Claude Fable 5.2: API, Benchmarks, Pricing and How to Use It

Built for Speed: ~10ms Latency, Even Under Load
Blazingly fast way to build, track and deploy your models!
- Handles 350+ RPS on just 1 vCPU — no tuning needed
- Production-ready with full enterprise support
Anthropic has shipped three Fable models in five months. Fable 5 arrived on June 9, 2026, Fable 5.1 on September 1, and Fable 5.2 on November 9. Each one is the model for the hardest coding, research and agentic work in Anthropic's lineup, so it is easy to read an older guide and use the wrong ID or the wrong price.
This guide covers what Fable is, how the three versions differ, current pricing and benchmarks, what changes when you move to 5.2, and how to run it in production. Figures come from Anthropic's documentation and launch materials, and benchmark scores are Anthropic's own unless noted.
What Is Claude Fable?
Fable is Anthropic's model for long-running, multi-step work: large coding projects, deep research, document and spreadsheet analysis, and agents that use tools over many steps. It is the generally available version of a Mythos-class model. Mythos is the same underlying model with more permissive safeguards, offered only to organizations verified through programs such as the Cyber Verification Program.
At launch, Stripe reported using Fable 5 for a codebase-wide migration in a 50-million-line Ruby codebase that took a day, work estimated at more than two months for a team doing it by hand. That is one customer's account, but it shows the kind of task Fable is built for.
Fable 5 vs 5.1 vs 5.2
Fable 5.2 keeps the 1M-token context window and 128K output limit, uses adaptive thinking that is always on, and has a reliable knowledge cutoff of June 2026. It takes text and images as input and returns text. Anthropic says output is up to 30% faster than the 5.1 models, and an optional fast mode (model ID claude-fable-5-2-fast on Anthropic-hosted deployments) is up to 2.5 times faster at double the price.
Claude Fable 5.2 Benchmarks
These are results Anthropic published with the 5.2 release. Treat them as vendor-reported, and note that one is an internal benchmark. The picture is strong on agentic and coding work and mixed against some rivals.
Two takeaways. The jump from 5.1 is large on the harder coding and agentic tests, and 5.2 beats Opus 5.5 on most of those. But it does not win everywhere: Gemini 4 Carbon leads on AutomationBench and Opus 5.5 edges ahead on AA-Briefcase. Run your own evaluations on your own tasks before committing.
Fable 5.2 Pricing
Standard pricing is $8 per million input tokens and $40 per million output tokens. That is 20% below the $10 and $50 charged for Fable 5 and 5.1. Cache reads cost $0.20 per million tokens, a 5-minute cache write costs $10 and a 1-hour write costs $16. The Batch API takes 50% off, bringing batch prices to $4 and $20. Fast mode costs $16 input and $80 output.
Here is where it sits among current Claude models:
Anthropic's own guidance is to start with Opus 5.5 for most workloads and use Fable 5.2 for demanding reasoning and long-horizon agentic work, or when Opus 5.5 at higher effort still falls short on your evals. Cost control with a premium model is mostly a routing problem, which we cover in our guide to LLM cost optimization.
Moving From Fable 5.1 to 5.2
The model ID change is one line, but Anthropic lists several behavior and API changes worth checking before you switch:
- Thinking is always on. Setting thinking to disabled, or giving it a token budget, returns a 400 error. Use the effort parameter instead. It has five levels, from low to max, and defaults to high.
- Forced tool use is rejected. Setting tool_choice to any or to a specific tool returns a 400. Use auto with strict tool schemas, or JSON outputs.
- Leave sampling parameters alone. Omit temperature, top_p and top_k, since some values return errors.
- Thinking blocks are bound to the model. Only Fable 5.2 and Mythos 5.2 can read 5.2 reasoning blocks. If a router or fallback switches models mid-conversation, that reasoning is dropped. Editing the system prompt, tools or earlier messages can also trigger a 400 error.
- Computer use and the advisor tool changed. The old computer use tool returns a 400 outside Amazon Bedrock, and a 5.2 executor accepts only a 5.2 advisor.
Behavior also shifts a little. At higher effort the model thinks before almost every reply, may end long tasks early with a list of remaining work, and delegates to subagents readily, which can make up a large share of session cost.
Safeguards, Fallbacks and Data Retention
Fable ships with safety classifiers for areas such as cybersecurity and biology. When one declines a request, the API returns HTTP 200 with a stop reason of refusal and a details object, rather than an error. Anthropic offers a server-side fallback (in beta) that retries the request on a fallback model. For Fable 5.2 the permitted targets are Opus 4.8 and Opus 5. Refusal billing depends on the category, and every refusal counts against your rate limits.
Fable 5.2 also requires 30-day data retention, and it is not available under zero data retention unless Anthropic expressly authorizes it. If you handle regulated data, check that with your security team before rollout.
How to Access the Fable API
Fable 5.2 is available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. The model IDs are claude-fable-5-2 on the Claude API, Google Cloud and Foundry, and global.anthropic.claude-fable-5-2 on Bedrock. To call it directly, create a key in the Claude Console and send requests with the model ID, as described in Anthropic's API quickstart.
Direct access is fine for a prototype. Once several teams use a premium model, you need to control who can call it, cap spend, handle refusals and fallbacks consistently, and see costs in one place. That is what a gateway is for.
Calling Fable through the TrueFoundry AI Gateway
With TrueFoundry's AI Gateway you connect your Anthropic account once, enable the Fable model, and use access control to decide which teams and virtual keys can call it. Applications then use one OpenAI-compatible endpoint, so swapping between Fable, Opus, Sonnet or another provider is a change of model name. Per-team budgets and rate limits keep spend predictable on a model priced at $8 and $40, and every call is logged with tokens, latency and cost.
Because thinking blocks are tied to the model that wrote them, a gateway also lets you decide deliberately when a conversation may change models, instead of finding out through dropped reasoning. If you want to see it with your own workload, book a demo.
When to Use Fable 5.2
Use Fable for tasks that are hard, long and multi-step: large migrations, complex coding projects, deep research, detailed document analysis, and agents that work through many stages with tools. It also suits jobs where one run can save days of human effort, because the higher token price is small next to that.
Don't use it for everything. Short summaries, simple questions and high-volume routine requests rarely need the strongest model, and Opus 5.5, Sonnet 5.5 or Haiku 5.5 will cost far less. The sensible pattern is to route by task, and to run your own evaluations with an effort sweep before you decide.
Limitations
- Cost. At $8 and $40 per million tokens, long agentic sessions add up, and subagent delegation can multiply usage.
- Refusals. Safety classifiers can decline legitimate requests on sensitive technical topics, so plan for refusals and fallbacks.
- Data retention. The 30-day retention requirement rules out some regulated workloads.
- API strictness. Several parameters that worked on earlier models now return errors.
Final Thoughts
Fable 5.2 is the one to evaluate for hard agentic and coding work: it is cheaper and faster than 5.1 and strong on Anthropic's reported tests. Treat the benchmarks as a starting point, migrate carefully, and put governance in front of it from the start. For the wider picture, see our guide to AI gateways.
TrueFoundry AI Gateway delivers ~3–4 ms latency, handles 350+ RPS on 1 vCPU, scales horizontally with ease, and is production-ready, while LiteLLM suffers from high latency, struggles beyond moderate RPS, lacks built-in scaling, and is best for light or prototype workloads.


