Anthropic released Claude Fable 5.1 on September 1, 2026. Its Claude API model ID is claude-fable-5-1, its context window is 1 million tokens, and its maximum output is 128,000 tokens. Standard API pricing is $10 per million input tokens and $50 per million output tokens. A successful prompt-cache read costs $0.25 per million tokens.
The direct answer for coding agents: Fable 5.1 is a premium specialist for difficult, long-horizon work, not the automatic default for every code change. Its low cache-read price can materially improve the economics of a loop that repeatedly reads the same repository context, instructions, and tool definitions. Cache writes, uncached input, generated output, tool results, cache misses, and repair attempts still count.
This article covers model specifications, API economics, and task fit. If you are comparing Pro, Max, Team, Enterprise, or Console billing, use the separate Claude Code pricing guide. To select Fable 5.1 in Claude Code and control its effort level, follow the focused Claude Fable 5.1 in Claude Code guide.
Facts and prices verified September 9, 2026. Model prices, specifications, product availability, and integrations can change. Check the linked Anthropic pages and the live product selector before relying on them.
Claude Fable 5.1 at a glance
Anthropic's official model page lists the following specifications:
The 1M context window is available at standard per-token pricing. That means there is no separate long-context rate, but a larger request still contains more billable tokens. Capacity is not the same as cost efficiency.
Anthropic's own selection guidance is also narrower than “use the most capable model.” It recommends starting with Claude Opus 5 for most workloads, then using Fable 5.1 for demanding reasoning and long-horizon agentic work, or when Opus 5 at higher effort still falls short. That is a sensible default unless your own evaluations show otherwise.
What prompt caching changes—and what it does not
Prompt caching lets an API application reuse a previously processed prompt prefix instead of paying the normal input rate every time. A stable prefix might include a system prompt, repository rules, tool definitions, reference documents, or earlier conversation history.
Anthropic's current prompt-caching prices separate three operations:
- a 5-minute cache write costs $12.50 per million tokens;
- a 1-hour cache write costs $20 per million tokens; and
- a successful cache read costs $0.25 per million tokens.
The write is more expensive than normal input because it creates the reusable entry. The later reads are much cheaper. The benefit therefore depends on how much of the prompt stays stable, how often it is reused before expiry, and whether the application actually produces cache hits.
A transparent cost example
Suppose a coding-agent loop has a stable 400,000-token prefix and sends ten requests within a valid 5-minute cache window. Ignore new input, output, tool charges, and any other costs for the moment.
Without caching, processing that prefix ten times would cost:
With one 5-minute write and nine successful reads, that same prefix would cost:
This is an illustration, not a task quote. The real bill also includes any uncached prompt growth, the model's output at $50 per million tokens, tool and search costs where applicable, refreshes or rewritten cache entries, and requests that miss the cache. A one-hour cache write would also cost more than the five-minute write in this example.
Anthropic estimates that the lower cache-read rate can reduce typical Fable 5.1 workload cost by about 25%, and highly agentic workload cost by up to approximately 45%. Those are Anthropic estimates, not a guarantee for a particular harness. Your own cache-hit and accepted-result data should decide the route.
Why Fable 5.1 may fit long-horizon coding agents
Coding agents often have a cost shape that is different from one-shot chat. They may read a large stable context, call a tool, append the result, reason again, and repeat the loop many times. Cheap cache reads matter most when the same useful prefix survives across those turns.
Based on Anthropic's specifications and stated positioning—not an Agent.Space benchmark—Fable 5.1 is a reasonable candidate for tasks such as:
- a repository-wide investigation with conflicting evidence across many files;
- a multi-stage implementation that must keep architecture constraints and acceptance checks active for hours;
- difficult code review or performance analysis in which a shallow answer creates expensive repair work;
- long-running research that repeatedly refers to the same source set; or
- a delegated task whose output can be verified with tests, diffs, logs, or a clear review rubric.
Anthropic describes Fable 5.1 as designed for ambitious coding work, including codebase-spanning features, code review, performance work, and multi-day autonomous sessions. Treat that as the vendor's capability claim. It is a reason to test the model, not proof that it will outperform a cheaper model on your repository.
The 1M context window can help when the relevant working set is genuinely large. It should not become permission to send an unfiltered codebase. Better retrieval, a smaller task boundary, and explicit source priority can improve both cost and reliability.
When Fable 5.1 is probably the wrong default
The model is less compelling when the work is bounded, easy to verify, and unlikely to benefit from a long reasoning chain. Examples include a known text replacement, a small format correction, or a deterministic update with narrow file scope.
