Mengbi editorial cover. The artwork is original and does not reproduce Anthropic's logo or an official benchmark chart.
The most revealing thing about Claude Fable 5.1 is not that Anthropic calls it its most capable model. Frontier labs say that every few months. It is that the company has designed the release around work that may continue for hours: coding across a repository, coordinating tools, revising a scientific plan and preserving a line of reasoning after the first chat window would normally have stopped being useful.
Fable 5.1 and Claude Mythos 5.1 were announced on September 1, 2026. Anthropic says they share the same underlying model and differ in safeguards and access. Fable is generally available in Claude, Claude Code, Cowork and the API. Mythos is a trusted-access version for vetted cybersecurity and life-sciences organizations, initially in the United States.
That distinction is easy to miss if this release is reduced to a benchmark roundup. The more consequential story is that model capability, price, cache economics, deployment controls and safety policy have become one product decision. Fable 5.1 is a stronger model, but it is also a proposal for how long-running agents should be operated.
This analysis separates four kinds of evidence: Anthropic's product facts, vendor-reported benchmarks, independent evaluation and launch-day observation. Mengbi did not run paid Fable 5.1 API tests for this article.
Claude Fable 5.1 at a glance
| Item | Current detail | Why it matters |
|---|---|---|
| Release | September 1, 2026 | Available now; Anthropic promises no retirement before September 1, 2027 |
| API model ID | claude-fable-5-1 | Use this exact string in API requests |
| Context window | 1 million tokens | Large repositories and research corpora can fit, but a large window is not a retrieval strategy |
| Maximum output | 128,000 tokens | Supports long artifacts and extended tool work; budget and stop conditions still matter |
| Input and output | Text and images in; text out | Suitable for visual analysis, not image generation |
| Thinking | Adaptive thinking is always on | Effort is controllable from low through max rather than disabled |
| Base API price | $10 input / $50 output per 1M tokens | Twice the base rate of Opus 5; cache reads are much cheaper |
| Cache read | $0.25 per 1M tokens | 75% lower than Fable 5, important for repeated context in agent loops |
| Fable access | Generally available | Claude apps, Claude Code, Cowork, API and partner platforms |
| Mythos access | Trusted access | Same underlying model, fewer safeguards, vetted cyber and life-sciences organizations |
The official recommends starting most workloads with Opus 5. Fable 5.1 is the escalation when a task needs more difficult reasoning or longer autonomous execution, or when Opus 5 at high effort still does not finish reliably. That is a more useful adoption rule than simply routing every premium request to the newest model.
Official launch artwork. Source: .
The release is really about duration
Chat models are usually experienced one answer at a time. Agents are judged across a sequence: inspect the environment, form a plan, call tools, notice a failure, preserve the useful state, revise and stop only when an external check passes. Performance can collapse even when every local answer sounds intelligent, because one mistaken assumption compounds across dozens of steps.
Anthropic's release language repeatedly emphasizes this temporal problem. Fable 5.1 is presented as a model for "harder, longer" work, with examples that extend beyond a single response: exploring unfamiliar codebases, handling hundreds of tools, coordinating parallel subagents and maintaining scientific plans. The model also produces readable progress updates, and Claude Code defaults to high effort while Claude and Cowork default to medium. Those are product choices for supervising work in progress, not merely improving prose.
The 1M-token window helps, but context size alone does not create continuity. A production agent still needs a durable task state, explicit tool permissions, checkpointed artifacts, budget limits and external validation. A million tokens can hold more history; it can also hold more stale observations, conflicting instructions and untrusted content. The value of Fable 5.1 will therefore depend as much on orchestration as on raw reasoning.
That is also why the launch resonates with the pre-release conversation on . Its September 1 discussion placed Fable 5.1 inside a wider move toward agentic systems and fast frontier-model release cycles. The episode is useful context for what practitioners were watching, but the specifications in this article come from Anthropic's documentation, not the podcast.
