The Numbers That Matter

On September 1, 2026, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 — updates to its most capable frontier models. The headline number is Terminal-Bench-Science: Fable 5.1 scored 52.6%, up from Fable 5’s 24.7% and more than double any other publicly tested model. GPT-5.6 Sol sits at 22.4% on the same benchmark. For teams evaluating AI agents for scientific research workflows, that gap is the whole conversation.

The second number is $0.25 per million tokens — the new price for Fable 5.1 cache reads, down 75% from the $1.00 Fable 5 charged. Standard input and output rates remain unchanged at $10 and $50 per million tokens. Anthropic estimates the cache cut delivers roughly 25% lower costs on typical workloads and up to 45% on context-heavy agentic pipelines. Cognition shifted Devin’s Opus 5 traffic to Fable 5.1 on launch day.

What Terminal-Bench-Science Actually Tests

Terminal-Bench-Science 0.1 is not a benchmark you can game by training on the answers. It consists of 70 tasks drawn from working scientists’ real research workflows — submitted by 376 contributors across 22 countries, vetted at a 7.6% acceptance rate. Tasks survived domain expert review, technical validation, and quality checks. The result is something closer to a competence audit than a leaderboard.

The five domains covered: life sciences (19 tasks), physical sciences (17), mathematical sciences (17), engineering sciences (9), and earth sciences (8). Task types include data analysis, statistical inference, simulation, optimization, and scientific machine learning. There are no multiple-choice shortcuts — the agent must plan, execute in a terminal environment, and self-verify the output.

What makes the benchmark revealing is the comparison with general coding benchmarks. Every model tested scored 10+ percentage points lower on Terminal-Bench-Science than on standard coding benchmarks, with most dropping 60+ points. Strong performance on SWE-bench or HumanEval does not transfer to a protein structure prediction or climate model calibration task. This benchmark exists to measure exactly that gap — and Fable 5.1 is the first model to close it meaningfully.

What Changed Under the Hood

Anthropic has been explicit that Fable 5.1 is not a full model retrain — it is a targeted capability upgrade with the same 1M-token context window and 128K maximum output. Adaptive thinking remains always-on. The improvements concentrate on sustained problem-solving across long investigations: better at holding a hypothesis through multiple tool calls, checking its own work, and recovering when a sub-task fails.

The Mythos 5.1 variant targets a narrower audience. Available only to vetted US organizations through Anthropic’s verification program, it scored 60.9% on Terminal-Bench 4.0 (the general coding version), versus Fable 5.1’s 55.8%. On protein binder design, Mythos 5.1 achieved roughly a 50% hit rate — a domain where Fable 5 had clear limitations. On Humanity’s Last Exam without tools, Fable 5.1 reached 60.9%; with tool access, 65.0%.

One breaking API change worth flagging: forced tool use via tool_choice set to any or tool now returns errors. Pipelines built on Fable 5 that relied on this pattern need to be updated before migrating. Thinking blocks from Fable 5 are also model-bound — they cannot be carried into Fable 5.1 sessions, which means conversation history with reasoning traces needs to be rebuilt from scratch on upgrade.

The Cache Pricing Math for Production Stacks

The 75% cache read cut changes the economics for context-heavy deployments more than the benchmark headline does for most operators. A typical agentic task with large system prompts and persistent context can send 80–90% of its tokens as cache reads. At the old $1.00/MTok rate, those tokens cost 10% of the base input price per read. At $0.25/MTok, that falls to 2.5% — four times cheaper per cached token.

For a production stack running 100 million cache-read tokens per day, the difference is $75,000 per month in savings at constant usage. For smaller operators the absolute number is lower, but the relative impact on agentic workload costs is identical. Anthropic’s bet is clear: cheaper cache reads encourage longer system prompts, richer tool definitions, and more sustained agent sessions — all of which increase total API spend while making per-task costs more predictable.

The cost calculus for research agents is a theme we have tracked over time — our piece on AI Research Agents: The Data Quality Reckoning noted that inference cost is rarely the binding constraint; data quality and task specification are. Fable 5.1’s cache cut reduces the cost of iterating on those harder problems, which is the more meaningful unlock.

Who Fable 5.1 Is Actually For

The Terminal-Bench-Science jump makes Fable 5.1 the obvious choice for any organization deploying agents into scientific workflows — not because 52.6% is a solved problem (it still means failing 47% of hard research tasks) but because it more than doubles the previous best-in-class result. Academic labs using AI for literature synthesis, climate model pipelines, or compound screening now have a clearer upgrade path than they did before September 1.

The cache cut matters most to operators with high context reuse: customer support agents with large knowledge bases, coding assistants carrying full codebase context, or research pipelines that send the same multi-thousand-token method sections across thousands of runs. The economics now justify context investments that would have been marginal at $1.00/MTok.

For context on the original release — including the two-tier deployment architecture and US export restrictions — see our earlier piece: Claude Fable 5 and Mythos 5: The First US AI Export Ban. Fable 5.1 continues that pattern, with biosecurity and cybersecurity remaining the primary drivers of Mythos-tier access restrictions.

One caveat before drawing conclusions: the Terminal-Bench-Science 52.6% figure comes from Anthropic’s own evaluation. Independent labs have not yet reported matching runs on the same benchmark version. The gap over GPT-5.6 Sol is large enough to survive some measurement variance, but teams making deployment decisions should weight the independence question accordingly.

Further Reading

• VentureBeat: Claude Fable 5.1 and Mythos 5.1 — Detailed pricing breakdown and enterprise context, including how Cognition responded to the launch-day announcement
• Terminal-Bench-Science on GitHub — The benchmark repository with task examples and scoring methodology for teams considering independent evaluation
• Vellum: Fable 5.1 and Mythos 5.1 Benchmarks Explained — Practical migration guide covering the API breaking changes and when the economics favor switching from Opus 5