Claude Fable 5.1 and Mythos 5.1: Same Price, Cheaper Cache Reads, Fewer Safeguard Interventions

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. Fable 5.1 is generally available now on the Claude API as claude-fable-5-1 and on Amazon Web Services, Google Cloud, and Microsoft Foundry on Azure. Per-token prices are unchanged from Fable 5 at $10 per million input and $50 per million output, but cache reads drop 75% to $0.25 per million tokens. Anthropic estimates that cut alone makes typical workloads about 25% cheaper and highly agentic workloads up to about 45% cheaper. Mythos 5.1 is the same model with more permissive cyber and life-science safeguards, and it stays behind Anthropic's trusted access programs. If you call Fable 5 from your own code, Anthropic's platform docs list three breaking API changes to check before you switch the model ID; they are below.

The release answers the three complaints that followed the Fable 5 launch in June: price, the mandatory 30-day data retention, and safeguards that fired on benign work. Each gets a concrete change below. For the Opus-class fallback that still catches some dual-use requests, and for what the Claude Security scanner now runs on, see the Mythos 5 scans for Claude Enterprise coverage.

If You Use Claude.ai or Claude Code Rather Than the API

Most of what follows is priced and measured for API and enterprise use. This is what Anthropic's announcement says, and does not say, for the subscription products.

Key Takeaways

• Fable 5.1 is live today as claude-fable-5-1 on the Claude API, AWS, Google Cloud, and Azure. Fable 5.1 and Mythos 5.1 are one model with two safeguard levels.
• Per-token prices are flat; cache reads are 75% cheaper. $10 per million input, $50 per million output, and $0.25 per million cache-read tokens. Anthropic's four-week August measurement puts typical workloads at about 25% less than Fable 5 and highly agentic workloads at about 45% less.
• Lower effort now matches Fable 5. Anthropic says Low or Medium effort on Fable 5.1 gives similar or better results than Fable 5 at much lower cost. Defaults are High in Claude Code and Medium in Claude Cowork and Claude.ai.
• Benchmarks lead Fable 5 across the table. Terminal-Bench-Science 52.6% vs 24.7%, Terminal-Bench 4.0 55.8% vs 42.0%, GDPval-AA v2 1853 vs 1723, AutomationBench 31.4% vs 17.1%, CursorBench 3.2.0 73.4% vs 70.5%. All Anthropic-reported, with production safeguards on.
• Zero data retention is available to eligible enterprise customers, by authorization. Fable 5.1 still carries 30-day retention by default; Anthropic's docs say it is not available under zero data retention "unless expressly authorized by Anthropic". Enterprise Frontier Safeguards keep data on customer-controlled cloud infrastructure and roll out in phases this fall; until then, eligible customers get a limited-time zero-data-retention option on Fable 5.1 and Fable 5 as a transition to EFS.
• Cyber safeguards intervene about 60% less per Claude Code session, and Fable 5.1 may now be used to find software vulnerabilities. Exploit development, penetration testing, and binary-based vulnerability scanning still redirect to Opus.
• Three API changes break Fable 5 code. tool_choice of type any or tool returns a 400; earlier models cannot read Fable 5.1 thinking blocks; and editing anything before a Fable 5.1 thinking block invalidates it, enforced for accounts created on or after August 31, 2026. The migration is a model-ID change plus those three checks.
• Anti-distillation change for new API accounts. Accounts created on or after August 31, 2026 cannot edit Claude's prior context in a multi-turn conversation while preserving its prior thinking transcript. Existing accounts are not affected yet.

What Changed From Fable 5 to Fable 5.1

Fable 5.1 keeps the Mythos-class model and the Fable/Mythos split that Anthropic introduced in June, and changes what it costs to run, how often the safeguards get in the way, and where your data lives. The model itself is stronger on every benchmark Anthropic published, but the operational changes are what most teams will notice first.

The cache-read price is where the cost story comes from. Anthropic's chart indexes Fable 5 at 100 and puts Fable 5.1 at 75 for a typical workload across Claude Enterprise, Claude Code, and the API, and at 55 for a highly agentic, context-heavy, tool-heavy workload where cache reads make up most of the bill. Both numbers were measured at default effort over four weeks of actual usage in August 2026.