It may also be a poor default when:
- interactive latency matters more than maximum reasoning depth;
- most prompt content changes every turn, so cache reuse is low;
- the task generates a large amount of output, which is billed at $50 per million tokens;
- the harness cannot expose reliable cache or usage data; or
- there is no acceptance test strong enough to distinguish a good result from a persuasive-looking one.
For these tasks, compare Fable 5.1 with Opus 5, Sonnet 5, or another compatible model under the same conditions. The useful measure is total cost per accepted result, including retries and reviewer time—not the model's family name or the price of one request.
The broader coding-agent model selection framework explains how to filter by compatibility, task, context, cost, and tools before running a controlled trial.
If the decision has narrowed to the two newest premium routes, use the separate GPT-6 Astra vs Claude Fable 5.1 comparison. It leaves each model's detailed pricing and cache rules in its dedicated guide, then compares only the task, harness, evidence, and accepted-result questions that actually change the choice.
Model versus harness: where Claude Code fits
Claude Fable 5.1 is a model. Claude Code is an Agent harness: the execution layer that gathers context, invokes tools, manages permissions, edits files, and organizes the work around a model. A Runtime and product surface add further constraints. These layers should not be treated as one interchangeable product.
If that distinction is new, start with the guide to why an Agent harness and a model are separate choices.
Anthropic's current Claude Code model configuration documentation says Fable 5.1 requires Claude Code v2.1.255 or newer and must be selected explicitly; it is not the default model for any account type. Actual access still depends on what the server reports for the organization and account. The separate Claude Code tutorial covers the picker, command-line selection, effort controls, and access checks without turning this model guide into a CLI manual.
Before routing work, confirm the exact combination:
- the harness and surface you will use;
- the provider and exact model ID;
- the context and effort settings the integration actually supports;
- how caching is enabled and measured; and
- who receives the bill.
Run a controlled evaluation before routing real work
Choose one representative task that is difficult enough to expose the expected benefit but small enough to review. Then keep the starting code, prompt, permissions, tools, and acceptance checks as stable as practical across candidate models.
Record at least:
- exact model and harness versions;
- input, cache-write, cache-read, and output tokens;
- cache-hit or miss information by turn;
- elapsed time and number of repair cycles;
- tests or checks passed; and
- reviewer time required before acceptance.
Do not judge the model only by the first answer. A slower, more expensive request can still win if it avoids several repair runs. A strong-looking result can still lose if it expands scope, ignores constraints, or leaves evidence unverified.
For long tasks, also test recovery. Interrupt a run, provide a contradictory tool result, or resume from an explicit handoff. The important question is whether the workflow reaches a reviewable result while preserving the task's constraints—not whether it produces the longest explanation.
Frequently asked questions
How much does Claude Fable 5.1 cost?
Through the Claude API, standard input costs $10 per million tokens and output costs $50 per million tokens. A 5-minute cache write costs $12.50, a 1-hour cache write costs $20, and a successful cache read costs $0.25 per million tokens. Subscription and Console billing are separate paths, so check the product that will receive the bill.
What is the Claude Fable 5.1 context window?
The official model specification lists a 1-million-token context window and a 128,000-token maximum output. That is capacity, not a recommendation to send a whole repository without retrieval or filtering.
How do I select Fable 5.1 in Claude Code?
Use /model fable in an active Claude Code session or launch it with claude --model fable. Fable 5.1 requires Claude Code v2.1.255 or newer, and access depends on the organization and account. The step-by-step Claude Code guide also explains effort settings and what to check when the model is missing.
Are Claude Fable 5.1 and Claude Code the same product?
No. Fable 5.1 is the model that produces the response. Claude Code is the Agent harness that gathers context, calls tools, manages permissions, and sends work to a selected model. Model access, context handling, effort controls, and billing still depend on the exact Claude Code route and account.
What to do in Agent.Space
Agent.Space is independent from Anthropic. This article does not claim that Claude Fable 5.1 is currently available on Agent.Space or compatible with a particular Agent harness. Model availability, provider routes, and supported combinations can change.
If you want to evaluate a model in a persistent cloud Workspace, choose the Agent harness first and then use the live model selector as the source of truth. The Claude Code on Agent.Space guide shows the surrounding workflow and its product boundaries.
When the live selector offers a compatible route, begin with one bounded, verifiable task. Start an Agent.Space Workspace, confirm the current model and pricing shown for your account, and expand only after the result justifies the added cost.