benchr
5.1New endpoint, not an aliasA new model ID, not a silent upgrade
The production endpoint is claude-fable-5-1. Treat it as a distinct deployment target: pin it in the pilot, preserve a Fable 5 control group, and record the request settings used for every comparison. Anthropic's current lifecycle documentation says Fable 5 remains active and will not retire before June 9, 2027; a newer model is not evidence that the earlier one is unavailable.
The access boundary also matters. Anthropic says Fable 5.1 is generally available across its supported API and cloud channels, while Mythos 5.1 is restricted to vetted Project Glasswing or trusted access. Do not present the Mythos ID as a public fallback merely because the two models share headline specifications.
The rate card: the cache line is the meaningful change
| Category | Price | What it covers |
|---|---|---|
| Standard input | $10.00 / 1M | Non-cached input tokens |
| Standard output | $50.00 / 1M | Output tokens |
| Cache read | $0.25 / 1M | Cached input tokens |
| Cache write, 5 minutes | $12.50 / 1M | Creating a five-minute cache entry |
| Cache write, 1 hour | $20.00 / 1M | Creating a one-hour cache entry |
| Batch input / output | $5.00 / $25.00 | Per 1M tokens; half of standard input and output |
The $0.25 figure is a cache-read price, not a blanket 75% reduction in prompt cost. A realistic estimate separates non-cached input, cache writes, cache reads, output, and any applicable Batch workload. The cost calculator can compare the standard token fields; keep cache duration and reuse assumptions in the deployment estimate.
Migration gates that can return an error
Fable 5.1 changes the request contract in ways a generic “swap the model ID” migration can miss. Anthropic's migration guide explicitly says that forced tool selection is invalid and that adaptive thinking is always on.
| Client behavior | Fable 5.1 behavior | Migration action |
|---|---|---|
tool_choice type any or a named tool | HTTP 400 | Use auto or none; keep application-side enforcement outside the request. |
Manual or disabled thinking configuration | Invalid | Remove it. Adaptive thinking is always on. |
| Assistant-message prefill | HTTP 400 | Move the instruction into the allowed prompt structure. |
| Reusing or editing prior thinking blocks | Compatibility and invalidation rules apply | Replay multi-turn test cases; do not assume a transcript edit preserves thinking blocks. |
| Priority Tier deployment | Not supported | Select a supported lane before committing a latency SLO. |
Those are integration facts, not reasons to avoid the model. They are reasons to make the pilot an interface test as well as a task-quality test. Save the request payload, response schema, tool trace, error status, token use, and human decision for matched cases. A benchmark score will not tell you whether a forced tool-choice request will fail in your client.
Retention and restricted access are rollout constraints
Anthropic states that Fable 5.1 normally requires 30-day data retention for safety monitoring. Zero-data-retention use requires authorization. That is a procurement and policy question, not a knob a team should infer from an older Claude configuration. Confirm the approved data path before routing production records to the new endpoint.
How to read the launch benchmarks
Anthropic reports Fable 5.1 results including 55.8% on Terminal-Bench 4.0, 52.6% on Terminal-Bench-Science 0.1, 77.9% partial success and 41.7% strict success on OSWorld 2.0, and 60.9% / 65.0% on Humanity's Last Exam without / with tools. These are provider-reported results from Anthropic's release material, not benchr measurements. The announcement also describes its configuration and safeguards, so they are not a substitute for a controlled comparison against a differently configured model or vendor run.
A small, reversible rollout
Keep representative prompts, tools, schemas, accepted outputs, and token use.
Remove invalid tool-choice and thinking settings, then replay error and multi-turn cases.
Separate cache reads, writes, uncached prompts, outputs, and any Batch jobs.
Approve the model route after task, policy, reliability, and cost evidence agree.