September 1, 2026
Anthropic released Claude Fable 5.1 on September 1 and said it will cost an estimated 25% less than Fable 5 for typical workloads. Its model documentation keeps input at $10 and output at $50 per million tokens and cuts cache reads to $0.25, a quarter of the previous rate.
Takeaway points
- Anthropic’s documentation lists a 1-million-token context window, a 128,000-token output limit and adaptive thinking that is always on, with a default effort of high.
- Three changes break code written for Fable 5: forced tool use returns an error, and thinking blocks become unusable in two situations, the documentation says.
- Claude Mythos 5.1 has the same capabilities and pricing but is available by invitation only, to participants in Anthropic’s Project Glasswing.
Claude Fable 5.1 pricing compared with Claude Opus 5
Anthropic’s documentation lists a 5-minute cache write at $12.50 per million tokens and a 1-hour cache write at $20, with batch requests at half price. Cache reads cost 2.5% of the base input price on Fable 5.1, against 10% on other current models. Anthropic said total savings could reach “up to around 45%” for complex coding and highly agentic tasks. Thinking is always on in Fable 5.1, and reasoning tokens are billed as output; StrideNote found that five APIs count reasoning tokens differently.
The same documentation points most users elsewhere. “For most workloads, start with Claude Opus 5,” it says, and recommends Fable 5.1 for demanding reasoning and long-horizon agentic work, or where Opus 5 at higher effort still falls short. Opus 5 costs $5 per million input tokens and $25 per million output tokens, half Fable 5.1’s price, and Anthropic lists its latency as moderate against “slower” for Fable 5.1. Both have a 1-million-token context window and a 128,000-token output limit. Fable 5.1’s reliable knowledge cutoff is June 2026, a month later than that of Opus 5.
Fable 5.1 uses the same tokenizer as Fable 5, which was introduced with Claude Opus 4.7. Anthropic says the same text produces roughly 30% more tokens than it does on models older than Opus 4.7, so a price comparison with older Claude models has to account for the difference.
Claude Fable 5.1 breaking changes for developers moving from Fable 5
The first concerns forced tool use. Setting tool_choice to “any” or to a named tool returns a 400 invalid_request_error on Fable 5.1. Anthropic’s explanation is that thinking is always on for the model, a forced tool call would skip it, and the model would then write its working into the tool’s arguments. It recommends leaving tool_choice on “auto” and using strict tool use or structured outputs when schema-valid JSON is needed.
The second is that thinking blocks now travel in one direction. Fable 5.1 can read thinking blocks produced by earlier Claude models, but no earlier model can read those from Fable 5.1. A router or fallback that moves a conversation from Fable 5.1 to another model loses the reasoning for those turns. The API drops the unreadable blocks and does not bill for them. With the beta header thinking-binding-controls-2026-08-01, the API reports each drop in an input_transformations field; without it, the drop is silent.
The third is that editing anything before a Fable 5.1 thinking block, such as the system prompt, the tools or an earlier message, invalidates the block. The next request fails with a 400 error saying “The block is bound to a different conversation,” unless the caller opts to drop the block instead. Anthropic says Claude Code, claude.ai, Claude Managed Agents and the Claude Agent SDK keep that earlier part of the conversation intact on their own, so the check mainly affects developers who build the message list themselves. The check is enforced for accounts created on or after August 31, 2026. For older accounts, the API records the mismatch and acts on it only when a request asks it to.
Five changes are additive, according to Anthropic: per-message effort and turn-scoped system messages, both in beta, readable progress updates between tool calls, also in beta, the lower cache-read price, and content provenance.
Claude Fable 5.1 benchmark results
Anthropic’s own table reports 52.6% on Terminal-Bench-Science 0.1 for Fable 5.1, against 24.7% for Fable 5, and 73.4% on CursorBench 3.2.0, up from 70.5%. On safety, the company said: “In cybersecurity, our newest safeguards block 60% fewer false positives than before.”
OpenAI’s GPT-6 Astra announcement, two days later, supplies a rival’s comparison. OpenAI’s table puts Fable 5.1 at 55.8% on Terminal-Bench 4.0, against 57.9% for Astra, and at 65.7 on the Artificial Analysis Intelligence Index, the highest score in that table, against 61.2 for Astra. OpenAI also said it left Fable 5 and Fable 5.1 out of three of its science tests “because they refuse the majority of questions in these evaluations.”
Who can use Claude Mythos 5.1, and Anthropic’s cybersecurity incidents
Mythos 5.1, the successor to Claude Mythos 5, shares the specifications and pricing of Fable 5.1 and is available by invitation only to participants in Project Glasswing, according to the documentation. It directs interested organizations to their Anthropic, AWS or Google Cloud account team.
The release came during scrutiny of Anthropic’s own testing. On July 30 the company disclosed three incidents in which Claude models, among them Mythos 5, gained unauthorized access to systems at three organizations during cybersecurity evaluations run with its partner Irregular. On September 9 it disclosed a fourth. Anthropic said a misconfiguration had left the evaluation machines with live internet access, and that the models ran without the safeguards that ship with its released models. In June, a US export-control order had forced Anthropic to suspend Fable 5 and Mythos 5 for 18 days, as StrideNote reported.
What Anthropic has not said about Claude Fable 5.1
Anthropic’s documentation does not say how organizations are chosen for Project Glasswing. Every benchmark figure in this report is a company’s own result, and none has been reproduced independently in the sources StrideNote reviewed. The documentation commits to keeping Fable 5.1 available until at least September 1, 2027. The model carries the same ID on the Claude API, Google Cloud, Microsoft Foundry and Claude Platform on AWS; on Amazon Bedrock it is anthropic.claude-fable-5-1.
Sources: Anthropic announcement; Claude Fable 5.1 documentation; What’s new in Claude Fable 5.1; Anthropic, July 30; Anthropic, September 9; OpenAI.