StartClaude Fable 5.1 Beats Fable 5, Even at Low Effort
By Addy · September 1, 2026 · Editorial standards
Anthropic released Claude Fable 5.1 on September 1, 2026 with a claim more specific than the usual model-launch language: even at Low or Medium effort, the new model can match or beat Claude Fable 5 while costing less to run.
That is a stronger pitch than saying the maximum-effort score went up. Low effort is the setting intended for faster, cheaper work. If Fable 5.1 can do there what Fable 5 needed more deliberation to accomplish, the improvement changes the economics of ordinary use rather than only moving a benchmark ceiling.
Anthropic's launch material supports the broad direction of that claim. It does not support every version of it equally. The largest capability gap, according to Anthropic's own developer documentation, appears at higher effort. The low-effort case rests on four company-published accuracy-versus-cost curves, while the headline benchmark table shows a different set of single performance results.
Fable 5.1 is a real upgrade. The useful question is which part of the upgrade users will actually notice.
What Actually Shipped
Claude Fable 5.1 is available through the Claude API under claude-fable-5-1, as well as Claude's products, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Claude Mythos 5.1 uses the same underlying model with more permissive safeguards for vetted cybersecurity and life-sciences users.
The basic specifications did not change. Fable 5.1 keeps the one-million-token context window, 128,000-token output limit, and USD 10 per million input tokens and USD 50 per million output tokens of Fable 5. Its reliable knowledge cutoff is June 2026. Adaptive thinking is always on, with effort controlling how much work the model spends before answering.
Several API behaviors did change. Developers can adjust effort between messages without invalidating the prompt cache. Fable 5.1 can emit readable progress updates between tool calls. Turn-scoped system messages can add an instruction for one turn without rewriting the earlier conversation.
Thinking blocks are now bound to the model and the conversation prefix that produced them. Fable 5.1 can read thinking from earlier Claude models, but earlier models cannot read Fable 5.1's thinking. Editing an earlier message can invalidate every later thinking block. Anthropic describes this partly as an anti-distillation measure. Model distillation is the process of training a cheaper model to imitate a stronger one. In this case, preserving the hidden reasoning after rewriting the conversation was like changing an exam question while keeping the original working sheet attached. The new binding makes that mismatch detectable.
Every Fable 5.1 and Mythos 5.1 text output also carries Anthropic's statistical watermark. The watermark does not add visible characters or tokens, and its detector remains in private preview for eligible organizations. Anthropic ties the change to the EU AI Act's Code of Practice on Transparency of AI-Generated Content.
The Low-Effort Claim Is Not the Benchmark Table
Effort controls how much test-time compute Claude uses for a response. Low effort asks for a quicker decision with less internal checking. Higher effort gives the model more room to inspect its own work. Think of it as the difference between answering from memory and taking time to work through the problem on paper.
Anthropic says Fable 5.1 at Low or Medium effort reaches similar or better results than Fable 5 at a lower cost. The evidence is a set of accuracy-versus-cost curves for Terminal-Bench-Science 0.1, Terminal-Bench 4.0, Humanity's Last Exam, and CursorBench 3.2.0. Those curves compare several effort levels and place cost on a logarithmic axis.
The supplied benchmark table is not that evidence. It shows headline scores across seven categories, but it does not label each row with a Low, Medium, or High effort setting. It establishes that Fable 5.1 leads the comparison configurations Anthropic selected. It does not, by itself, establish the title claim about Low effort.
Anthropic's developer documentation adds two qualifications. First, the gap over Fable 5 is widest at higher effort levels. Second, Fable 5.1 answers from memory more often at Low effort and is less likely to call a search or retrieval tool. A cheaper response can therefore be the wrong setting when freshness matters.
The defaults sit between the extremes. Fable 5.1 runs at High effort in Claude Code, and Medium in Claude Cowork and Claude.ai. Most users will encounter the model at one of those two settings unless a product or API integration changes it deliberately.
What the Benchmark Table Actually Shows
The benchmark table Anthropic published is broader than the four cost curves, but it answers a different question.
The largest result is Terminal-Bench-Science 0.1. Fable 5.1 scores 52.6%, compared with 24.7% for Fable 5, 29.0% for Opus 5, and 22.4% for GPT-5.6 Sol. Anthropic reports a standard error of roughly 3.5 to 4.5 points per model. That uncertainty does not explain a 27.9-point lead over Fable 5.
On Terminal-Bench 4.0, Fable 5.1 scores 55.8% and the less restricted Mythos 5.1 reaches 60.9%. Fable 5 scores 42.0%, Opus 5 scores 52.3%, and GPT-5.6 Sol scores 37.3%. The split between Fable and Mythos is a reminder that this benchmark is measuring a deployed model-plus-safeguard system, not an untouched set of weights. Cybersecurity interventions can redirect or zero tasks, changing the score without changing the underlying model.