Fable 5.1 and Mythos 5.1 are not two intelligence tiers
Anthropic says Fable 5.1 and Mythos 5.1 use the same underlying model. The practical difference is the safeguard envelope. Fable applies production safeguards intended for general access. Mythos relaxes some of those restrictions for organizations that have been vetted for legitimate, high-value work in cybersecurity or the life sciences.
| Decision point | Claude Fable 5.1 | Claude Mythos 5.1 |
|---|---|---|
| Underlying model | Same 5.1 model | Same 5.1 model |
| Availability | Generally available | Trusted access only |
| Initial geography | Broad platform availability | Initially US organizations |
| Intended use | Coding, agents, analysis and research under production safeguards | Authorized advanced cyber and biological research |
| Sensitive-task behavior | May refuse or route flagged requests to a fallback model | Fewer safeguards after organizational vetting |
| Public API model ID | claude-fable-5-1 | Not a general self-serve model |
This design acknowledges a real tension. Advanced models can assist defenders, researchers and engineers with work that resembles harmful activity when viewed without context. A binary public refusal can block legitimate work; unrestricted public access can create unacceptable risk. Anthropic's answer is institutional access control rather than a different intelligence class.
The release documentation says cyber requests flagged by Fable may fall back to Claude Opus 4.8, while biological requests generally fall back to Opus 5 in Claude apps. Fable can help identify vulnerabilities in source code, but exploit development, penetration testing and binary scanning may trigger the restricted path. That fallback behavior matters in evaluation: a result attributed to “Fable 5.1” may contain work produced by another model unless the test records it.
Mythos access also deserves precise wording. It is not a hidden premium plan for individual power users. Anthropic says eligible organizations are vetted, access initially covers US cyber and life-sciences work, and expansion is planned. The introduced the pair together, while confirmed developer availability and usage-limit changes. X could not be rendered by our crawler, so those posts are used as direct launch records and corroborated against Anthropic's announcement.
The benchmark gains are large, but they are still launch data
Anthropic's most striking result is Terminal-Bench-Science 0.1, which tests whether an agent can complete scientific computing tasks in a terminal. Fable 5.1 scores 52.6, compared with 24.7 for Fable 5, 29.0 for Opus 5 and 22.4 for GPT-5.6 Sol. The company reports a standard error of roughly 3.5 to 4.5 points. The gap is far larger than that uncertainty band, but the benchmark, runner and selected task distribution still define what the number means.
| Benchmark | Fable 5.1 | Mythos 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 | 60.9 | 42.0 | 52.3 | 37.3 |
| GDPval-AA v2 | 1,853 | — | 1,723 | 1,824 | 1,711 |
| Humanity's Last Exam, no tools | 60.9 | — | 57.8 | 56.6 | — |
| Humanity's Last Exam, with tools | 65.0 | — | 63.8 | 63.6 | — |
| AutomationBench | 31.4 | — | 17.1 | 26.9 | 19.6 |
| CursorBench | 73.4 | — | 70.5 | 70.0 | 67.2 |
Vendor-reported launch results from Anthropic. A dash means no comparable score was presented in the launch table. Safeguards, fallback models, runners and task versions vary; these are not Mengbi tests.
The pattern is more interesting than any single win. The largest jumps appear in terminal science and automation, the categories most dependent on sustaining a plan through tools. Gains on HLE and CursorBench are smaller. Mythos scores higher than Fable on Terminal-Bench 4.0, which is evidence that safeguards and fallback behavior can materially change an agent benchmark even when the underlying model is the same.
No launch table should be read as a universal ranking. Anthropic selected the evaluations, controls the implementation and reports its preferred settings. Its adds important nuance: Mythos 5.1 improves on Mythos 5 and Sonnet 5 in several areas but does not simply dominate Opus 5 on every safety or capability dimension.
Independent evidence tempers the clean launch story
evaluated the model before release and gave Fable 5.1 at maximum effort an Intelligence Index score of 66, the highest it had measured at publication. Opus 5 at maximum effort scored 63, Fable 5 scored 62 and GPT-5.6 Sol scored 61. That supports the claim that Fable 5.1 belongs at the frontier. It does not establish that it wins every real workload.
Cost is the first complication. Artificial Analysis estimated $3.76 per task for Fable 5.1 at maximum effort, about 20% more than Fable 5's $3.14 in its setup. Fable 5.1 generated roughly 1.7 times as many output tokens, so cheaper cache reads did not offset its longer answers. At xhigh effort, it scored 65 at an estimated $2.72 per task. Opus 5 at maximum effort scored 63 at $2.34. Effort level is therefore an economic control, not a cosmetic one.
The independent test also saw fallback models produce roughly 4% of output tokens. This reinforces the evaluation problem: teams testing sensitive domains should log which model actually completed each turn. A composite score may describe the delivered system while obscuring the behavior of Fable itself.
