Astra |
| FrontierSWE | 56.3% | #4 / 20 | #4 / 20 | 65.5% GPT-6 Astra |
| CursorBench | 51.8% | #4 / 16 | #4 / 16 | 57.8% Claude Opus 5.5 |
| Rli | 17.9% | #3 / 10 | #3 / 15 | 21.3% Claude Opus 5.5 |
| Weirdml V3 | 26.0% | #3 / 10 | #3 / 11 | 42.2% GPT-6 Astra |
| MirrorCode | 73.3% | #2 / 9 | #2 / 9 | 77.4% Claude Opus 5.5 |
| Frontiermath Erdos | 0.0% | #4 / 7 (tied with 3) | #4 / 7 | 2.9% GPT-6 Astra |
| Spireclimb | 2.0% | #5 / 7 | #5 / 7 | 25.7% GPT-6 Astra |
| Spireclimb Sts1 A20 Heart | 0.0% | #4 / 7 (tied with 3) | #4 / 7 | 20.0% GPT-6.1 Sol |
| Spireclimb Sts2 A10 | 4.0% | #5 / 7 | #5 / 7 | 34.7% GPT-6 Astra |
For each test, the model's best effort level counts, as in the ranking. Source of the scores: Epoch AI, under CC BY 4.0.
What to compare Claude Fable 5.1 with
The models from other labs closest in capability and released in the same period, at most two per lab. Click any of them to see its page.
| Model | Lab | ECI | ELO | LiveBench | Release |
|---|---|---|---|---|---|
| GPT-6 Astra | OpenAI | 166.41 | n/a | 82.2 | Sep 3, 2026 |
| GPT-6.1 Sol | OpenAI | 166.32 | n/a | 81.6 | Sep 29, 2026 |
| Claude Fable 5.1 | Anthropic | 164.65 | 1,586 | 83.4 | Sep 1, 2026 |
| Gemini 4 Argon | Google DeepMind | 164.56 | 1,608 | n/a | Sep 30, 2026 |
| Kimi K3 | Moonshot | 157.42 | 1,528 | 79.2 | Jul 16, 2026 |
| Gemini 3.7 Flash | Google DeepMind | 157.22 | 1,551 | 78.8 | Aug 13, 2026 |
| Qwen 3.8 Max | Alibaba | 156.88 | 1,536 | 78.5 | Aug 2, 2026 |
Claude Fable 5.1 or Gemini 4 Argon?
Gemini 4 Argon is the model from another lab closest to Claude Fable 5.1 in capability. Across the 19 tests both took, this is the score:
| Test | Claude Fable 5.1 | Gemini 4 Argon |
|---|---|---|
| Vending Bench 2 | 5,422 | 13,718 |
| Blueprint Bench 2 | 41.9% | 54.4% |
| Chess puzzles | 48.0% | 59.0% |
| Apex Agents | 68.6% | 82.4% |
| CritPt | 31.1% | 27.1% |
| Ebr Bench | 57.1% | 64.8% |
| Frontiermath Tier 4 V2 | 87.8% | 78.0% |
| Game puzzles | 58.0% | 64.0% |
| SimpleQA Verified | 70.8% | 76.0% |
| FrontierCode | 50.9% | 47.6% |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
What Claude Fable 5.1 costs to deliver each task
Price per token is what is on the price sheet; cost per task is what lands on the invoice, because it depends on how many tokens the model spends. Each point is an effort level, published by whoever ran the test and compiled by Epoch AI. The further up and to the left, the more score per dollar.
CursorBench
The score scale does not start at zero.
Other models from Anthropic
The models from the same lab closest in time, newest first.
| Model | Release | ECI | ELO | LiveBench | Access |
|---|---|---|---|---|---|
| Claude Haiku 5.5 | Oct 7, 2026 | 158.98 | n/a | 72.1 | API access |
| Claude Sonnet 5.5 | Sep 28, 2026 | 164.45 | 1,537 | 77.8 | API access |
| Claude Opus 5.5 | Sep 22, 2026 | 166.82 | 1,584 | 83.2 | API access |
| Claude Fable 5.1 | Sep 1, 2026 | 164.65 | 1,586 | 83.4 | API access |
| Claude Opus 5 | Jul 24, 2026 | 162.60 | 1,573 | 80.1 | API access |
| Claude Sonnet 5 | Jun 30, 2026 | 156.36 | 1,491 | 76.0 | API access |
| Claude Fable 5 | Jun 9, 2026 | 162.13 | 1,542 | 83.0 | API access |
| Claude Opus 4.8 | May 28, 2026 | 158.26 | 1,507 | 76.2 | API access |
| Claude Opus 4.7 | Apr 16, 2026 | 156.32 | 1,532 | 76.5 | API access |
Have you used Claude Fable 5.1? Rate it
Users' opinion, not measurement: the rating is kept apart from the ranking and does not move anyone's position. One vote per browser, no sign-up, and you can change your rating later.
