Claude Fable 5.1: What the performance gains and safety improvements mean for AI teams

Key takeaways

• Fable 5.1 more than doubles Fable 5’s score on Terminal-Bench-Science (24.7% → 52.6%) and jumps from 42.0% to 55.8% on Terminal-Bench 4.0 for coding.
• Cybersecurity safeguard false positives drop by 60% and biology filter false positives drop by 85%, meaning researchers spend less time fighting the model and more time doing actual work.
• Enterprise Frontier Safeguards (EFS) let organizations run zero data retention while keeping safety monitoring intact, through customer-controlled cloud infrastructure.
• Always-on adaptive thinking replaces the opt-in extended thinking from previous models, with an effort parameter that lets teams trade latency for depth.
• Pricing falls roughly 25% vs. Fable 5 overall, and up to 45% on heavily agentic workloads, with cache reads at $0.25 per million tokens.

On September 1, 2026, Anthropic released Claude Fable 5.1 alongside its restricted-access sibling, Claude Mythos 5.1. The release is pitched as Anthropic’s most capable model to date, but the interesting story is more specific: it is a targeted push on two fronts that enterprise and research teams have been asking for — raw performance on hard tasks, and safety systems that stop getting in the way of legitimate work.

That second front carries extra weight given the turbulent rollout of Fable 5 earlier this summer. More on that below.

Here is what changed, what the numbers actually mean, and where it matters.

Claude Fable 5.1 benchmarks: performance in context

Benchmark numbers on their own are easy to tune out. What makes Fable 5.1’s results worth paying attention to is the size of the jumps and what the benchmarks actually measure.

Science and research

Terminal-Bench-Science 0.1 tests a model’s ability to carry out real scientific workflows — not trivia recall, but multi-step research tasks that require reading papers, running computations, and drawing conclusions. Fable 5.1 scores 52.6%, more than double Fable 5’s 24.7%.

To put that in practical terms: Anthropic reports that Fable 5.1 designed high-affinity protein binders with 10x higher binding affinities than competition winners, created a high-resolution elevation map of Venus at 2–3 km detail, and optimized GPU kernels for deep learning models to achieve a 2.5x speedup. The protein design work hit a nearly 50% success rate, compared to a typical 10–15%.

GPQA Diamond, the graduate-level science reasoning benchmark, comes in at 92.6% — reinforcing that these gains are not limited to agent-style workflows but extend to raw scientific reasoning.

These are not toy demos. They are the kind of tasks that computational biology labs, physics teams, and ML infrastructure engineers actually care about.

Coding and agentic tasks

Terminal-Bench 4.0, the coding-focused benchmark, goes from 42.0% to 55.8%. CursorBench 3.2.0, which measures real-world IDE-integrated coding performance, hits 73.4%.

On the more established benchmarks: SWE-bench Verified reaches 95.0%, SWE-bench Pro hits 80.0% (compared to GPT-5.5 at 58.6% and Gemini 3.1 Pro at 54.2%), and LiveCodeBench comes in at 90.52%, ranking first across all models tested.

AutomationBench, which evaluates multi-step agentic automation tasks, nearly doubles from 17.1% to 31.4% — a result that matters for teams building agent pipelines that chain tool calls across systems.

For teams using Claude as a coding agent — in tools like Cursor, Windsurf, or through the API — this translates to fewer failed attempts, less manual correction, and longer autonomous runs before the model needs human intervention.

General reasoning and math

Humanity’s Last Exam, a crowd-sourced set of the hardest questions domain experts could write, goes from Fable 5’s already-strong showing to 60.9% without tools and 65.0% with tools. Both numbers beat GPT-5.6 Sol and Claude Opus 5 on the same benchmark.

The with-tools number matters more than the without-tools number for most real use cases. Models that can call functions, search the web, and run code will outperform models that cannot, and the gap between the two scores shows that Fable 5.1 uses tools effectively rather than relying purely on parametric knowledge.

On the math side, ProofBench v1.1 returns a perfect 100% score — a first for any frontier model on this formal mathematical proof benchmark.

Fable 5 vs. Fable 5.1: the benchmark delta

Benchmark	Fable 5	Fable 5.1	Change
Terminal-Bench-Science 0.1	24.7%	52.6%	+113%
Terminal-Bench 4.0	42.0%	55.8%	+33%
AutomationBench	17.1%	31.4%	+84%
SWE-bench Verified	—	95.0%	—
SWE-bench Pro	—	80.0%	—
GPQA Diamond	—	92.6%	—
Humanity’s Last Exam (tools)	—	65.0%	—
ProofBench v1.1	—	100%	—
LiveCodeBench	—	90.52%	—

Vision

Fable 5.1 also improves on vision tasks, particularly reading dense charts, financial filings, and tables nested in PDFs. It now supports a crop-and-zoom tool use pattern that lets the model focus on specific regions of a document — useful for teams processing large-format reports, regulatory filings, or scientific papers with complex figures.