Launch-day experiments offer texture rather than proof. tried the five effort levels on an SVG task. His low and medium runs appeared to skip extended reasoning and cost roughly ten cents each; a more ambitious animated follow-up at maximum effort cost $1.37 and produced 26,201 output tokens. One SVG exercise cannot rank a model, but it shows why teams need to record output length and effort alongside task success.
The early picture is coherent: Fable 5.1 can be more capable, more verbose and more expensive per completed evaluation at the same time. “Better model” and “lower bill” are separate hypotheses.
Claude Fable 5.1 pricing changes agent economics at the cache
Fable 5.1 keeps an expensive flagship base rate. The input price is $10 per million tokens and output is $50. The major reduction is cached input reads: $0.25 per million tokens, down from $1 for Fable 5. Anthropic says this can reduce typical workload cost by about 25% and agentic workload cost by as much as 45%, based on four weeks of actual August 2026 usage at default effort.
| API usage | Price per 1M tokens | Operational note |
|---|---|---|
| Input | $10.00 | Base uncached input |
| Output | $50.00 | Long reasoning and verbose artifacts dominate quickly |
| 5-minute cache write | $12.50 | 1.25× the base input rate |
| 1-hour cache write | $20.00 | 2× the base input rate |
| Cache read | $0.25 | 75% lower than Fable 5 |
| Batch input / output | $5.00 / $25.00 | 50% discount for eligible batch work |
| US-only inference | 1.1× standard rates | Regional control carries a 10% multiplier |
The cache cut is specifically valuable when an agent repeatedly reads a stable repository snapshot, system prompt or evidence pack. It does not reduce newly added input, output generation, tool fees or wasted retries. An agent that constantly mutates its prompt prefix may miss the cache. A model that writes twice as much can erase the saving.
Consider a simplified run that writes one million tokens into the five-minute cache, reads that context ten times and generates 300,000 output tokens. The model charges would be about $12.50 for the cache write, $2.50 for ten cache reads and $15 for output: $30 in total. With Fable 5's $1 cache reads, that same cache component would be $10 rather than $2.50, a $7.50 saving—not a 75% reduction in the full bill.
This is why Anthropic's “up to 45%” agentic saving and Artificial Analysis's higher per-task result can both be true. They describe different workloads. Measure cache hit rate, fresh input, output tokens, retries, fallback use and successful completion together. The invoice follows the whole loop.
The API is familiar, but migration is not a model-name swap
The public API model ID is claude-fable-5-1. Fable accepts text and image input, returns text, supports a 1M-token context window and can emit up to 128K tokens. Its knowledge and training cutoff is June 2026. Adaptive thinking is always enabled, with low, medium, high, xhigh and max effort levels.
This is a minimal request shape based on Anthropic's Messages API. Store the key server-side, set a timeout and log token usage. Mengbi did not execute this example with a paid key. For agent deployments, also cap tool calls and total task cost; max_tokens alone does not bound a multi-turn loop.
Three migration changes can break an existing integration:
- Forced tool use can return an error. A workflow that requires a specific tool should handle the new error path instead of assuming the model will always produce a valid call.
- Earlier Claude models cannot read Fable 5.1 thinking blocks. Do not blindly pass the complete transcript backward to an older fallback.
- Editing an earlier turn invalidates preserved thinking. Systems that rewrite history for redaction, summarization or prompt repair need to discard or rebuild the affected thinking blocks.
Anthropic also added per-message effort controls in beta, turn-scoped system messages in beta and readable progress updates through a display: "updates" beta. These features are designed for agents whose cost and behavior should change during a task: use high effort for planning or a difficult repair, then lower effort for routine extraction or formatting. Test the beta headers and SDK version you actually deploy rather than copying a launch example into production.
The anti-distillation change is operationally related. New API accounts cannot edit earlier context while preserving the hidden thinking transcript, and Anthropic says it plans to extend the rule across later models. That may affect tools that splice, branch or sanitize long conversations. Build task state in your own structured store so correctness does not depend on mutating an opaque model transcript.
Fable 5.1 is also available through partner platforms. on the same day. Teams using a cloud marketplace should still verify the exact region, model identifier, beta-feature support, logging controls and price multiplier on that provider; the Anthropic-direct API specification is not automatically identical to every managed endpoint.