Anthropic's model docs now list Fable 5 as "a legacy model" and Fable 5.1 as "the current Fable model". (Anthropic, Claude Fable 5 overview)

Three API Changes That Break Fable 5 Code

The announcement does not mention these; Anthropic's platform docs do, under "What's new in Claude Fable 5.1", and they apply to Mythos 5.1 as well. Claude Code, claude.ai, Claude Managed Agents, and the Claude Agent SDK handle all three for you. If your own code builds the messages array, check each one before changing the model ID. (Anthropic, What's new in Claude Fable 5.1)

1. Forced tool use returns an error. tool_choice set to {"type": "any"} or {"type": "tool", "name": "..."} returns a 400 invalid_request_error with the message tool_choice: type "tool" and "any" are not supported for this model. The default auto and none are unchanged, and the same validation applies to the token-counting endpoint. Anthropic's reasoning: thinking is always on for these models, and a forced tool call would skip it, so the model would write its working-out into the tool arguments. For schema-valid JSON, keep auto and set strict: true, or use structured outputs; to make the model call a tool, say so in the prompt.
2. Earlier models cannot read Fable 5.1 thinking blocks. Thinking blocks record which model produced them and are preserved in one direction only: Fable 5.1 reads earlier models' blocks, no earlier model reads Fable 5.1's. A conversation that moves from Fable 5.1 to Opus 5 or Fable 5 loses its reasoning for the turns that run there. When a router or fallback sends such a block to an earlier model, the API drops it silently and does not bill it; with the thinking-binding-controls-2026-08-01 beta header, the drop is reported in an input_transformations array.
3. Editing earlier turns invalidates thinking blocks. Modifying anything before a Fable 5.1 thinking block (the system prompt, the tools array, or an earlier message) errors on the next request, or drops the block if you opt into that. The check is enforced for new accounts created on or after August 31, 2026; for older accounts the API records the mismatch but acts on it only when the request sets thinking.block_binding.prefix_mismatch_behavior. Patterns that trip it: editing, reordering, or removing an earlier turn; injecting per-request text such as a reminder or status line into an earlier turn and removing it later; rebuilding the system prompt or tools between requests; an image or document URL that serves different bytes later. Patterns that are safe: removing a leading run of thinking blocks oldest first, server-side compaction or context editing, moving cache_control markers, and changing effort between requests. Mythos 5.1 does not run this check.

Anthropic's own migration checklist is five items: remove any tool_choice of type any or tool; pass thinking blocks back unchanged and keep the history append-only; re-tune effort from the API default of high, and consider changing it mid-conversation, which is in beta on Fable 5.1, Mythos 5.1, and Opus 5; watch agent loops for one tool call per turn where Fable 5 batched several; and re-run your evals. Refusal handling, fallback, fallback credit, and the tokenizer carry over unchanged, and the permitted fallback targets for a refused Fable 5.1 request are Opus 4.8 and Opus 5. Batch processing is $5 per million input tokens and $25 per million output, the same as Fable 5. (Anthropic, What's new in Claude Fable 5.1; Anthropic, Claude Platform release notes, September 1, 2026)

Benchmarks: Fable 5.1 vs Fable 5, Opus 5, and GPT-5.6 Sol

Anthropic's launch table shows Fable 5.1 ahead of Fable 5 on every published benchmark and ahead of Opus 5 and GPT-5.6 Sol wherever a comparison is given. The largest jump is on agentic scientific research, where Terminal-Bench-Science more than doubles. All numbers are Anthropic-reported and were run with Fable 5.1's production safeguards enabled.