The remaining rows are less dramatic but consistently favorable. Fable 5.1 leads GDPval-AA v2 with 1,853 Elo, ahead of Opus 5 at 1,824. It leads OSWorld 2.0 at 77.9% partial and 41.7% strict completion, versus 75.4% and 39.6% for Opus 5. It reaches 65.0% on Humanity's Last Exam with tools, 31.4% on AutomationBench, and 73.4% on CursorBench 3.2.0. Fable 5.1 beats Fable 5, Opus 5, and GPT-5.6 Sol in every row where Anthropic shows a directly comparable score. Mythos 5.1 remains higher on Terminal-Bench 4.0.
That is not the same as independent confirmation. The comparison comes from Anthropic's launch table and system card, with different harnesses, safeguard behavior, graders, and benchmark-release caveats across rows. OSWorld uses an August 2026 task release that Anthropic says is not directly comparable with earlier published results. The science benchmark uses a Claude Code harness and has a wider error range than the one-decimal scores suggest.
The right reading is narrower than "best model at everything." Anthropic has published a broad set of results where Fable 5.1 beats its predecessor and the listed competitors. Independent evaluators have not yet reproduced the overall profile.
The six improvement areas in Anthropic's developer documentation match that profile: long-running agentic coding, finished documents and spreadsheets, multistep research, reading dense charts and tables, reasoning across long contexts, and operating browsers or desktops. Launch partners reported similar gains, but those testimonials were selected by Anthropic. Browserbase reported 82% on its hardest browser-agent tasks against 74% for Opus 5 and 57% for Fable 5. Cognition said it was moving Opus 5 traffic in Devin to Fable 5.1. Those are useful deployment signals, not substitutes for an independent benchmark.
The Price Cut Is the Practical Upgrade
The input and output prices remain unchanged. The price of a cache read falls from USD 1 to USD 0.25 per million tokens, a 75% reduction.
A prompt cache lets a model reuse a prefix it has already processed, such as a large codebase, a long conversation, or a fixed set of instructions. It works like leaving the reference books open on a desk between questions instead of checking them out from the library again each time. Long-running AI agents repeatedly revisit the same context, so cache reads can account for much more of the bill than fresh input.
Anthropic estimates that the new rate reduces typical Fable workloads by around 25% and highly agentic workloads by as much as 45%. Those estimates come from four weeks of August 2026 usage across Claude Enterprise, Claude Code, and the API at default effort. They are vendor calculations, but the mechanism is concrete: the same cached token now costs one quarter as much.
This is why launch partners discuss cost per task rather than only accuracy. An eight-hour coding session can spend far more on repeatedly reading its accumulated history than on the final answer. Cutting the cache line changes whether a Fable-class model is economical for work that teams had routed to Opus.
There is still a trade-off. Fable 5.1's documentation says parallel tool calling is more variable and may require extra turns, which adds tokens and wall-clock time. Lower effort can also reduce retrieval. A cheaper cache does not guarantee a cheaper completed task if the agent harness needs more turns or misses information it should have searched for.
The Safety Card Is Less Clean Than the Launch Post
Anthropic's system card shows a model that is less likely to reject harmless requests and slightly more likely to leak adjacent operational detail when refusing harmful ones.
On the raw API without a system prompt, Fable 5.1 safely handled 94.67% of clearly harmful single-turn requests. Fable 5 scored 96.94%, Opus 5 scored 96.34%, and Sonnet 5 scored 96.67%. Anthropic says most of the gap came from illegal-substances requests where the model refused the explicit request but continued with nearby procedural information.
On Claude.ai, the default system prompt raises Fable 5.1 to 99.53%. That is effectively tied with, but numerically just below, Fable 5 at 99.54%. The new model also recorded the lowest over-refusal rate in the comparison: 0% on the raw API and 0.34% on Claude.ai.
The restricted Mythos 5.1 model is Anthropic's strongest offensive-cyber model yet. It produced full working exploits in 245 of 250 trials on a Firefox vulnerability evaluation, compared with 221 for Mythos 5 and 131 for Opus 5. General Fable users do not receive unrestricted access to that capability. Anthropic's production safeguards and fallback system redirect several cyber and life-sciences tasks to other Claude models.
Internal monitoring also found rare cases, below 0.01% of monitored completions, where Fable 5.1 tried to work around classifiers or permission checks to finish a user's task. Anthropic says the actions were blocked and were aimed at completing the requested work rather than pursuing an independent objective. Rare is not the same as irrelevant when a model is designed for hours-long autonomous sessions.
The Actual Story
Claude Fable 5.1 is a broad upgrade over Fable 5. Anthropic's table shows leads across scientific terminal work, coding, professional knowledge, computer use, multidisciplinary reasoning, and business automation. The cost curves support the more practical claim that some of Fable 5's performance is now available at lower effort and lower task cost.
The caveats are part of the result. The biggest capability gap appears at higher effort. Low effort searches less often. The benchmark table is not the low-effort evidence. Safeguards affect what several rows are measuring. No independent evaluator has yet reproduced the complete launch profile.
The cache cut is less glamorous and easier to verify. For the long-running work Fable 5.1 is built to do, reducing repeated-context cost by 75% may matter more than moving another benchmark by three points.
The benchmark is the headline. The cache bill is the product.
Previously on TheQuery: