Claude Fable 5.1 review, pricing and limits

Developers who already deploy Fable 5 for long coding and research agents get the natural upgrade path here, since Fable 5.1 now ranks first on MMLU Pro at 92.38% among evaluated frontier models. Anthropic recommends most workloads start with the cheaper Opus 5 instead, reserving Fable 5.1 for tasks where Opus 5 still falls short at high effort.

Claude Fable 5.1 launched September 1, 2026 as Anthropic's follow-up Mythos-class model, topping the Vals Index at 67.87% among 51 evaluated models. It keeps Fable 5's 1 million token context window and newly allows defensive vulnerability discovery that its predecessor blocked outright.

Provider: Anthropic · Family: Claude Fable

More about Anthropic on HokAI

Context window: 1,000,000 tokens · Max output: 128,000

Input modalities: text, image, pdf, tool-calls · Output: text, tool-calls

About Claude Fable 5.1

Claude Fable 5.1 is Anthropic's second Mythos-class model, released September 1, 2026, just under three months after Fable 5. It sits above Claude Opus 5 in Anthropic's lineup for the most demanding reasoning and long-horizon agentic work. Fable 5.1 shares its underlying weights with Claude Mythos 5.1, a configuration with reduced safeguards restricted to vetted cybersecurity and life-sciences organizations through Anthropic's Project Glasswing program. Anthropic has not disclosed the parameter count or architecture type for either model. On SWE-bench Pro, Fable 5.1 scores 81.2%, the top result on BenchLM.ai's September 2026 leaderboard, ahead of Fable 5 at 80.0%, Claude Opus 5 at 79.2%, GPT-5.6 Sol at 64.6%, and Gemini 3.5 Flash at 55.1%. On the Vals Index, a composite benchmark spanning coding, legal, tax, and medical tasks, Fable 5.1 ranks first among 51 evaluated models at 67.87%, just ahead of Opus 5 at 67.21% and Fable 5 at 66.04%. It leads MMLU Pro at 92.38%, LiveCodeBench at 90.52%, and MMMU Pro at 90.64%, and reaches 85.02% on Terminal-Bench 2.1, up from Fable 5's 80.52%. Anthropic's own release comparison table shows Terminal-Bench-Science 0.1 rising to 52.6% from Fable 5's 24.7%, and AutomationBench rising to 31.4% from 17.1%. The model is not perfect everywhere: it scores just 6.67% on Harvey's Legal Agent Benchmark, its weakest published result. The context window holds at 1 million tokens with a 128,000-token maximum output, unchanged from Fable 5. Adaptive thinking stays always-on, with reasoning depth set through the effort parameter, defaulting to high. A new per-message effort control, in beta, lets a running conversation change its reasoning depth partway through a session without losing the prompt cache, a first for the Claude API. Fable 5.1 accepts text and images, including PDFs read as images, and returns text and tool calls; there is no audio input and no video support in either direction. Vision handles charts, screenshots, and scanned documents natively. Native tool support includes function calling, parallel tool calls, and the memory tool. Two more beta features shipped alongside 5.1: turn-scoped system messages, which apply instructions to a single turn instead of the whole conversation, and a progress-update display mode that streams readable notes between tool calls during long agentic runs. What changed most from Fable 5 is prompt-cache pricing: cache-read cost dropped sharply, while per-token input and output pricing and the standard cache-write rates carry over unchanged. Anthropic frames the cache-read cut as meaningfully lowering typical workload cost, with the largest savings going to cache-heavy agentic automation that replays the same context repeatedly. The Batch API keeps its standard discount on input and output, with results still returned asynchronously. Fable 5.1 is generally available on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry on Azure, and Claude Platform on AWS, alongside Claude.ai, Claude Code, and Claude Enterprise. Anthropic commits to keeping it active for at least one year, with retirement not scheduled before September 1, 2027. Three changes break code written for Fable 5 without modification: a forced tool-use parameter now returns an API error instead of executing, thinking blocks Fable 5.1 produces cannot be read back by earlier Claude models, and editing an earlier turn in a conversation invalidates any thinking blocks generated after that point. Fable 5.1 keeps Fable 5's three-classifier safety system but retunes two of them. The cybersecurity classifier now permits the model to flag software vulnerabilities during defensive research, a change from Fable 5's broader block, while it still refuses to generate exploit code or assist offensive hacking. The biology and chemistry classifier fires about 85% less often on ordinary medical questions while keeping its block on high-risk biological research. Every output the model generates carries an invisible statistical content watermark, a requirement for post-August-2026 Anthropic models under the EU AI Act; Anthropic opened a private-preview API letting regulators, law enforcement, media, and researchers check for that watermark. A separate new program, Enterprise Frontier Safeguards, lets qualifying enterprise customers run misuse detection on their own infrastructure instead of Anthropic's, developed with more than 100 customers across finance, healthcare, and other regulated industries. Teams already running Fable 5 in autonomous coding agents gain the SWE-bench Pro and Terminal-Bench improvements at the same base price, plus the cache-read savings. Anthropic's own model-choice guidance says most workloads should start with Opus 5 instead, at a markedly lower price, and move up to Fable 5.1 only when Opus 5 falls short at high effort on a team's own evaluations. Teams needing audio or video input have no path on either model, and regulated teams requiring zero data retention outside the new Enterprise Frontier Safeguards contract should also look elsewhere, since standard access still carries mandatory retention. Knowledge is reliable through June 2026, about five months later than Fable 5's cutoff. Anthropic has not disclosed training data composition for either model. The Responsible Scaling Policy governs deployment, and the September 1, 2026 system card documents red-teaming for the retuned cybersecurity and biosecurity classifiers. SOC 2, HIPAA-eligible, and GDPR-aligned deployments remain available through enterprise contracts on Anthropic-operated and partner cloud platforms. The API model ID is claude-fable-5-1.