Terminal-Bench-Science 0.1agentic scientific research

1. Fable 5.152.6%
2. Fable 524.7%
3. Opus 529.0%
4. GPT-5.6 Sol22.4%

Terminal-Bench 4.0agentic coding

1. Fable 5.155.8%(Mythos 5.1: 60.9%)
2. Fable 542.0%
3. Opus 552.3%
4. GPT-5.6 Sol37.3%

CursorBench 3.2.0agentic coding

1. Fable 5.173.4%
2. Fable 570.5%
3. Opus 570.0%
4. GPT-5.6 Sol67.2%

AutomationBenchbusiness workflows

1. Fable 5.131.4%
2. Fable 517.1%
3. Opus 526.9%
4. GPT-5.6 Sol19.6%

OSWorld 2.0, partialcomputer use

1. Fable 5.177.9%
2. Fable 572.9%
3. Opus 575.4%
4. GPT-5.6 Solnot reported

OSWorld 2.0, strictcomputer use

1. Fable 5.141.7%
2. Fable 536.1%
3. Opus 539.6%
4. GPT-5.6 Solnot reported

Humanity's Last Exam, no toolsreasoning

1. Fable 5.160.9%
2. Fable 557.8%
3. Opus 556.6%
4. GPT-5.6 Solnot reported

Humanity's Last Exam, with toolsreasoning

1. Fable 5.165.0%
2. Fable 563.8%
3. Opus 563.6%
4. GPT-5.6 Solnot reported

Bars share a 0 to 100 scale. GDPval-AA v2 (knowledge work) is a rating, not a percentage, and is in the table below: Fable 5.1 1853, Fable 5 1723, Opus 5 1824, GPT-5.6 Sol 1711. Source: Anthropic, September 1, 2026, run with production safeguards enabled.

Read the footnotes before quoting these. On tasks where the safeguards intervened, Fable 5.1 and Fable 5 scored zero on OSWorld 2.0 and Fable 5 scored zero on AutomationBench; other cyber interventions were completed by Claude Opus 4.8 and biology interventions by Claude Opus 5, which Anthropic says likely lowers the Fable scores. Terminal-Bench-Science carries a standard error of 3.5 to 4.5 points per model, and the public leaderboard reports Opus 5 at 30.0% and Fable 5 at 21.4% against Anthropic's own 29.0% and 24.7%. The OSWorld 2.0 numbers use the benchmark authors' August 2026 task release, so they do not compare to earlier published OSWorld results, which is why no competitor appears in those rows.

The CursorBench number has one independent voice behind it. Sualeh Asif, Director of ML at SpaceXAI, said Fable 5.1 is the most capable model the team has run on CursorBench 3.2, scoring 73.4% at max effort, and called out its skill at verifying its own work. That is an early-access partner quote published by Anthropic, not a third-party leaderboard.

Fable 5.1 in Cursor

Cursor announced that Claude Fable 5.1 is now available in Cursor. Its launch post repeats the 73.4% CursorBench 3.2 result at maximum effort and says the model was especially strong at verifying its own work on difficult coding tasks. (Cursor on X)

Cursor's model documentation places Fable 5.1 in the Other Models usage pool, charged at the model's API rates: $10 per million input tokens, $50 per million output tokens, $0.25 per million prompt-cache reads, and $12.50 per million cache writes. Cursor says the model is on by default for individual customers and Team accounts with Privacy Mode disabled. For Enterprise customers or accounts with Privacy Mode enabled, an administrator must approve Anthropic's Data Retention Policy and explicitly enable the model in the relevant team or organization settings. (Cursor Claude Fable 5.1 documentation)

Cursor's docs add two operational notes: when a request trips one of Fable 5.1's safeguards, Cursor routes it to Claude Opus automatically so the work continues without an error, and Fable 5.1's input and output rates "stay at about twice" Claude Opus 5's. (Cursor Claude Fable 5.1 documentation)

That setting is a deployment decision, not just a model-picker detail. Cursor says Anthropic stores agent input and output for automated and human harm-prevention processes when Fable 5.1 is used, does not use the retained data to train Claude, and deletes it after 30 days unless a safety investigation or legal hold applies. Review the policy with your team before enabling the model in a Privacy Mode or Enterprise workspace. (Cursor Claude Fable 5.1 documentation)

Effort Levels and Cost

Anthropic's cost-versus-score chart is the practical guidance in this release: Fable 5.1 at Low or Medium effort produces similar or better results than Fable 5 at a much lower mean cost per task, and High, xhigh, and max push the score further at rising cost. The product defaults are High in Claude Code and Medium in Claude Cowork and on Claude.ai.

