Claude Fable 5.1 launches: pricing, access, benchmarks, and what changes in Claude Code
Anthropic has released Claude Fable 5.1, its new generally available frontier model for coding, computer use, and long-running knowledge work. The release is designed for the kind of multi-step assignments that increasingly define agentic software development: exploring an unfamiliar repository, implementing a change, running tools, checking the result, and recovering when the first approach fails.[1]
Fable 5.1 shares its underlying model with Claude Mythos 5.1, but the two products are not interchangeable. Fable is the broadly available release. Mythos uses a different safeguards configuration and is limited to Anthropic’s trusted-access programme. For most developers and teams, Fable 5.1 is therefore the version that can be evaluated and deployed today.[1]
What Fable 5.1 changes
Anthropic is positioning Fable 5.1 as an upgrade to both coding quality and sustained autonomous work. The company says the model is better at following long chains of instructions, navigating software interfaces, using visual feedback, and completing tasks that span many tool calls. In practical development workflows, that means Claude can write or modify code, run tests, inspect the result, and use screenshots or rendered output to verify its own work.[1]
Claude Code selects **High** effort by default for Fable 5.1. Cowork and Claude.ai use **Medium** effort by default. That distinction matters: effort is a trade-off between depth, latency, and token consumption, so a comparison made in Claude Code may not reproduce the behaviour seen in the chat product without matching the setting.[1]
The release is more evolutionary than disruptive at the API boundary. Developers can call the model as `claude-fable-5-1`, while existing Anthropic tooling and deployment channels remain in place. The meaningful changes are inside the model and its economics: stronger task completion, cheaper cache reuse, and more explicit safeguards for sensitive capabilities.[1][2]
Price and availability
Fable 5.1 costs **$10 per million input tokens** and **$50 per million output tokens** through the Claude API. Anthropic lists cache reads at **$0.25 per million tokens**, a 75% reduction from Fable 5. The company estimates that the lower cache price reduces the cost of a typical workload by about 25%, with savings approaching 45% for highly agentic work that repeatedly reuses a large context.[1][2]
| Item | Current information |
| --- | --- |
| API model | `claude-fable-5-1` |
| Input | $10 per million tokens |
| Output | $50 per million tokens |
| Cache reads | $0.25 per million tokens |
| Claude Code default effort | High |
| Cowork and Claude.ai default effort | Medium |
The model is available on Claude Pro, Max, Team, and Enterprise plans, as well as through the Claude Platform. Anthropic also lists AWS, Google Cloud, Microsoft Foundry, and partner marketplaces as distribution channels. Customers who choose US-only inference pay a 1.1× multiplier on input and output tokens.[1][2]
For engineering teams, the cache change may prove as consequential as the headline benchmark gains. Repository agents often send the same system prompt, code context, and tool definitions across many turns. Reducing the price of those repeated tokens can materially lower the cost of a completed task, even though the base input and output rates remain premium.
The benchmark claims
Anthropic reports substantial gains over Fable 5 across coding, computer use, science, and automation evaluations. Fable 5.1 scores **55.8% on Terminal-Bench 4.0**, compared with 42.0% for Fable 5, and **73.4% on CursorBench 3.2**, up from 70.5%. On AutomationBench, the reported result rises from 17.1% to **31.4%**.[1]
| Benchmark | Fable 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
| --- | ---: | ---: | ---: | ---: |
| Terminal-Bench 4.0 | 55.8% | 42.0% | 52.3% | 37.3% |
| Terminal-Bench-Science 0.1 | 52.6% | 24.7% | 29.0% | 22.4% |
| AutomationBench | 31.4% | 17.1% | 26.9% | 19.6% |
| CursorBench 3.2 | 73.4% | 70.5% | 70.0% | 67.2% |
Computer-use results move in the same direction. Anthropic reports 77.9% on its partial-credit OSWorld 2.0 evaluation and 41.7% under strict scoring, versus 72.9% and 36.1% for Fable 5. On Humanity’s Last Exam without tools, Fable 5.1 reaches 60.9%; with tools, it reaches 65.0%.[1]
These figures are vendor-published launch results, not independent measurements. Anthropic also notes that the evaluations ran with production safeguards enabled and that a safeguarded response can receive a score of zero. The OSWorld result uses the August 2026 task release, so it should not be compared directly with numbers from older versions of that benchmark.[1]
The safety boundary developers will notice
Fable 5.1 can help identify and explain software vulnerabilities, but Anthropic draws a harder line around exploit development and other high-risk cyber activity. The company says it reduced false-positive cyber blocks by 60%, while requests involving penetration testing, exploit generation, or binary-based vulnerability scanning can still be redirected to a less capable model. Mythos 5.1 exposes more of the underlying capability only to approved organisations operating under additional controls.[1]
That boundary has a practical consequence: teams should test their actual security and debugging workflows instead of assuming the benchmark score describes every request the production model will answer. A legitimate defensive task can look similar to an offensive one, and routing or refusals may affect completion rates in ways a general coding benchmark does not capture.
Anthropic says interactions are retained for 30 days by default. Eligible Enterprise customers can receive temporary zero-data-retention access, and the company plans to begin rolling out Enterprise Frontier Safeguards later in the fall. Those controls add monitoring and governance for organisations deploying frontier models in sensitive environments.[1][3]
Why it matters for developers
Fable 5.1 is most interesting when the unit of comparison is a finished engineering task rather than a single response. Higher benchmark scores are useful, but the larger operational story is a model intended to remain effective through long sessions while making repeated context substantially cheaper. That combination targets the real cost centre of coding agents: retries, repeated repository context, tool calls, and human review.
Teams evaluating the model should measure completion rate, wall-clock time, cache utilisation, review burden, and total cost on their own repositories. They should also record when safety routing changes the model used for a task. Anthropic’s numbers suggest a meaningful step forward, but production value will depend on how often Fable 5.1 completes the entire loop—from understanding the request to validating the result—without requiring a developer to take over.[1][2]