Pricing

$10/$50 per 1M input/output tokens, unchanged from Fable 5. Cache reads cut 75% to $0.25/MTok (from $1). 5-minute cache write $12.50/MTok, 1-hour cache write $20/MTok. Batch API: 50% off input and output, results within 24h.

Key Features

• SWE-bench Pro Leaderboard Lead: Tops the SWE-bench Pro leaderboard among evaluated frontier models as of September 2026, ahead of Fable 5, Opus 5, and GPT-5.6 Sol.
• Sharply Cheaper Cache Reads: Prompt-cache read pricing dropped sharply from Fable 5's rate, meaningfully cutting the cost of agentic workflows that reuse the same context many times.
• Defensive Vulnerability Discovery: Security researchers can now ask the model to hunt for bugs in their own codebase, something Fable 5's classifier used to shut down outright regardless of intent.
• Content Provenance Watermarking: Every output carries an invisible statistical watermark for EU AI Act compliance, and Anthropic opened a private-preview API letting regulators and researchers check for that watermark.
• Per-Message Effort Control: A beta setting lets an ongoing conversation switch its adaptive-thinking depth partway through, something earlier Claude models could not do without breaking the prompt cache.

Pros

• Leads rival frontier models on the SWE-bench Pro leaderboard and the composite Vals Index, at the same base per-token price as Fable 5.
• The cache-read discount specifically rewards long, cache-heavy agentic sessions, the workload pattern where Fable 5 was previously most expensive to run continuously.
• The retuned classifiers matter most for security teams and clinicians who were previously misclassified as risks under Fable 5's broader rules, not just a headline safety metric.

Cons

• Mandatory data retention still applies outside the new Enterprise Frontier Safeguards program; a standalone zero-retention option remains unavailable.
• Costs twice as much per token as Claude Opus 5, and Anthropic's own documentation recommends starting with Opus 5 for most workloads.
• Has no audio or video support, and always-on adaptive thinking makes it the slowest model in Anthropic's current lineup by published latency.

Benchmarks

• medcode: 53.51
• mmlu pro: 92.38
• mmmu pro: 90.64
• medscribe: 91.29
• vals index: 67.87
• tax eval v2: 75.96
• gdpval aa v2: 1853
• swe bench pro: 81.2
• live code bench: 90.52
• automation bench: 31.4
• proof bench v1 1: 100
• cursor bench 3 2 0: 73.4
• osworld 2 0 strict: 41.7
• terminal bench 2 1: 85.02
• terminal bench 4 0: 55.8
• humanitys last exam: 60.9
• osworld 2 0 partial: 77.9
• legal research bench: 55.29
• terminal bench science 0 1: 52.6
• harveys legal agent benchmark: 6.67
• humanitys last exam with tools: 65
• artificial analysis intelligence index: 66
• artificial analysis price blended per m: 7.17
• artificial analysis speed tokens per sec: 66.2

Visit Claude Fable 5.1 Official Page