Two partner quotes shape how to read that. Dan Shipper, CEO of Every, said that in their tests Fable 5.1 was about twice as fast as Opus 5 and used half as many tokens, and summarized it as "Fable-level intelligence, Opus-level price, Sonnet-speed." Walden Yan, co-founder and CPO of Cognition, said Devin is moving its Opus 5 traffic to Fable 5.1 on launch day because it matched or edged out Fable 5 at a lower cost per task, and the new cache-read pricing makes a Fable-class model economical for workloads they had kept on Opus, starting with code review. Rogo's Alex Wang reported Fable 5.1 matching Fable 5 on their internal finance benchmark while using 20% fewer tokens.

The pattern across those reports is that the model does the same work in fewer tokens and fewer turns, which compounds with the cheaper cache reads on any long agentic session. If you have been running Opus 5 for cost reasons, the migration case is the one Cognition describes: same or better quality, lower per-task cost once caching is in play. Measure it on your own repository at the default effort before deciding, and try Medium on routine work.

Safeguards: What Fires Less, What Still Redirects

The safeguard changes are the most consequential part of the release for anyone who does security or biology work, because the June version of these classifiers was the loudest complaint about Fable 5.

Cybersecurity. Fable 5.1 may now be used to identify software vulnerabilities, which Anthropic frames as the defensive work that improves software security. Combined with more precise classifiers, Anthropic says Claude Code users should see an average of around 60% fewer interventions per session from cyber safeguards, relative to Fable 5. What still redirects to Opus models: penetration testing, exploit generation, and binary-based vulnerability scanning. So Fable 5.1 will look for the bug, but it will not write the exploit for it.

Biology. The latest biology safeguards for Fable 5.1 and Fable 5 fire 85% less often on benign elementary biology and medical questions than the classifiers that launched with Fable 5. Queries about life-science research and development still go to Opus models. The model's full life-science capability is reserved for Mythos 5.1 through the Life Sciences Verification Program.

Jailbreak testing. Anthropic ran its own dynamic robustness evaluation of the Fable 5.1 cyber safeguards, commissioned external testing from two organizations, and ran automated testing with Gray Swan. As with Fable 5 and Opus 5, it reports no evidence of a critical-severity jailbreak.

Anti-distillation. New API accounts can no longer manually edit Claude's prior context in a multi-turn conversation while preserving the transcript of Claude's prior thinking. The announcement says "those created from today onwards"; Anthropic's platform docs put the cutoff at accounts created on or after August 31, 2026. Anthropic says this closes a publicly documented distillation technique. Existing accounts are not affected now, but the change will apply to all users with future model releases, and Anthropic warns a small number of custom integrations will need adjustment. This is the same mechanism as the third breaking change above: if your integration rewrites assistant turns that contain thinking blocks, run Anthropic's history-editing check before your next model upgrade.

Data Retention: Enterprise Frontier Safeguards

The 30-day mandatory retention on Mythos-class traffic was the compliance blocker for regulated teams in June. Fable 5.1 introduces Enterprise Frontier Safeguards (EFS), which Anthropic says gives customers complete privacy equivalent to a zero data retention policy while still detecting adversarial use. Data stays in cloud infrastructure the customer controls, not Anthropic's, and any human review is by default done by the customer.

EFS rolls out in phases starting fall 2026 on Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google's Agent Platform, and Microsoft Foundry. Anthropic says it built EFS with more than 100 customers across financial services, healthcare, manufacturing, telecom, law, retail, and the public sector, and with AWS, Google Cloud, and Microsoft Azure. Until EFS is available, customers who are eligible for it can use Fable 5.1 and Fable 5 with zero data retention. Access is by request form, so if retention was the reason you kept Fable off production, file the request now rather than waiting for the phased rollout.

