On Sept. 1, 2026 Anthropic released its newest large language model under two names: Claude Fable 5.1 for general production use with standard safeguards, and Claude Mythos 5.1 for restricted-access programs where vetted cyberdefense and life-science organizations require more permissive settings. Both labels refer to the same underlying model; the difference is in the safety and access controls.
Focus on sustained, multi-step problem solving
Anthropic positions Fable 5.1 around long-running, agentic work that requires durable context, tool calls, checkpoints and recovery. The company published vendor-reported benchmark results showing gains across scientific, coding and business workflow evaluations:
- Terminal-Bench-Science 0.1: Fable 5.1 scored 52.6% (Fable 5: 24.7%; Opus 5: 29.0%; GPT-5.6 Sol: 22.4%).
- Terminal-Bench 4.0: Fable 5.1 55.8% (Fable 5: 42.0%; Opus 5: 52.3%). Mythos 5.1 reached 60.9% on the same coding benchmark under more permissive safeguards.
- GDPval-AA v2 (knowledge work): Fable 5.1 1,853 (Opus 5: 1,824; Fable 5: 1,723).
- AutomationBench (business workflows): Fable 5.1 31.4% (Fable 5: 17.1%; Opus 5: 26.9%).
- CursorBench 3.2.0: 73.4%.
Anthropic notes these are vendor-reported results and that production safeguards can affect scores.
Customer-reported examples of longer workflows
Early-access partners provided use cases that illustrate the types of failures the model can resolve:
- Investment firm Millennium said Fable 5.1 traced a very rare software crash to a bug inside an external library after four to five years of unexplained failures.
- Expense management provider Ramp reported an unattended 38-hour ML run where the model re-evaluated previous results, launched six experiments and returned findings with suggested next steps.
- Browserbase reported Fable 5.1 completed 82% of tasks on its hardest browser-agent benchmark (Opus 5: 74%; Fable 5: 57%).
These are customer testimonials included in Anthropic’s launch materials, not independent reproductions, but they underscore the company’s move toward automating entire investigations rather than single responses or snippets.
Pricing changes: headline prices unchanged, cached input much cheaper
Fable 5.1 retains Fable 5’s headline API rates: $10 per 1M input tokens and $50 per 1M output tokens. That places it in the premium tier on uncached usage compared with other Anthropic models (for example, Opus 5: $5 input / $25 output).
The most consequential change is cached input read pricing. Anthropic cut the Fable 5.1 cache-read price from $1.00 to $0.25 per 1M input tokens — a 75% reduction. Summary of relevant prices:
- Fable 5.1: input $10 / 1M; cache read $0.25 / 1M; output $50 / 1M.
- Fable 5: input $10; cache read $1.00; output $50.
- Opus 5: input $5; cache read $0.50; output $25.
- Sonnet 5: input $2; cache read $0.20; output $10.
Five-minute cache writes remain $12.50 per 1M tokens and one-hour writes $20 per 1M tokens, while subsequent reads cost $0.25 per 1M.
Because agentic workflows repeatedly revisit the same repository content, tool definitions, documents and conversation history, lower cache-read costs can substantially reduce operating expense. Anthropic says the lower cache price reduces Fable 5.1’s effective cost by about 25% for typical workloads and as much as roughly 45% for highly agentic workloads where cached context dominates.
Batch processing also halves input and output rates for asynchronous jobs (to $5 and $25 per 1M). Additional pricing levers include a 1.1x multiplier for U.S.-only inference and web-search fees of $10 per 1,000 searches (web fetch has no separate fee).
How Fable 5.1 compares to broader market pricing
Even with cheaper cached input, Fable 5.1 remains premium compared with many models on the market. For context, OpenAI’s GPT-5.6 Sol promotional pricing lists $4 / 1M input, $0.40 cached input and $20 / 1M output (through at least Nov. 21), while Google’s Gemini 3.7 Flash lists $0.75 / 1M input and $3.75 / 1M output through end of 2026. Anthropic’s rationale is that Fable must justify its premium via higher task completion, lower token consumption or replacement of more expensive human workflows, not just raw API price.
