{"slug": "anthropic-launches-claude-fable-5-1-with-cheaper-cached-inputs-and-new-migration", "title": "Anthropic launches Claude Fable 5.1 with cheaper cached inputs and new migration requirements", "summary": "Anthropic released Claude Fable 5.1 and restricted-access Claude Mythos 5.1 on September 1, cutting cache-read pricing to $0.25 per million tokens, a 75% reduction from Fable 5, while keeping input and output prices at $10 and $50 per million tokens. The company reports Fable 5.1 scored 52.6% on Terminal-Bench-Science 0.1 versus 24.7% for Fable 5, but notes a standard error of roughly 3.5 to 4.5 points, and the release introduces breaking changes for developers, including forced tool use and edited conversation histories.", "body_md": "# Anthropic launches Claude Fable 5.1 with cheaper cached inputs and new migration requirements\n\n- Claude Fable 5.1 keeps input and output pricing at $10 and $50 per million tokens, while cutting cache-read pricing to $0.25 per million tokens, down 75% from Fable 5. <sup>[\\[1\\]](https://www.anthropic.com/claude/fable)</sup>\n- Anthropic reports a 52.6% score on Terminal-Bench-Science 0.1, compared with 24.7% for Fable 5, but says the result carries a standard error of roughly 3.5 to 4.5 points. <sup>[\\[2\\]](https://www.anthropic.com/claude/fable)</sup>\n- Fable 5.1 is generally available through Anthropic, Amazon Web Services, Google Cloud and Microsoft Foundry; Mythos 5.1 remains limited to vetted organizations, currently in the United States. <sup>[\\[3\\]](https://www.anthropic.com/claude/mythos)</sup>\n- The migration includes breaking changes involving forced tool use, thinking blocks and edited conversation histories. <sup>[\\[4\\]](https://platform.claude.com/docs/en/models/fable-5-1/overview)</sup>\n\nAnthropic released Claude Fable 5.1 and the restricted-access Claude Mythos 5.1 on September 1, pairing claimed improvements in long-running coding and agentic work with a substantial cut in the price of cached inputs. The company kept Fable 5.1’s standard rates at $10 per million input tokens and $50 per million output tokens, but reduced cache reads to $0.25 per million tokens. Anthropic estimates that the change lowers the cost of typical workloads by about 25% and highly agentic workloads by as much as 45%. [\\[1\\]](https://www.anthropic.com/claude/fable)\n\nThe release also changes how developers handle safety interventions and model state. Fable 5.1 can identify software vulnerabilities in source code, Anthropic says, but it blocks penetration testing, exploit generation and binary-based vulnerability scanning. Queries that trigger cybersecurity or biology safeguards can be routed to Opus models instead. [\\[3\\]](https://www.anthropic.com/claude/mythos)\n\n## The largest reported gains are on long-running tasks\n\nAnthropic’s comparison table shows Fable 5.1 scoring 52.6% on Terminal-Bench-Science 0.1, versus 24.7% for Fable 5 and 29.0% for Opus 5. On Terminal-Bench 4.0, an agentic coding evaluation, Fable 5.1 scored 55.8%, compared with 42.0% for Fable 5 and 52.3% for Opus 5. The company also reports improvements on AutomationBench, GDPval-AA v2, OSWorld 2.0 and CursorBench. [\\[2\\]](https://www.anthropic.com/claude/fable)\n\nThose figures are vendor-run launch evaluations, not independent post-release tests. Anthropic says the Terminal-Bench-Science results have a standard error of approximately 3.5 to 4.5 points per model. It also used the August 2026 release of OSWorld 2.0, whose task files differ from earlier versions, making direct comparisons with older published results inappropriate. [\\[2\\]](https://www.anthropic.com/claude/fable)\n\nA third-party review of the published table found that the gains were concentrated in tasks involving tools and long-running work. It described the improvement on CursorBench as comparatively modest, with Fable 5.1 at 73.4% versus 70.5% for Fable 5. <sup>[\\[5\\]](https://www.vellum.ai/blog/claude-fable-5-1-mythos-5-1-benchmarks-explained)</sup> No independent evaluation of Fable 5.1 itself was identified in the sources reviewed for this report. A June study of the preceding Fable 5 model found that sustained automated jailbreak attacks could still produce harmful completions, including 702 panel-confirmed cases across its test set. That study does not establish the security performance of Fable 5.1, whose safeguards Anthropic says were retested before launch. [\\[6\\]](https://arxiv.org/abs/2606.18193)\n\n## Same model, different access rules\n\nFable 5.1 and Mythos 5.1 use the same underlying model. Fable is the generally available version with stronger restrictions in cybersecurity, biology and chemistry. Mythos 5.1 is available through Anthropic’s