What people ask about Claude Fable 5.1
How much does Claude Fable 5.1 cost?
US$ 10 per million input tokens and US$ 50 per million output, the same as Fable 5. Cache reads cost US$ 0.25 (down from US$ 1), cache writes US$ 12.50 (5 minutes) or US$ 20 (1 hour), and batch is US$ 5 and US$ 25. There is no fast mode and no long context surcharge. Since Sep 24, 2026, a refusal that arrives before any output is billed in categories with few false positives.
Is Fable 5.1 better than Opus 5.5?
On the tests Anthropic published with Opus 5.5, no: Opus 5.5 is ahead on terminal work (66.4% against 55.8%) and office work (1,846 against 1,735), at US$ 4 and US$ 20. Anthropic itself says the gap is narrower than the scores suggest and recommends Fable 5.1 when Opus 5.5 at high effort falls short. The independent scores are at the top of this page.
Why do Fable 5.1 scores change from one page to another?
Because tests get new versions (73.4% on CursorBench 3.2.0, 51.8% on 4.0) and because each maker publishes its own figures for competitors: xAI gives it 57.9% on Terminal-Bench 4.0, and Anthropic 55.8%. Compare the same test version from the same source.
What is the difference between Fable 5.1 and Mythos 5.1?
They are the same model, with the same weights and the same price. Mythos 5.1 has more permissive safeguards on cybersecurity and life sciences and is therefore only available to a set of US organizations. Fable 5.1 is available to all API customers.
Where does Claude Fable 5.1 rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, Claude Fable 5.1 ranks #4 out of 267 models, with an ECI of 164.65. The ECI is a composite index that combines several hard benchmarks into a single scale. The position changes when a new model comes out, and the page is refreshed six times a day.
Is Claude Fable 5.1 better than Gemini 4 Argon?
Across the 19 tests both models took and Epoch AI compiled, Claude Fable 5.1 comes out ahead on 9, behind on 9 and tied on 1. It is a test-by-test count with no weighting: a win in math counts the same as a win in coding. To choose between the two, look at the tests that resemble your own use.
What is Claude Fable 5.1 best at?
Against models of the same generation (released up to 9 months before and up to 3 months after it), the best relative results of Claude Fable 5.1 are: Gso (#1 of 13); SciCode (#2 of 75); ProofBench (#1 of 67). The yardstick is its position among the models of that period that took the same test.
Where do the numbers on this page come from?
From the same three sources as the ranking, all openly licensed: Epoch AI (ECI and individual tests, CC BY 4.0), Arena (human-preference ELO, CC BY 4.0) and LiveBench (overall average, CC BY-SA 4.0). ROO3 does not measure any model. When the page shows a figure published by the maker itself, it appears in the written analysis, labeled as the maker's, and never enters the ranking.
Where each number comes from
- Anthropic, Claude Fable 5.1 and Claude Mythos 5.1 announcement
- Anthropic, Claude Fable 5.1 and Claude Mythos 5.1 system card
- Anthropic, Claude Fable 5.1 overview (documentation)
- Anthropic, what's new in Claude Fable 5.1 (documentation)
- Anthropic, pricing (documentation)
- Anthropic, models overview (documentation)
- Anthropic, Claude Fable page
- Anthropic, Claude Opus 5.5 announcement (comparison with Fable 5.1)
- xAI, Grok 4.7 announcement (figures xAI attributes to Fable 5.1)
- OpenAI, GPT-6 Astra announcement (figures OpenAI attributes to Fable 5.1)
- OpenAI, API pricing (GPT-6 Astra)
Epoch AI, "AI Benchmarking Hub". Published at epoch.ai. Data under the CC BY 4.0 licence, reordered by roo3.co.
The Arena (LMArena) leaderboard, from the public lmarena-ai/leaderboard-dataset on Hugging Face, under the CC BY 4.0 licence. We are not affiliated with the Arena.
LiveBench, by the LiveBench team, published at livebench.ai under the CC BY-SA 4.0 license. The overall average is recomputed by roo3.co with the same formula as the website (average of the categories), and those numbers stay under CC BY-SA 4.0. We are not affiliated with LiveBench.
← See the full ranking