Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. The two names describe the same underlying model capabilities with different safeguards and access rules: Fable 5.1 is generally available, while Mythos 5.1 is reserved for vetted cybersecurity and life-sciences users.
The release is aimed at long-running coding, research, and knowledge-work agents. Its headline savings come from a 75% cut in the price of cached input reads—not from lower standard token rates. That distinction matters because independent tests found that Fable 5.1 can still cost more per completed task when it generates more output. Teams should compare finished-work cost and success rate on their own tasks before moving production traffic.
What is Claude Fable 5.1?
Claude Fable 5.1 is Anthropic's generally available model for demanding reasoning and long-horizon agentic work. It supports a 1-million-token context window, up to 128,000 output tokens, text and image input, adaptive thinking, and five effort levels. Standard API pricing is $10 per million input tokens and $50 per million output tokens.
Anthropic's own model-selection guidance says most workloads should start with Claude Opus 5, which costs half as much at standard token rates. Fable 5.1 is the escalation option when higher-effort Opus 5 evaluations still miss the target, or when long, cache-heavy agents can benefit from much cheaper repeated context.
Claude Fable 5.1 specifications and access
| Item | Confirmed detail |
|---|---|
| Release date | September 1, 2026 |
| Claude API model ID | claude-fable-5-1 |
| Context window | 1 million tokens |
| Maximum synchronous output | 128,000 tokens |
| Input and output | Text and images in; text out |
| Thinking | Adaptive thinking is always on; effort controls depth |
| Knowledge cutoff | June 2026 |
| Standard API price | $10 input and $50 output per million tokens |
| Cache-read price | $0.25 per million cached input tokens |
| Platforms | Claude products and API, AWS, Google Cloud, and Microsoft cloud platforms |
Anthropic lists Fable 5.1 as active and says it will not retire before September 1, 2027. Cloud model identifiers vary: Amazon Bedrock uses anthropic.claude-fable-5-1, while Anthropic's API and the other documented platforms use claude-fable-5-1. Confirm the current identifier in the provider's model catalog before changing deployment configuration.
Fable 5.1 versus Mythos 5.1
Fable 5.1 and Mythos 5.1 share model specifications and pricing. The operational difference is the safeguard layer. Fable is intended for general use and can route or restrict sensitive cybersecurity and biology requests. Mythos offers more permissive safeguards to vetted professionals whose legitimate work would otherwise be constrained.
- Choose Fable 5.1 for normal coding, analysis, research, document work, and production agents.
- Consider Mythos 5.1 only if your organization performs qualifying cybersecurity or life-sciences work and can complete Anthropic's trusted-access process.
- Do not treat Mythos as a premium public tier. Anthropic says access is invitation-only through Project Glasswing programs and initially limited to a set of U.S. organizations.
The safer assumption for product planning is that Mythos is unavailable unless Anthropic or a supported cloud account team has explicitly approved the organization.
What changed from Claude Fable 5?
Cheaper cached context
Standard input and output rates remain $10 and $50 per million tokens. Cache reads fell from $1 to $0.25 per million tokens. Anthropic estimates about 25% lower cost for typical billed workloads and as much as roughly 45% for highly agentic workloads where cache reads dominate.
Those are workload estimates, not a universal discount. Artificial Analysis measured Fable 5.1 at max effort costing $3.76 per task in its Intelligence Index, 20% more than Fable 5's $3.14, because 5.1 used about 1.7 times as many output tokens. TECHSY's 14 short matched tasks also showed no blanket saving: all three tested models passed, while Fable 5.1 cost more than Fable 5 and Opus 5 in that small test. The useful metric is therefore cost per accepted result, including retries and review time.
Five effort levels and always-on thinking
Fable 5.1 uses adaptive thinking at low, medium, high, xhigh, or max effort. Anthropic says Claude Code defaults to high, while Claude.ai and Claude Cowork default to medium. Independent measurements show output volume and cost can vary dramatically by effort, so do not leave every request at the default without testing. Use a lower level for bounded tasks that already pass, then raise it for difficult planning, debugging, or research steps.
Breaking API behavior
Three migration details can break an existing agent:
- Forced tool choice is unsupported. Requests using a forced
tooloranychoice return an HTTP 400 error. Keep tool choice on automatic and use strict tool schemas or structured outputs when schema compliance matters. - Thinking blocks are model-bound. Earlier Claude models cannot read thinking blocks created by Fable 5.1. A router that falls back to an older model should remove incompatible thinking blocks.
- Conversation prefixes must remain stable. Editing the system prompt, tool definitions, or earlier messages before a preserved Fable 5.1 thinking block can invalidate the block or cause an error, depending on account and request settings.