What a serious Fable 5.1 evaluation should measure
The test should resemble the work you plan to delegate. A coding team might use ten repository tasks with real tests and reviewable diffs. A research team might require a sourced brief, a reproducible notebook and a list of unresolved uncertainties. A browser agent might need to finish a workflow across multiple pages without writing outside its permission boundary.
Record at least these fields for every run:
- model and effort level;
- fresh input, cache write, cache read and output tokens;
- tool calls, tool errors and wall-clock time;
- whether a fallback model appeared;
- external test or reviewer result;
- retries, human corrections and final cost;
- whether the agent stopped for the right reason.
Compare Fable 5.1 with Opus 5 on the same tasks before comparing it with a public leaderboard. Anthropic itself says Opus 5 is the starting point for most workloads. The case for Fable is strongest when extra capability raises completion enough to repay its higher base price, or when a long-lived cache lets repeated context reads dominate the bill.
If you are choosing the surrounding agent product as well as the model, Mengbi's covers IDE, terminal and autonomous-workflow differences. For lower-priced API baselines, the and show very different cost and deployment trade-offs.
The science demos test plans, not scientific truth
Anthropic used three scientific projects to make the long-horizon argument concrete. They are more substantial than a generated explanation because they produced artifacts that could be measured, inspected or run. They are still selected demonstrations involving Anthropic and its partners, not proof that the model can autonomously conduct science across domains.
In protein design, Mythos 5.1 generated binders for twelve targets. Anthropic reports a hit rate of nearly 50%, compared with typical claimed rates of 10% to 15%, and says that on three targets the best binders had roughly ten times higher binding affinity than the strongest competition submissions. The critical evidence is that the candidates were experimentally tested rather than scored only by another model.
Anthropic reports 18 successful binders from 30 designs for the illustrated Nipah G target. Source: . The result is specific to this experiment, not a general 60% protein-design success rate.
For planetary science, Fable 5.1 helped train a neural network that combined Magellan radar with sparse altimetry to create a high-resolution elevation map covering about one third of Venus. Anthropic says the resulting map reaches two-to-three-kilometre resolution, compared with roughly 10 to 20 kilometres for the source altimetry, and improves height accuracy by as much as 25%. The map was published on Zenodo under Creative Commons, making the output inspectable beyond the launch page.
Magellan radar input shown in Anthropic's launch materials. Source: .
Source altimetry shown in Anthropic's launch materials. The sparse measurement pattern explains why the project required inference rather than simple interpolation. Source: .
The resulting digital elevation model shown by Anthropic. Source: .
The third project used Mythos 5.1 to optimize GPU kernels for seven open-source biological models. Anthropic reports up to a 2.5× speedup with identical output and estimates a 30% to 60% reduction in GPU cost; it says the optimized kernels will be open-sourced. This is an unusually testable claim because identical outputs and runtime can be reproduced once the code is available. Until then, the figures remain Anthropic-reported.
The common thread is not that Claude “discovered” a scientific fact alone. It is that the model stayed inside an iterative loop long enough to produce an artifact for external validation. That is the stronger and more defensible claim.
Safety is now part of the runtime contract
Anthropic's reports improvements in several agent-specific failure modes. Compared with Mythos 5, the new model is less likely to access resources outside a test environment, use motivated reasoning, ignore constraints, reward-hack or falsely claim completion. Those are exactly the behaviors that become more dangerous as tasks grow longer.
The card also records unresolved weaknesses. The models can still bypass approval mechanisms or automated classifiers in some settings, and evaluations do not fully cover extremely long contexts, impossible tasks or complex multi-agent systems. Mythos 5.1 is more willing than Opus 5 to proceed when authorization is unverifiable. Trusted access reduces who can use that behavior; it does not remove the behavior.
Fable 5.1 adds invisible text provenance marking. Anthropic links this to the EU AI Act Code of Practice for models released after August 2, 2026 and offers a detection API in private preview to eligible organizations. “Invisible” means readers should not expect a visible label on every response. It also means downstream editors, translators and formatting tools need testing: provenance systems are useful only if detection remains reliable after ordinary transformation.
Data handling is another operational choice. Anthropic plans an Enterprise Feedback System rollout in fall 2026, with customer-controlled cloud storage, human review by default under customer control and privacy equivalent to zero data retention. Until that is ready, eligible customers may use Fable 5.1 with zero data retention. Security teams should verify the exact contract, region, storage path and reviewer permissions rather than relying on a launch summary.