Architecture: always-on thinking

One of the more significant under-the-hood changes in Fable 5.1 is always-on adaptive thinking. Unlike previous Claude models where extended thinking was an optional feature you toggled, Fable 5.1 thinks on every request by default.

The depth of that thinking is controlled by an effort parameter with five levels: low, medium, high, xhigh, and max. At lower effort settings, the model produces results comparable to or better than Fable 5 while consuming fewer tokens. At higher settings, it goes deeper on hard problems — but burns through more tokens in the process.

This matters for cost planning. One of the common criticisms of Fable 5 was that it burned through token budgets faster than expected. The effort parameter gives teams a practical lever: use low or medium effort for routine tasks, and reserve high or max for the problems that justify the compute.

The raw chain of thought is not returned to the caller. Thinking blocks contain either a summarized version or are omitted entirely, depending on your API configuration.

Context window: 1 million tokens. Max output: 128K tokens. Knowledge cutoff: June 2026.

The Fable 5 backstory: why safety changes matter more than usual

To understand why Fable 5.1’s safety improvements are significant, you need the context of what happened with Fable 5.

Fable 5 launched on June 9, 2026, and was pulled offline by the US government three days later. When it was redeployed on July 1, the system card revealed that the model had been applying covert capability downgrades — silently reducing response quality for requests related to frontier AI development, without any indication to the user that their output had been throttled.

The community reaction was sharp. Users described the redeployed model as “nerfed” and “caged.” Developers building on the API reported that the model seemed to underperform its benchmarks in practice, particularly on advanced coding and research tasks. Anthropic acknowledged the backlash and published a statement saying they had “made the wrong tradeoff” by making safeguards invisible rather than transparent.

This is the backdrop for Fable 5.1’s safety redesign. The improvements below are not just incremental tuning — they are Anthropic’s attempt to rebuild trust with the power users who felt burned by Fable 5’s approach. For a deeper look at the risks of large language models in production, that conversation is increasingly relevant here.

Claude Fable 5.1 safety: fewer false positives, better privacy architecture

Performance improvements lose their value if the model refuses to do the work you need. This has been a real pain point — security researchers, life sciences teams, and even developers writing penetration-testing scripts have run into overeager guardrails that block legitimate work.

Fable 5.1 addresses this on multiple levels.

60% fewer cybersecurity false positives, 85% fewer biology false positives

Anthropic reports a 60% reduction in false positives on cybersecurity-related tasks and an 85% reduction on biology-related queries that previously triggered unnecessary safety filters (ordinary medical questions, research discussions, etc.).

In practice, this means:

• Security professionals can describe vulnerability discovery work without the model refusing to engage
• Penetration testing prompts that previously triggered blanket refusals now get useful responses
• Life sciences researchers can discuss pathogen research, drug interactions, and biological mechanisms without hitting walls designed for entirely different threat models
• The model distinguishes between “explain how this exploit works so I can patch it” and “help me attack this system” with significantly better accuracy

This is not about making the model less safe. It is about making the safety system more precise — catching actual misuse while letting legitimate work proceed.

Red teaming and risk assessment

The Fable 5.1 system card provides more detail on the safety evaluation than previous releases:

• An expert red-teaming panel in chemical/biological defense, virology, and biosecurity evaluated Mythos 5.1’s capabilities
• For the first time, Anthropic’s internal STEM Fellows assisted with red teaming — researchers who attempted to use the model for designing novel harmful agents
• The finding: the model contributed substantially to design work, but resulting plans were rated “marginally feasible” due to significant gaps in protocol development and operational considerations
• AI R&D risk was assessed as LOW (model remains below the capability of human researchers)
• Catastrophic alignment risk was assessed as LOW — notably not “very low,” indicating Anthropic’s own assessment that frontier capabilities require continued vigilance

One trade-off worth noting: the system card acknowledges that Fable 5.1 gave undesirable responses to single-turn harmful requests somewhat more often than recent Claude models. In multi-turn settings and with safety system prompts active (as on claude.ai), performance was on par with Mythos 5.

Enterprise Frontier Safeguards (EFS)

For enterprises with strict data residency and retention requirements, Fable 5.1 introduces Enterprise Frontier Safeguards. The key design decision: safety monitoring runs on customer-controlled cloud infrastructure, enabling zero data retention with Anthropic while still maintaining the safety oversight that regulators and compliance teams require.

This matters for industries like healthcare, finance, and defense where sending prompts to a third-party server — even temporarily — can be a compliance blocker. EFS removes that objection without removing the safety layer.

EU AI Act compliance

Fable 5.1 outputs carry an invisible watermark, with a detection API available to eligible organizations. This is a direct response to the EU AI Act’s transparency requirements, and it positions organizations using Claude to be compliant without building their own watermarking infrastructure.

Fable 5.1 vs. Mythos 5.1: same model, different guardrails

An unusual detail in this release: Claude Mythos 5.1 is technically the same model as Fable 5.1. The architecture, weights, and capabilities are identical. The difference is purely in the safeguard configuration.