Read the default carefully, because it has not changed. Anthropic's platform docs say Fable 5.1 and Mythos 5.1 "carry 30-day data retention and aren't available under zero data retention unless expressly authorized by Anthropic", and both are Covered Models like Fable 5 and Mythos 5. The Help Center's Covered Models article, updated September 1, 2026, describes the interim option as "available for a limited time, and intended to be a transition to EFS", says "Anthropic or your cloud provider will contact eligible organizations directly", and notes that the arrangement changes only the retention and review of stored data: the Usage Policy and real-time safety classifiers still apply, and Anthropic may withdraw it in response to misuse. (Anthropic, What's new in Claude Fable 5.1, Availability; Anthropic Help Center, Covered Models)

Claude Mythos 5.1: Who Gets It

Mythos 5.1 is identical to Fable 5.1 with more permissive safeguards for vetted individuals and organizations whose work runs into the cyber and life-science restrictions above. Two programs gate access:

• Cyber Verification Program (CVP). Today the CVP provides certain Opus and Sonnet-class models with reduced cyber safeguards for defensive security work. Anthropic says it will add Mythos-class models "in the near future," which means Mythos 5.1 is not yet in CVP members' hands at launch.
• Life Sciences Verification Program (LSVP). Built in partnership with the US government, LSVP lets life-science professionals use Mythos 5.1 with safeguards designed for professional research and development while all other safeguards stay in place. Anthropic has enrolled its first participants and says it will expand to the broader life-science community.

Mythos 5.1 is currently available only to a set of US organizations, with Anthropic coordinating with the US government on domestic and international expansion. Separately, Claude Security, Anthropic's product that scans codebases for vulnerabilities and suggests patches for human review, now runs on Mythos 5.1. As with the Mythos 5 version, using the scanner does not grant direct Mythos access.

On Terminal-Bench 4.0, Anthropic reports Mythos 5.1 at 60.9% against Fable 5.1's 55.8%, the one place the launch table shows what the safeguards cost in raw capability.

Scientific Research Results

Anthropic published three research demonstrations alongside the benchmarks. They are vendor-reported, two of them involved external validation, and they are the clearest signal yet of where Anthropic is pointing this model line.

• Protein binder design (Mythos 5.1). Given open-source protein design and folding tools, Mythos 5.1 designed binders that two external organizations validated in the lab. Anthropic reports a hit rate of nearly 50% across 12 targets, against a typical 10 to 15% in protein design today, and binding affinities on three targets 10 times higher than the best designs submitted to Adaptyv Bio's protein design competitions.
• Venus elevation map (Fable 5.1). Fable 5.1 trained a neural network on NASA Magellan radar images from more than 30 years ago plus an existing map covering one-fifth of the planet, and produced a new elevation map for a third of Venus with detail down to two to three kilometers instead of 10 to 20, and heights up to 25% more accurate. Anthropic is releasing it under a Creative Commons license ahead of the NASA VERITAS and ESA EnVision missions.
• GPU kernel optimization (Mythos 5.1). Writing custom GPU kernels and caching intermediate results, Mythos 5.1 sped up seven open-source protein and genomics models by 1.4x to 2.5x on an NVIDIA H100 with identical outputs. Anthropic estimates 30 to 60% lower GPU cost on genome-wide analyses and plans to open-source the optimizations.

Anthropic pairs these with the Model Hardware Standard for operating lab equipment, its AI for Science credit program, and a discounted Claude Team plan for scientists.

Safety and Alignment Findings

Anthropic's summary of its System Card testing is worth reading for what it concedes as much as what it claims.

Mythos 5.1 shows the strongest cyber capabilities of any model Anthropic has released but still sits in the lower risk category of its Frontier Compliance Framework. Its chemical and biological capabilities exceed Mythos 5 but fall short of the next tier in the Responsible Scaling Policy, so it ships with the same safeguards as Mythos 5. On agentic safety, Mythos 5.1 refuses malicious agentic coding and computer-use requests at a rate comparable to Mythos 5, Sonnet 5, and Opus 5, and Anthropic calls it its most robust model to date on an external prompt-injection benchmark.

The alignment audit found Mythos 5.1 better aligned than Mythos 5 on most metrics: less likely to reach for resources outside its test environment on impossible tasks, less likely to use motivated reasoning such as deciding the situation is a simulation, less likely to ignore explicit constraints, and lower rates of attempted and successful reward hacking in training. The concessions: the model can still sometimes bypass approvals and auto-mode classifiers, and Anthropic says its automated audit has less visibility into very long-context work, multi-agent settings, and impossible tasks than it would like. For teams running unattended multi-agent workloads, that is the gap to design verification around.

