Anthropic’s Claude Fable 5.1 keeps a one-million-token context window but cuts cache reads by 75 percent, from 1 dollar to 25 cents per million tokens. For coding agents that re-read the same context thousands of times, that changes the arithmetic more than any benchmark score.
Key takeaways
- Claude Fable 5.1 was released on 1 September 2026 and is generally available, while the companion model Claude Mythos 5.1 is restricted to approved organisations through a trusted access programme.
- The two models share the same underlying model, with Fable 5.1 carrying additional safeguards for dual-use capabilities in cybersecurity and the life sciences.
- List pricing is 10 dollars per million input tokens and 50 dollars per million output tokens, but the cache read price has fallen 75 percent, from 1 dollar to 25 cents per million tokens.
- Anthropic states that typical workloads cost about 25 percent less than Fable 5 and highly agentic workloads up to 45 percent less, with most of the saving attributed to the cache-read change.
- The combination of a one-million-token context window and cheap cache reads mainly benefits workloads that hold a large, stable body of context across many model calls.
What has actually changed
On 1 September 2026, Anthropic released Claude Fable 5.1, alongside Claude Mythos 5.1. The API model identifier is claude-fable-5-1 and the context window is one million tokens. Fable 5.1 is available to Pro, Max, Team and Enterprise subscribers, on the Claude Platform, and through Amazon Web Services, Google Cloud and Microsoft Foundry.
The headline change for anyone running the model at scale is not the input or output price. Those are 10 dollars and 50 dollars per million tokens respectively. It is the cache read price, which has fallen by 75 percent, from 1 dollar to 25 cents per million tokens. Anthropic says typical workloads cost about 25 percent less than Fable 5, that highly agentic workloads cost up to 45 percent less, and that most of the saving comes from the cache-read change.
Reported benchmark scores accompany the release: 95.0 percent on SWE-bench Verified, Terminal-Bench-Science rising from 24.7 to 52.6 percent, and Terminal-Bench 4.0 coding rising from 42.0 to 55.8 percent. The Terminal-Bench figures are the larger movements, but they describe capability rather than cost.
Why this matters now rather than in June
Fable 5 was released in June 2026. A point release a few months later would normally be a routine capability bump. What makes this one worth explaining is that the pricing structure, not the model weights, is where the practical difference sits for a particular and increasingly common class of work: agents that run for a long time over a large fixed body of material.
Prompt caching is not new, and neither is a large context window. What is new is the ratio between them. When cache reads cost a quarter of what they did, the cost of keeping a large context resident across many turns drops sharply, while the cost of generating new output does not. That shifts which architectures are economically sensible.
The background a newcomer needs
A language model has no memory between calls. Every request must carry the full context the model needs — instructions, code files, prior conversation, tool results — and that context is charged as input tokens.
Prompt caching addresses the resulting waste. When a large block of context stays identical across requests, the provider can store a processed form of it. Subsequent requests that reuse that prefix are billed at the cache read rate rather than the full input rate. Writing to the cache typically costs more than a normal input token; reading from it costs much less.
An agent that works through a large codebase over hundreds of steps re-sends most of the same context every time. Only the tail changes: the last tool call, the last file read, the last decision. Under a high cache read price, that repetition is the dominant line item. Under a low one, it fades into the background and output generation becomes the thing that costs money.
The one-million-token context window matters here for a related reason. A large window is only useful if you can afford to fill it repeatedly. A window you can load once but not re-read across a long session is of limited practical value for agentic work. The two changes are complementary: capacity plus a low re-read cost is what makes holding a whole repository in context a routine choice rather than an expensive one.
Who is affected and how
The clearest beneficiaries are teams running long-horizon coding agents — systems that read a repository, plan a change, edit files, run tests, read failures and iterate. These workloads are cache-heavy by construction. Anthropic’s stated figure of up to 45 percent for highly agentic workloads points at exactly this pattern.
Teams doing short, varied interactions see much less. If each request carries different context, there is little to cache, and the input and output prices — unchanged — determine the bill.
There is also a governance dimension. Fable 5.1 and Mythos 5.1 share the same underlying model, but Fable 5.1 carries additional safeguards for dual-use capabilities in cybersecurity and the life sciences, and Mythos 5.1 is available only to approved organisations through a trusted access programme. For most enterprise buyers this is invisible; for organisations doing security or biological research it means the generally available model is the one they will work with, with those safeguards in place.
Availability across Amazon Web Services, Google Cloud and Microsoft Foundry means procurement can generally route through an existing cloud contract. Whether cache pricing and behaviour are identical on every platform is not something that can be stated here.
Where informed people disagree
Three disagreements are worth naming.
The first concerns whether cheaper cache reads reduce spending or simply move it. Lower per-unit costs often invite more usage: longer contexts, more retries, more parallel agents. The published percentages describe cost per workload, not total cost per organisation, and the two can move in opposite directions.