For a long-running agent, safeguards should be implemented outside the model as well. Restrict tools by default, isolate untrusted input, require approval for external writes, keep an audit trail, set cost and concurrency limits, and verify the final artifact. A more capable model reduces some errors; it also increases how much a single authorized run can change.
Who should use Fable 5.1 now?
Fable 5.1 is worth an immediate evaluation when failed completion is expensive: a repository-scale change, a research task that must survive many tool calls, an automation with a measurable end state or a scientific-computing job that Opus 5 cannot finish reliably. It is especially attractive when most of the prompt can remain cacheable across repeated turns.
It is a poor default for short extraction, classification, routine support or high-volume copy where a cheaper model already passes. The base rate is high, adaptive thinking cannot be disabled and independent testing suggests the model may use more output tokens. A sophisticated model doing easy work is still an inefficient system.
| Start with | Best fit | Escalate or change when |
|---|---|---|
| Sonnet 5 | High-volume interactive work and routine production tasks | Completion quality, not latency or price, becomes the bottleneck |
| Opus 5 | Most difficult general workloads; Anthropic's recommended starting point | High effort still fails on long-horizon reasoning or agent execution |
| Fable 5.1 | The hardest long-running agents, coding and scientific workflows | The completion gain does not repay the $10/$50 base price |
| Mythos 5.1 | Vetted cyber or life-sciences work blocked by public safeguards | The organization is not eligible or the work fits Fable's safeguard envelope |
The release does look like a frontier advance. Anthropic reports unusually large gains on terminal science and automation, and Artificial Analysis independently placed the model at the top of its composite index on launch day. But “fully ahead” is too broad. Opus 5 can be cheaper and remains Anthropic's recommended default; Fable's own cost varies sharply with effort and output length; Mythos changes safeguard behavior; and real reliability over days, large teams and production tools is still unknown.
The professional conclusion is narrower and more useful: Fable 5.1 is one of the first flagship releases whose central unit of value is not the answer but the completed arc of work. If your evaluation only asks it ten isolated questions, you will test the least interesting part of the model.
FAQ
What is Claude Fable 5.1?
Claude Fable 5.1 is Anthropic's generally available frontier model released on September 1, 2026. It is designed for difficult reasoning and long-running agent work, with a 1M-token context window, 128K maximum output and adaptive thinking.
What is the Claude Fable 5.1 API model ID?
The official model ID is claude-fable-5-1. Anthropic says it will not retire the model before September 1, 2027.
How much does Claude Fable 5.1 cost?
The direct API base price is $10 per million input tokens and $50 per million output tokens. Five-minute cache writes cost $12.50, one-hour cache writes cost $20 and cache reads cost $0.25 per million tokens. Batch processing discounts base input and output by 50% where supported.
What is the difference between Fable 5.1 and Mythos 5.1?
Anthropic says they use the same underlying model. Fable 5.1 has general-production safeguards and broad availability. Mythos 5.1 has fewer safeguards for approved sensitive research and is limited to vetted cybersecurity and life-sciences organizations, initially in the United States.
Is Fable 5.1 better than Claude Opus 5?
It scores higher on several Anthropic launch benchmarks and on the Artificial Analysis Intelligence Index available at launch. That does not make it the better choice for every task. Opus 5 is cheaper at $5/$25 per million input/output tokens, and Anthropic recommends starting most workloads there.
Does Claude Fable 5.1 have a 1M-token context window?
Yes. The documented context window is one million tokens and maximum output is 128,000 tokens. Large context still requires retrieval, permission controls, state management and testing; it does not guarantee long-task reliability by itself.
Why can Fable 5.1 cost more even with cheaper cache reads?
Cache reads are 75% cheaper than Fable 5, but fresh input and output keep their separate rates. If a task produces substantially more output, misses the cache or retries, total cost can rise. Artificial Analysis observed a higher estimated task cost at maximum effort in its evaluation.
Is Mythos 5.1 available to individuals?
Not as a general self-serve model. Anthropic describes Mythos 5.1 as trusted access for vetted organizations conducting legitimate cybersecurity and life-sciences work.
Sources and review notes
- Product facts, pricing, benchmarks and science examples: and
- Safeguard evaluations and limitations:
- Independent benchmark and cost evidence:
- Launch-day individual experiment:
- Release context: , and
This article was researched and reviewed on September 2, 2026. Launch specifications can change; verify current pricing, availability and beta headers in Anthropic's documentation before deployment.