What Claude Fable 5.1 changes for long-running work
On September 1, Anthropic released Claude Fable 5.1, three months after Fable 5. The list price didn't move. The bill for agent work did, and the model is built for the kind of task that runs for hours with nobody watching.
What shipped
Fable 5.1 is generally available on claude.ai, the API (asclaude-fable-5-1), AWS, Google Cloud, and Microsoft Azure. It shipped alongside Claude Mythos 5.1, which is the same model with fewer safeguards, offered only to verified organizations doing cybersecurity and life-sciences work, and currently only in the US.
Anthropic's pitch is about behaviour more than raw capability. The model is meant to fix the root cause of a problem instead of the symptom, avoid shortcuts that look like progress, keep a written record as it works, and report what it verified rather than what it did.
The numbers
Anthropic's published results, Fable 5 against Fable 5.1:
| Benchmark | Fable 5 | Fable 5.1 |
|---|---|---|
| Terminal-Bench 4.0 | 42.0% | 55.8% |
| Terminal-Bench-Science 0.1 | 24.7% | 52.6% |
| CursorBench 3.2.0 | 70.5% | 73.4% |
| OSWorld 2.0 (strict) | 36.1% | 41.7% |
| Humanity's Last Exam (with tools) | 63.8% | 65.0% |
The gap is widest on the terminal benchmarks, which is where an agent actually lives. CursorBench, the closest thing to editor-driven coding, moves less. Read that as: the model got better at working through a task, not just at writing a function.
These are the vendor's own figures. We treat them as a reason to test, not a conclusion.
The pricing change is the story
Input stays at $10 per million tokens and output at $50. Cache reads drop 75%, from $1 to $0.25 per million. Anthropic says that lands as roughly 25% lower cost on typical workloads and up to 45% on heavily agentic ones.
Cache reads matter more than the list price because an agent re-reads its context every turn: the spec, the files it has opened, the test output, the conversation so far. On a long run, most of the tokens billed are that context being read back, not new tokens being produced. Cut the price of the re-read and the long run gets cheaper in proportion to how long it is.
Anthropic also says the model at low or medium effort matches Fable 5 at a fraction of the cost, so the routine tier gets cheaper too.
What we're changing
We run coding agents on client work from a written spec, with a verification pass before anything ships. Three things in this release change that work.
- Unattended runs become reviewable. A model that documents its own work and reprioritizes as it goes is what makes an overnight run something a person can audit in the morning. That is the difference between an agent that saves time and an agent that produces a diff nobody trusts.
- Root cause over symptom. Anthropic's example is a rare crash a team couldn't explain for years. We've had the smaller version of that on every migration: the bug that lives three layers below where it shows. A model that keeps digging is worth more than a model that patches faster.
- The economics of long tasks. At up to 45% off agentic work, the question of whether to let the agent run the whole migration script and check it afterwards has a different answer. We're moving longer, well-specified tasks to Fable 5.1 as the default and keeping the effort setting low where the task is routine.
Caveats
"Avoids shortcuts" is a claim about behaviour, not a benchmark. We'll know whether it holds on real repositories, not from a launch post. Mythos 5.1 is gated and US-only, so the security-research capability is out of reach for most teams. And OpenAI shipped GPT-6 Astra two days later; we cover it inthe next post.
