Claude Fable 5.1: price, limits, and what changes for developers

Anthropic announced Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026, described by the company as its most advanced models for code and knowledge work. For those who build software, what actually changes is in the wallet, in caching, and in a set of API breaking changes that require code review before switching the model ID.

The key point: input and output token pricing remains the same as Fable 5. What dropped, and dropped sharply, was the cache read price, from US$1.00 to US$0.25 per million tokens, a 75% cut according to the platform documentation. In previous models, cache read cost 0.1x the input price; in Fable 5.1 it now costs 0.025x.

Pricing and Specs Side by Side

The values below come from Anthropic's official documentation and the model overview. Where the source doesn't provide data for the specific version, the cell shows "not provided."

| Item | Fable 5.1 | Opus 5 | Sonnet 5 | Mythos 5.1 | |---|---|---|---|---| | Input (US$/MTok) | 10 | not provided | not provided | 10 | | Output (US$/MTok) | 50 | not provided | not provided | 50 | | Cache write 5m (US$/MTok) | 12.50 | not provided | not provided | 12.50 | | Cache read (US$/MTok) | 0.25 | not provided | not provided | 0.25 | | Context window | 1M tokens | not provided | not provided | 1M tokens | | Max output | 128k tokens | not provided | not provided | 128k tokens | | Anthropic API | yes (claude-fable-5-1) | yes | yes | Project Glasswing only | | Amazon Bedrock | yes | not provided | not provided | via account team | | Google Cloud (Vertex) | yes | not provided | not provided | via account team | | GitHub Copilot | yes | not provided | not provided | no |

Fable 5.1 and Mythos 5.1 share specs and pricing, according to the documentation: in practice, they are the same model with different levels of safeguards. A community analysis cited by Latent Space goes further and claims they are "exactly the same weights," with the difference being the safety classifier threshold and fallback routing. This isn't an official statement from Anthropic, but it fits with Artificial Analysis's note that ~4% of output tokens in its evaluations were served by fallback models (Opus 4.8 or Opus 5).

Reference Conversion to Brazilian Reais

Anthropic charges in dollars, with no pricing table in reais. To size the cost in local currency, you need to apply the exchange rate on the day you check against the US$/MTok values listed in the table above; official sources show no fixed real conversion practiced by the company, and the value floats with the dollar.

What Changes for Developers in Brazil

The cache read cut is the headline for anyone running agents. Agentic loops resend the repository, scratchpad, previous steps, and tool transcripts on every round, and most of those tokens are read from cache. With the read price at US$0.25/MTok, an agent session that rereads 1 million tokens of code context on every step pays about US$0.25 per reread, versus US$1.00 before. Anthropic estimates savings of ~25% on typical workloads and up to ~45% on heavily agentic work.

But there's an important counterweight, and it's what Anthropic's announcement doesn't highlight. According to Artificial Analysis, via Latent Space, Fable 5.1 uses about 1.7x more output tokens per task than Fable 5. The measured net result was a per-task cost ~20% HIGHER: US$3.76/task on Fable 5.1 max versus a lower value on Fable 5. The cache cut saves ~US$1.40 per task, but the extra output tokens eat up the difference. In other words: cheaper where you reread context, more expensive where the model writes a lot. It's worth measuring your own case before assuming it got cheaper.

For those without an international card, the practical route is running it through clouds already billed in reais: Fable 5.1 is available on Amazon Bedrock (as anthropic.claude-fable-5-1), on Google Cloud/Vertex (claude-fable-5-1), on Microsoft Foundry, and on GitHub Copilot. In Copilot it arrives for Pro+, Max, Business, and Enterprise plans, selectable in the model picker of VS Code, Visual Studio, JetBrains, Xcode, Copilot CLI, and coding agent.

Watch Out: Data Retention and Breaking Changes

A detail that changes the compliance decision: unlike other Claude models in Copilot, Fable 5.1 requires data retention by default to run Anthropic's safety classifiers, and the policy comes turned off by default for Business and Enterprise admins. Zero data retention (ZDR) is only available for eligible companies, on a temporary basis, until the end of the year, while Anthropic rolls out Enterprise Frontier Safeguards (EFS). Anthropic states that retained data is not used to train its models.

On the code side, swapping claude-fable-5 for claude-fable-5-1 isn't painless. The documentation lists three breaking changes:

model = "claude-fable-5"    # before
model = "claude-fable-5-1"  # after

• Forced tool use is not supported. tool_choice with {"type": "any"} or {"type": "tool", ...} returns 400. Use {"type": "auto"} with strict: true or structured outputs.
• Thinking blocks are one-directional. Fable 5.1 reads the thinking from previous models, but no previous model reads its own. Routers and fallbacks that switch models mid-conversation lose the reasoning.
• Editing previous turns invalidates thinking blocks. Treat the conversation as append-only. The check is mandatory for accounts created on or after August 31, 2026.

There's also a behavior change that catches anyone with an agent in production: Fable 5.1 tends to make one tool call per turn where Fable 5 fired several in parallel, costing more round trips. The documentation suggests adding an explicit batching instruction to the prompt.

What the Announcement Doesn't Say

The published benchmarks (Terminal-Bench 4.0, Humanity's Last Exam, CursorBench) are Anthropic's own. TechCrunch sums up the release as "cheaper and less restrictive": the "less restrictive" part refers to reducing false positives from the safeguards (60% fewer in cybersecurity, according to Anthropic), including allowing the model to be used for discovering vulnerabilities, though not for developing exploits. In practice, user reports cited by Latent Space still point to annoying false positives and complaints about rate limits on launch day.

Mythos 5.1, the twin version with safeguards geared toward cybersecurity and life sciences, remains restricted to approved partners in Project Glasswing, via an Anthropic, AWS, or Google Cloud account. For the average developer, what's on the table is Fable 5.1: faster, strong on long, autonomous tasks, with cheap caching for agents, and a per-task cost that demands its own measurement before any promise of savings.