Recent Blogs
Frequently asked questions
What is the Claude Fable 5 API model name?
The original Claude Fable 5 API model used claude-fable-5. The current version is Claude Fable 5.1, with the API model ID claude-fable-5-1. It is available through the Claude API and supported cloud platforms, including Amazon Bedrock and Google Cloud.
How much does Claude Fable 5 cost?
Claude Fable 5.1 costs $10 per million input tokens and $50 per million output tokens. Five-minute cache writes cost $12.50/MTok, one-hour cache writes cost $20/MTok, while cache reads cost $0.25/MTok. Batch processing offers 50% lower input and output pricing.
What's the context window?
Claude Fable 5.1 has a 1-million-token context window and supports up to 128,000 output tokens. It accepts text and image inputs and produces text output. The model also uses adaptive thinking, allowing it to adjust reasoning effort for more demanding tasks.
Is Claude Fable 5 better than GPT-5.5?
There is no single benchmark that establishes one model as better for every task. Anthropic's current evaluations compare Fable 5.1 with competing models across coding, reasoning, research, and knowledge work. Results vary by benchmark, so testing both on your workload provides a more useful comparison.
How is Fable 5 different from Mythos 5?
Fable 5 and Mythos 5 share the same underlying model, but have different safeguards and access requirements. The current versions are Fable 5.1 and Mythos 5.1. Fable is generally available, while Mythos is restricted to vetted organisations through trusted-access programmes.
Why did my request return an Opus 4.8 response?
A Fable request can be handled by another Claude model when safety restrictions prevent Fable from completing it. Anthropic supports fallback mechanisms that can retry refused requests on permitted models, including Opus. Therefore, receiving an Opus response does not necessarily mean Fable was unavailable.
How do you use Claude Fable 5?
You can use Claude Fable through the Claude API, supported cloud platforms, or compatible Anthropic products. For new projects, use the current `claude-fable-5-1` model ID. Choose Fable when your workload needs extended reasoning, large context, or long-running agentic workflows.
Can I use Claude Fable for free?
You may access Claude Fable through eligible Claude plans, but availability and usage limits depend on your plan and Anthropic's current terms. API usage is separately billed based on tokens. Check Anthropic's current pricing and plan details before relying on free access.
Can Fable 5 be jailbroken?
No model can be guaranteed completely resistant to jailbreak attempts. Anthropic continuously evaluates and improves Fable's safeguards against attempts to bypass its safety controls. Avoid treating individual jailbreak demonstrations as evidence of a permanent vulnerability, since model behaviour and safeguards can change over time.
Can I use Claude Fable 5 in Claude Code?
Yes. Fable 5.1 can be used for demanding coding and agentic software-development workflows. It is particularly suited to complex coding tasks that require extended reasoning, tool use, and multiple steps. Availability and model-selection options can depend on your Claude Code setup and plan.
What is the difference between Claude Fable 5 and Fable 5.1?
Claude Fable 5.1 is the newer version of Fable 5. It keeps the 1-million-token context window while improving long-running agentic coding, research, and document workflows. It also supports up to 128,000 output tokens and updated caching economics, making it the current version.















.webp)
.webp)
.webp)


.png)
.png)
.png)
.webp)