trusted-access programs for vetted cyberdefenders and life-sciences organizations; the company says access is currently limited to U.S. organizations. [\\[3\\]](https://www.anthropic.com/claude/mythos)\n\nAnthropic says Fable 5.1’s biology safeguards intervene on benign requests 85% less often than the safeguards introduced with Fable 5. It still routes dual-use biology and chemistry questions to Opus models. Mythos 5.1 has reduced safeguards for approved research programs, but Anthropic requires a default 30-day retention period for safety monitoring. [\\[3\\]](https://www.anthropic.com/claude/mythos)\n\nFor Fable, 30-day retention is also the default. Enterprise customers eligible for Anthropic’s Enterprise Frontier Safeguards can keep data on infrastructure they control and conduct human review themselves. Until that program is available to an eligible customer, Anthropic says those customers may use Fable 5.1 with zero data retention. [\\[1\\]](https://www.anthropic.com/claude/fable)\n\n## The cheaper workload depends on caching\n\nThe headline discount comes from cache reads rather than lower base token rates. Fable 5.1’s input and output prices remain unchanged from Fable 5, while cache reads fall from $1 to $0.25 per million tokens. Cache writes remain $12.50 per million tokens for five-minute storage and $20 per million tokens for one-hour storage. Anthropic also offers a 50% discount on input and output through its Batch API. [\\[4\\]](https://platform.claude.com/docs/en/models/fable-5-1/overview)\n\nThe model has a one-million-token context window, a maximum output of 128,000 tokens and adaptive thinking that is always enabled. Anthropic’s documentation lists a June 2026 knowledge cutoff and describes Fable 5.1 as slower than Opus 5, which costs half as much per input and output token. [\\[4\\]](https://platform.claude.com/docs/en/models/fable-5-1/overview)\n\nThat pricing structure favors applications that repeatedly send the same long context, such as coding agents, research workflows and document-processing systems. The actual savings will depend on cache-hit rates, output volume, tool calls and how often an agent has to restart work. Anthropic’s 25% and 45% estimates are workload estimates rather than across-the-board reductions. [\\[1\\]](https://www.anthropic.com/claude/fable)\n\n## Developers face several migration changes\n\nAnthropic’s migration documentation lists three breaking changes for applications moving from Fable 5. Forced tool use now returns an error, earlier models cannot read Fable 5.1’s thinking blocks, and editing earlier turns invalidates later thinking blocks. The release adds per-message effort controls in beta, turn-scoped system messages, readable progress updates between tool calls and content-provenance features. [\\[4\\]](https://platform.claude.com/docs/en/models/fable-5-1/overview)\n\nDevelopers can call the model as `claude-fable-5-1` through Anthropic’s API. The model is also available through Amazon Bedrock, Google Cloud and Microsoft Foundry. Anthropic lists Fable 5.1 as active and says it will not retire the model before September 1, 2027. [\\[4\\]](https://platform.claude.com/docs/en/models/fable-5-1/overview)\n\nThe launch gives enterprise buyers a cheaper route to repeated, long-context agent work, but it also shifts evaluation work onto customers. Teams will need to test cache economics, fallback behavior, safety refusals and thinking-block handling inside their own tools before replacing Fable 5 in production.\n\n## Companies mentioned\n\n## Further sources\n\nThe stories that matter, in one email. Free — unsubscribe anytime.", "url": "https://wpnews.pro/news/anthropic-launches-claude-fable-5-1-with-cheaper-cached-inputs-and-new-migration", "canonical_source": "https://mlq.ai/news/anthropic-launches-claude-fable-51-with-cheaper-cached-inputs-and-new-migration-requirements/", "published_at": "2026-09-04 16:37:59+00:00", "updated_at": "2026-09-07 02:02:24.632948+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-products", "ai-policy"], "entities": ["Anthropic", "Claude Fable 5.1", "Claude Mythos 5.1", "Amazon Web Services", "Google Cloud", "Microsoft Foundry", "Terminal-Bench-Science 0.1", "OSWorld 2.0"], "alternates": {"html": "https://wpnews.pro/news/anthropic-launches-claude-fable-5-1-with-cheaper-cached-inputs-and-new-migration", "markdown": "https://wpnews.pro/news/anthropic-launches-claude-fable-5-1-with-cheaper-cached-inputs-and-new-migration.md", "text": "https://wpnews.pro/news/anthropic-launches-claude-fable-5-1-with-cheaper-cached-inputs-and-new-migration.txt", "jsonld": "https://wpnews.pro/news/anthropic-launches-claude-fable-5-1-with-cheaper-cached-inputs-and-new-migration.jsonld"}}