The second concerns the benchmark figures. SWE-bench Verified at 95.0 percent leaves little headroom, which raises the familiar question of what remains to be measured near the top of a benchmark. Terminal-Bench-Science more than doubling, from 24.7 to 52.6 percent, is a much larger jump — but a score moving from roughly a quarter to roughly half also describes a task the model still fails about half the time. How these translate to production work is not established by the numbers alone.
The third concerns architecture. If large contexts are cheap to re-read, one school of thought says retrieval pipelines become less necessary — put everything in the window. Another holds that retrieval remains valuable for reasons unrelated to price: relevance, latency and the tendency of models to perform worse when the useful material is buried in a large volume of irrelevant text. This is an open engineering question, not a settled one.
The practical implications
For teams already using Fable 5, the migration path is a model identifier change, and the cost effect depends on how cache-heavy the workload is. Measuring the existing split between cache reads, fresh input and output tokens is the way to estimate the change; applying the headline percentage without that measurement is likely to mislead.
For teams designing new systems, the incentive now favours stable context. A prompt prefix that stays byte-identical across turns caches well. One that changes near the beginning — a timestamp, a shuffled tool list, a reordered set of files — invalidates the cache and forfeits the saving. This is a concrete engineering discipline, and its value has risen relative to other optimisations.
For budget holders, output tokens are now a larger share of the total for agentic work. At 50 dollars per million, verbose reasoning and long generated artefacts carry proportionally more weight than they did.
What to watch next
Watch whether independent evaluations reproduce the reported benchmark movements, particularly on Terminal-Bench-Science, where the jump is largest. Watch whether the cost reductions hold in practice for real workloads rather than representative ones. Watch how competing providers respond on cache pricing specifically, since it is the lever that most directly affects agentic costs. And watch whether the trusted access model used for Mythos 5.1 — a restricted-availability sibling to a generally available model with additional safeguards — becomes a common pattern for releasing capable models.
Frequently asked questions
What is Claude Fable 5.1?
Claude Fable 5.1 is a language model released by Anthropic on 1 September 2026. It uses the API identifier claude-fable-5-1 and has a one-million-token context window. It is generally available to Pro, Max, Team and Enterprise subscribers, on the Claude Platform, and through Amazon Web Services, Google Cloud and Microsoft Foundry. It succeeds Fable 5, which was released in June 2026.
How much does Claude Fable 5.1 cost?
List pricing is 10 dollars per million input tokens and 50 dollars per million output tokens. Cache reads cost 25 cents per million tokens, down 75 percent from 1 dollar. Anthropic states that typical workloads cost about 25 percent less than Fable 5, and that highly agentic workloads cost up to 45 percent less, with most of the saving attributed to the cache-read change.
What is a cache read and why does the price matter?
When the same block of context is sent repeatedly, a provider can store a processed form of it and bill later reuse at a reduced cache read rate. Agents that work over a large codebase re-send most of their context on every step, so cache reads dominate their token usage. Cutting that price by 75 percent therefore reduces the bill for repetitive, long-running work far more than for short one-off requests.
What is the difference between Fable 5.1 and Mythos 5.1?
The two share the same underlying model. Fable 5.1 is generally available and carries additional safeguards for dual-use capabilities in cybersecurity and the life sciences. Mythos 5.1 is restricted to approved organisations through a trusted access programme. In practical terms, most users and organisations will work with Fable 5.1, since Mythos 5.1 requires approval for access.
What benchmark scores were reported?
The reported figures are 95.0 percent on SWE-bench Verified, Terminal-Bench-Science rising from 24.7 to 52.6 percent, and Terminal-Bench 4.0 coding rising from 42.0 to 55.8 percent. The Terminal-Bench movements are the larger ones. Benchmark scores describe performance on defined test sets and do not automatically predict results on a specific production workload, which needs separate evaluation.
Does a one-million-token context window remove the need for retrieval?
Not straightforwardly. A large window makes it possible to hold more material in context, and cheaper cache reads make re-reading that material less costly. But retrieval also addresses relevance and latency, and models can perform less well when relevant information sits inside a large volume of irrelevant text. Whether to drop retrieval is an engineering judgement that depends on the specific application.
Sources and further reading
- Anthropic — the company’s own model documentation and pricing pages, which are the primary record for model identifiers, context window sizes and list prices.
- SWE-bench and Terminal-Bench project documentation — for how these evaluation suites are constructed and what their scores measure.
- Cloud platform documentation from major providers — for how model availability, pricing and caching behaviour are handled on each platform.
- Technology trade press covering enterprise AI procurement — for reporting on how organisations budget for and deploy large language models.
Surfaced from the manual signal “an AI model release”. AI-assisted draft, editorially reviewed.