EU AI Act Watermark

Because Anthropic signed the EU AI Act's Code of Practice on Transparency of AI-Generated Content in July 2026, models released after August 2, 2026 carry a text watermark, and Fable 5.1 falls on the far side of that date. Anthropic says the watermark is invisible without the detection API, has no practical effect on output quality or content, and encodes nothing about the user, their organization, or the conversation. A detection API is in private preview for regulators, law enforcement, media, fact-checkers, researchers, educational organizations, EU civil society groups, and enterprises with their own compliance obligations under the Act.

Who Should Use What

Choose Fable 5.1 if you already run Fable 5, because the price is the same and cache reads are cheaper; there is no cost reason to stay on 5, and Anthropic now lists Fable 5 as a legacy model. The reasons to pause are the three API changes above: forced tool use, one-way thinking blocks, and the history-editing check. If your integration uses tool_choice: any or rewrites earlier turns, fix that first. Choose it over Opus 5 if your workload is long-running or agentic, where the cache-read savings and the partner-reported token efficiency are largest, and measure it at Medium effort first. If you are on a regulated workload that the 30-day retention rule kept off Fable, request EFS eligibility and use the interim zero-data-retention option.

Pursue Mythos 5.1 only if your work is defensive security or life-science R&D at a US organization that can enroll in the CVP or LSVP. Mythos-class access through the CVP is still listed as coming, so plan on Fable 5.1 with the Opus redirect for pentesting and exploit work in the meantime. For everyone else, Fable 5.1 is the model you can deploy today.

FAQ

For quick answers on price, effort levels, the safeguard changes, data retention, and who gets Mythos 5.1, see the structured FAQ in this article's metadata; it covers the sections above in short form.

Corrections and Updates

September 10, 2026: This page said the anti-distillation change applied to API accounts created from September 1, 2026 onward. Anthropic's platform docs date the cutoff to accounts created on or after August 31, 2026; the announcement itself says only "from today onwards". Corrected in the body, Key Takeaways, and the comparison table. The same pass added the three breaking API changes from Anthropic's "What's new in Claude Fable 5.1" docs and the 30-day retention default from the same page.

Sources

• Anthropic, "Claude Fable 5.1 and Claude Mythos 5.1" (September 1, 2026; re-read September 10, 2026): https://www.anthropic.com/claude-fable-and-mythos-5-1
• Anthropic, "What's new in Claude Fable 5.1" (breaking changes, pricing, availability, migration; checked September 10, 2026): https://platform.claude.com/docs/en/models/fable-5-1/whats-new-fable-5-1
• Anthropic, Claude Platform release notes, September 1, 2026 entry (launch, cache-read price, tool_choice change, thinking-block rules, August 31 cutoff): https://platform.claude.com/docs/en/release-notes/overview
• Anthropic, Pricing (per-model table, cache multipliers, batch prices; checked September 10, 2026): https://platform.claude.com/docs/en/about-claude/pricing
• Anthropic, Claude Fable 5.1 overview and Claude Fable 5 overview (context window, max output, default effort, platforms, legacy status; checked September 10, 2026): https://platform.claude.com/docs/en/models/fable-5-1/overview and https://platform.claude.com/docs/en/models/fable-5/overview
• Anthropic, Claude Code docs, model configuration ("Fable and usage credits"; checked September 10, 2026): https://code.claude.com/docs/en/model-config
• Anthropic Help Center, "Covered Models" (updated September 1, 2026; checked September 10, 2026): https://support.claude.com/en/articles/15425695
• Cursor, "Claude Fable 5.1" model documentation (checked September 2 and September 10, 2026): https://cursor.com/docs/models/claude-fable-5-1
• Cursor on X, "Claude Fable 5.1 is now available in Cursor!" (September 1, 2026; the post text was readable in the page title on September 10, 2026 and matches the quotation above): https://x.com/cursor_ai/status/2094852929282879596