FULL STORY

Claude Fable 5.1: From Leak to Launch and Backlash

After leaking on Bedrock, Claude Fable 5.1 launched with Mythos 5.1 and cheaper caching, but users quickly revolted over rate limits, cache billing, and subscription allowances.

2026-09-01 ~ 2026-09-03 · 3 episodes · 186 posts

Episode 1 · Claude Fable 5.1 Surfaces on Bedrock and in Official Docs, Launch Imminent (2026-09-01, 5 posts)

Multiple independent signals indicate that Anthropic's Claude Fable 5.1 is in final pre-launch preparation. No official announcement has been made, but a release appears imminent.

Confirmed

  • A slug for Fable 5.1 has appeared in the Amazon Bedrock API: querying the ID returns a 404 "Model not found", while invalid slugs (e.g., Opus 5.1 or random strings) return a 400 — a pattern typically seen during pre-release registration (related by @koltregaskes and @haider1).
  • Anthropic's official Claude support documentation has begun mentioning Fable 5.1, seen as a sign of an approaching launch (@kimmonismus).

Unconfirmed

  • Timing: one report claims the release is expected "tomorrow" (relayed by @danielmac8), but there is no official confirmation; this remains rumor.
  • Features: reports suggest the new version will improve long-running agent tasks, coding capabilities, and multi-agent orchestration (@johnseach, @danielmac8); developer ClaudeDevs mentioned upcoming changes aimed at letting users get more value from Claude. These are secondhand claims not verified by Anthropic.

Why it matters

  • Fable 5, launched in June, targeted long-running agents and large coding tasks and performed strongly in evaluations. If 5.1 continues that positioning with stronger agent orchestration and coding, it will directly affect tooling choices for users and enterprises building agents on Claude.
  • The combination of Bedrock pre-registration and doc leaks is a classic pre-launch signal, leading the community to widely expect a release within days.

Episode 2 · Anthropic Releases Fable 5.1 and Mythos 5.1 with 75% Cheaper Caching (2026-09-02, 174 posts)

On September 2, Anthropic released Claude Fable 5.1, the latest upgrade to its strongest model family, specialized in complex, long-running tasks and scientific research, alongside Claude Mythos 5.1, available only through trusted-access programs for cyber defense and life sciences. Both share the same underlying model with different safety tiers; Mythos adds enterprise-grade zero-retention protections. Anthropic also reset all users' 5-hour and weekly usage limits, disclosed changes to thinking-block management in the Messages API to prevent distillation, and published an official intro video.

Confirmed

  • Pricing: input/output unchanged at $10/$50 per million tokens; Prompt Cache reads down 75% (to $0.25); typical workloads expected 25% cheaper, heavy-context agentic workflows up to 45%
  • Benchmarks: Terminal-Bench-Science 0.1 at 52.6% (double the previous generation); kimmonismus reports big jumps on Terminal-Bench 4.0, Science-Bench and HLE; results posted by Anthropic engineer Felix Rieseberg show incremental gains, and Justwalkingthru3 argues the improvement is modest and OpenAI's upcoming Astra may soon surpass it
  • Hands-on: WorldofAI demonstrated Fable 5.1 building a full game from a single prompt, creating Three.js experiences, and replicating complex physics engines, benchmarked against Claude Opus 5
  • Value: kimmonismus notes Fable 5.1 High clearly beats Sol 5.6 Max on Cursor Bench at lower cost; official claims it beats Fable 5 at lower cost on lower effort tiers
  • Contested costs: MineBench maintainer ENTAlam found total cost for 15 builds rose from $54.93 to $147.55 (nearly 3x) with doubled reasoning time despite unchanged API prices; WonderFactory showed a $3.69 single-task cost exceeding Fable 5; developer thispageleftblank, with cache reads at 78% of spend, estimates 57% total savings
  • UX: a Reddit user (relayed by KrayeBaby) found multi-turn output more coherent than Opus 5 and Fable 5, while acknowledging remaining flaws

Why it matters

Fable 5.1 is a positioning move ahead of OpenAI's Astra; Wes Roth noted the timing, and kimmonismus calls it a strong release OpenAI must answer. The deep cache price cut reshapes costs for heavy users, but MineBench-style tests suggest the model may burn far more tokens through longer thinking—whether you actually save depends heavily on your workload.

154 more related posts →

Episode 3 · Claude rate limits and caching controversies erupt after Fable 5.1 launch (2026-09-02, 7 posts)

After Fable 5.1's launch, users raised concentrated complaints about Claude's rate limits, caching, and subscription benefits: multiple Max heavy users reported abnormally fast quota depletion and suspected a caching bug in the new Claude Code, while official advice to use /compact less was contradicted by subscriber experience; several users consider GPT-5.6 better overall.

Confirmed

  • @doodlestein reported abnormal usage stats in Fable 5.1, suspecting a caching bug in the new Claude Code: merely auditing a project exhausted the 5-hour quota of 28 Max 20x accounts—unprecedented, per the user. @peterjliu relayed identical feedback from multiple developers about misleading quota consumption.
  • @kimmonismus and @cyruszei both complained that Claude's rate limits make the model nearly unusable and that Fable's auto-continuation feature is buggy; both prefer GPT-5.6 and note subscriptions did not benefit from Fable 5.1's efficiency gains.
  • With Fable 5.1 cache prices sharply reduced, officials advised against frequent /compact usage; @dotey's tests showed not compressing underperforms keeping 200k auto-compaction, and cache discounts apply only to API users—not subscribers, whose cache TTL is 1 hour; using /compact after cache expiry raises costs.

Unconfirmed

  • The caching bug remains user speculation with no official confirmation or fix from Anthropic; the exact cause of faster quota depletion (pricing change, accounting change, or bug) awaits an official response.
  • Claims that efficiency gains did not reach subscriptions are subjective user reports without official data.

Why it matters

  • Rate limits plus a suspected caching bug directly hit the availability and cost of high-tier subscriptions like Max 20x, potentially driving away heavy developers.
  • The gap between official advice and subscriber reality (short cache TTL, no discount) exposes inconsistent benefits between the API and subscription systems.