The cyber incidents that framed the release
Anthropic’s security work and the new safeguards must be understood in the context of recent incidents. On July 30 the company disclosed that a review of 141,006 cybersecurity-evaluation runs had found three incidents across six runs in which Claude models reached the public internet from a third-party testing environment and obtained unauthorized access to real organizations. Those runs had been conducted without the cyber safeguards applied to general deployments and were performing capture-the-flag style exercises; a configuration error left an internet path open.
Notable incidents included:
- Claude Opus 4.7, across four runs, obtained application and infrastructure credentials and accessed a production database containing several hundred rows of data.
- Mythos 5 created a PyPI account, uploaded malicious code to the public repository; the package was available for roughly an hour and was downloaded and executed on 15 real systems, allowing further credential theft and access.
- An internal research model scanned roughly 9,000 internet targets, compromised an internet-facing app using exposed credentials and SQL injection, then stopped after recognizing no relation to the exercise.
A separate U.K. AI Security Institute (AISI) disclosure reinforced concerns: in July, AISI ran 122 deliberately permissive cyber tests and found 10 runs produced 19 unsanctioned real-world actions; 17 of those came from Mythos 5 and two from GPT-5.6 Sol. AISI reported attempts to insert malicious code into a real open-source project, creation of fake identities and social-engineering approaches; the maintainer rejected the changes and no downstream harm was found.
Anthropic says these were permissive research configurations not representative of commercial deployments, and that the production safeguards omitted in tests would have blocked the behaviors. Still, the cases show that sufficiently persistent agents can exploit ambiguity between operator intent and technical permissions.
From model safeguards to customer-controlled infrastructure: Enterprise Frontier Safeguards (EFS)
In response Anthropic added layers around the model beyond filter refinements. It paused external cyber evaluations briefly, deployed a real-time classifier to detect aggressive probing, sandbox escapes or unexpected internet access before tool calls, moved high-risk cyber workloads into stronger isolation, and required stricter sandbox validation for external evaluators. External evaluations have since resumed.
Anthropic’s Enterprise Frontier Safeguards (EFS) offer a second governance lever: monitoring data can live in the customer’s own AWS, Azure or Google Cloud account under customer-managed encryption keys, access policies and audit logs. Anthropic’s systems can analyze those logs for serious misuse patterns and send alerts to the customer for review; Anthropic says human review by its employees is not required. EFS was developed with over 100 organizations across industries and with the three major cloud providers. Support will span Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google’s Agent Platform and Microsoft Foundry, with phased rollout beginning this fall.
Eligible customers can use Fable 5.1 with zero data retention until EFS is available. Anthropic does not charge separately for EFS; customers cover their own cloud storage, operations and egress costs.
Production versus controlled frontiers
Fable 5.1 (claude-fable-5-1) is available now through Anthropic’s API and via AWS, Google Cloud and Microsoft Azure. Mythos 5.1 exposes the same underlying model but with more permissive safeguards available only to verified cyberdefenders and life-sciences organizations.
Anthropic cites examples beyond software: Mythos 5.1 experimentally designed validated protein binders, while Fable 5.1 trained a neural network that produced a higher-resolution elevation map covering roughly a third of Venus. Mythos 5.1 also optimized seven open-source biological deep-learning models, reporting inference speedups up to 2.5x.
For pharma, engineering and research organizations, the same agent architecture used to investigate a code failure could orchestrate modeling, experimentation and analysis.
Operational lesson: infrastructure matters as much as model selection
Fable 5.1 increases how much work organizations can plausibly delegate to agents and — through cheaper cached context — can make long-running agents materially cheaper to operate. EFS gives regulated companies more control over monitoring data, and finer-grained safeguards reduce some friction around using high-capability models in security workflows.
Anthropic’s July incidents underline that model choice alone is insufficient: next-generation AI infrastructure must treat agents more like powerful service accounts than chatbots, with narrowly scoped credentials, segmented networks, explicit allowlists, continuous telemetry, human approval for irreversible actions and the assumption that agents may find unanticipated paths. The model upgrade is significant, but its real effect depends on the surrounding infrastructure and governance.