Non-default temperature, top-p, or top-k settings also return an error, according to Anthropic's current migration documentation. Re-run integration tests instead of assuming an old Fable configuration remains valid after swapping the model ID.
When should you use Fable 5.1?
Fable 5.1 is a plausible fit for long coding sessions, complex repository work, multistep research, document or spreadsheet agents, and other jobs where a model must preserve a large working context while using tools repeatedly. Anthropic reports large gains on several agentic and science-oriented evaluations, but vendor benchmarks should be treated as evidence about specific test harnesses rather than guaranteed product performance.
Independent evidence is mixed in a useful way. Artificial Analysis placed Fable 5.1 at the top of its Intelligence Index at max effort, while noting higher per-task cost and a higher tendency to attempt answers it did not get right. A hands-on review by Thomas Wiegold reported strong results on a difficult codec experiment and a design task, but clearly labeled the codec sample as too small for production claims. These results support testing the model on hard work; they do not justify replacing every cheaper model.
A practical evaluation before switching
- Create a representative test set. Include routine, difficult, long-horizon, and failure-prone tasks drawn from real work. Remove secrets and personal data.
- Define acceptance criteria before running. Use unit tests, factual source checks, formatting rules, or a human rubric rather than judging which output merely sounds best.
- Compare at matched effort. Test Opus 5 and Fable 5.1 at more than one effort level. Record input, cached input, output, latency, retries, and reviewer time.
- Test tools and history handling. Exercise automatic tool choice, structured output, long conversations, context trimming, model fallback, and error recovery.
- Calculate cost per accepted task. A more expensive call can still be cheaper if it eliminates retries; a cheaper cache line does not help a short, uncached request.
- Use a limited rollout. Send a small percentage of eligible work to Fable 5.1, monitor failures and cost, and preserve an explicit rollback path.
If you are designing an agent rather than choosing only a model, ToolNexa's practical Agents API guide explains broader approval, logging, and least-privilege controls that also apply to model-driven automation.
Safety and privacy checks
- Keep approval gates for consequential actions. Higher benchmark scores do not remove the need to approve deployments, payments, messages, access changes, or security-sensitive actions.
- Expect occasional confident errors. Open cited sources, run tests, and let the model say when evidence is missing.
- Review retention terms. Anthropic's current documentation classifies Fable 5.1 and Mythos 5.1 as covered models with 30-day retention unless Anthropic expressly authorizes another arrangement. Enterprise Frontier Safeguards are scheduled to roll out in phases.
- Use least-privilege tools. Restrict file paths, network destinations, credentials, and write permissions to what each task requires.
- Log model and effort. Save the stable model ID, effort level, tool results, and approval decisions so incidents can be reproduced without storing hidden reasoning.
Frequently asked questions
Is Claude Fable 5.1 available now?
Yes. Anthropic released Fable 5.1 on September 1, 2026 and lists it as generally available across its products, API, and supported cloud platforms. Account, region, and product-plan access can still vary.
How much does Claude Fable 5.1 cost?
The Claude API base price is $10 per million input tokens and $50 per million output tokens. Five-minute cache writes cost $12.50 per million, one-hour cache writes cost $20, and cache reads cost $0.25. Cloud providers and consumer subscriptions may use different billing structures.
Is Claude Mythos 5.1 better than Fable 5.1?
No general capability difference is claimed. Anthropic describes them as the same model with different safeguards. Mythos is restricted to vetted cybersecurity and life-sciences users, while Fable is the generally available version.
Should every Claude user upgrade to Fable 5.1?
No. Anthropic recommends starting most workloads with the cheaper Opus 5. Test Fable 5.1 when demanding or long-running work still fails, or when cache-heavy agents may benefit from its lower cache-read price.
What can break when migrating from Fable 5?
Forced tool choice, cross-model reuse of thinking blocks, edited conversation prefixes, and non-default sampling parameters can cause errors. Update the model ID, keep tool choice automatic, handle thinking blocks carefully, and run regression tests before production rollout.
Conclusion
Claude Fable 5.1 is best treated as a specialized escalation model, not an automatic default. Its large context window, agentic performance, and cheaper cache reads are valuable when work is difficult and repetitive context dominates the bill. Start with a representative evaluation, measure cost per accepted result, verify tool and history behavior, and roll out only where the evidence beats a cheaper alternative.
Sources
- Anthropic: Introducing Claude Fable 5.1 and Claude Mythos 5.1
- Anthropic Platform Docs: Fable 5.1 specifications and pricing
- Anthropic Platform Docs: New features and migration changes
- Artificial Analysis: Independent benchmark and cost analysis
- Thomas Wiegold: Hands-on Fable 5.1 review
- TECHSY: Metered Fable 5.1 comparison