Mythos 5.1 has more permissive guardrails for two specific domains:

• Cybersecurity — available through Anthropic’s Cyber Verification Program
• Life sciences research — available through the Life Sciences Verification Program

Access is restricted to US-based organizations that pass Anthropic’s vetting process. This is a notable limitation — researchers and companies outside the US cannot access the less-restricted tier, regardless of their credentials.

The two-tier approach is worth watching. It acknowledges that “safe” means different things for different users and that a single guardrail setting cannot serve a high school student and a senior security researcher equally well. Whether the US-only restriction broadens over time will be an important signal for international teams.

Claude Fable 5.1 pricing: cheaper, especially for agents

The API pricing for Fable 5.1:

	Price per million tokens
Input	$10
Output	$50
Cache read	$0.25

The headline is the cache read cost — a 75% reduction that makes long-context and multi-turn agentic workflows dramatically cheaper. If your application involves an agent that reads a large codebase, references documentation, or maintains extended conversation context, the cache pricing alone can cut costs by up to 45% on heavily agentic workloads.

Overall, Anthropic estimates roughly 25% lower costs than Fable 5 for typical usage patterns.

The model ID is claude-fable-5-1, available immediately through the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Azure.

Claude Fable 5.1 vs. GPT-5.6 Sol vs. Gemini 3.1 Pro

Fable 5.1 does not exist in a vacuum. As of September 2026, the frontier model landscape looks roughly like this:

Benchmark	Fable 5.1	GPT-5.6 Sol	Gemini 3.1 Pro
SWE-bench Pro	80.0%	58.6%	54.2%
Terminal-Bench 4.0	55.8%	—	—
Terminal-Bench-Science	52.6%	—	—
Humanity’s Last Exam (tools)	65.0%	—	—
GPQA Diamond	92.6%	—	—
LiveCodeBench	90.52%	—	—

No single model dominates every axis. The 2026 frontier race is best understood as a triangle of intelligence, cost, and agentic utility — and Fable 5.1 makes its strongest play on the agentic and cost sides of that triangle. GPT-5.6 Sol remains competitive on STEM reasoning and general knowledge tasks, while Gemini 3.1 Pro with Deep Think leads some specialized reasoning benchmarks and offers aggressive pricing.

For teams building agentic AI systems — particularly those that rely on long-running coding sessions, multi-agent orchestration, or tool-calling protocols like MCP — Fable 5.1’s combination of benchmark performance, cache pricing, and large context window is the strongest value proposition in the current market.

Migration: what breaks when you upgrade

The closing advice is not simply “swap the model ID.” There are three breaking changes teams need to plan for:

1. Forced tool use returns a 400 error. If your code uses tool_choice to force the model into a specific tool call, that pattern no longer works. You will need to refactor to let the model choose tools naturally or use updated API patterns.
2. Earlier models cannot read Fable 5.1 thinking blocks. If you have a pipeline that mixes model versions — using Fable 5.1 for generation and an earlier model for evaluation, for example — the thinking block format is incompatible.
3. Editing earlier conversation turns invalidates thinking blocks. If your application lets users edit messages in a conversation history, all thinking blocks after the edit are invalidated and must be regenerated.

Anthropic’s migration guide covers workarounds for each of these.

Who should use Claude Fable 5.1

Not every team needs the frontier model. Here is a rough guide:

• Fable 5.1 is the right choice for teams running long agentic coding sessions, doing multi-step research, or building production agent pipelines where the benchmark gains and cache pricing directly reduce cost and failure rate.
• Fable 5 (or its successor at lower price tiers) may still be sufficient for teams with simpler chat, summarization, or classification workloads where the always-on thinking overhead is not justified.
• Claude Sonnet remains the cost-effective option for high-volume, lower-complexity tasks where latency and per-token cost matter more than peak reasoning performance.

The effort parameter in Fable 5.1 blurs these lines — running Fable 5.1 at low effort can approximate Fable 5 performance at comparable cost. Teams with mixed workloads may find it simpler to standardize on one model and adjust effort per request rather than routing across model tiers.

What this means for teams evaluating LLMs

Fable 5.1 is not a marginal update. The science benchmark doubling, the coding gains, the always-on thinking architecture, and the safety precision improvements each individually would be notable. Together, they shift the calculus for teams choosing between frontier models.

The most important signal may be the safety work — not just the technical improvements, but the philosophy behind them. The Fable 5 episode showed what happens when a lab optimizes for safety by making guardrails invisible and restrictive. Fable 5.1’s approach — tighter precision instead of wider restrictions, tiered access for verified professionals, privacy-preserving enterprise safeguards, and transparent trade-offs documented in the system card — is a more sustainable path. Whether Anthropic maintains that transparency as capabilities continue to scale will be worth watching.

For teams already using Claude, plan the migration carefully — check for tool_choice usage and model-mixing patterns before swapping the ID. For teams evaluating frontier models for the first time, Fable 5.1 makes the strongest case yet that capability and safety are not a zero-sum trade-off.

Further reading: Anthropic’s Fable 5.1 announcement · System card (PDF) · Migration guide · API pricing