Claude Fable 5.1 Ships Straight to GA at $10/$50 per Million
No preview stage, no waitlist — Anthropic's new top tier went live everywhere on day one. Here's what to re-test before migrating.
Jump to 7 sections
Anthropic shipped Claude Fable 5.1 on September 1, 2026, and did something unusual with it: the model went generally available on every platform the same day, with no preview stage and no waitlist. For a frontier-tier release that is still the exception rather than the rule.
Fable 5.1 sits at the top of the Claude 5 family, above the Opus and Sonnet tiers. It is a point release rather than a new family, which is worth taking literally — the interesting changes in a 5.0 to 5.1 step are usually behavioral rather than architectural.
What it costs
$10 per million input tokens and $50 per million output. That is the same top-tier price point OpenAI set for its own flagship, GPT-6 Astra, and the convergence is not a coincidence — the frontier tier has settled into a recognized price band.
Anthropic's published pricing puts the rest of the ladder beneath it: Opus 5.5 at $4 input and $20 output, Sonnet 5 at $2 and $10, and Haiku 4.5 at $1 and $5. The page also lists context windows up to 1 million tokens depending on model, US-only inference at 1.1x standard pricing, and a fast mode for Opus 5.5 at twice the standard rate.
| Model | Input / 1M | Output / 1M |
|---|---|---|
| Fable 5.1 | $10 | $50 |
| Opus 5.5 | $4 | $20 |
| Sonnet 5 | $2 | $10 |
| Haiku 4.5 | $1 | $5 |
That tenfold spread from top to bottom is the part worth planning around. The routing question — which requests genuinely need the $50 tier — matters more to a monthly bill than the choice of vendor, as we work through in our breakdown of LLM API pricing across the 2026 tiers.
Day-one general availability is the actual news
Shipping a flagship straight to GA changes what a launch means operationally. There is no window in which the benchmark numbers describe a preview build that behaves differently from what you can call, and no staged rollout during which half your team has access and half does not.
It also removes a familiar excuse. When a model is preview-gated, anomalous behavior gets attributed to the preview. A same-day GA release owns its behavior immediately.
The tradeoff is that independent evaluation has had no head start. Nobody outside Anthropic had weeks of preview access to build a considered view before the marketing landed, which means the first month of public commentary is running on launch-day material.
What changed from Fable 5
Anthropic released Fable 5.1 alongside Mythos 5.1, and a NetChoice summary of the announcement says the two are the same model with different levels of safeguards. Fable 5.1 is generally available, while Mythos 5.1 is limited to Anthropic's trusted access programs. The announcement frames the release as addressing customer feedback on price, data retention and safeguards.
The Batch, the newsletter from DeepLearning.AI, reported the specifics. Cached input pricing fell 75% from Fable 5, yet cost per task still rose about 20%. Anthropic says Claude Code sessions should see roughly 60% fewer cyber interventions, though the model still falls back to Opus 4.8 or Opus 5 for penetration testing, exploit writing and scanning compiled binaries.
Data retention is less changed than the feedback might suggest. The Batch reports a 30-day retention requirement for most users, with Zero Data Retention available to eligible enterprises only with Anthropic's authorization.
What to check before migrating
Treat the point release as a behavior change, not a free upgrade. In our experience at Model Drop, the failures that follow a minor version bump cluster in three places, and none of them show up on a capability chart.
Tool-calling format is the first. Argument serialization, parallel call behavior, and how strictly the model adheres to a schema all shift between versions often enough to break a production agent. Re-run your tool tests rather than assuming parity.
Output formatting is the second. Prompts tuned to produce terse JSON against one version frequently get prose wrappers or extra commentary from the next. Anyone parsing with a regex will find out in production.
Refusal and hedging behavior is the third, and it is the least measured. A model that declines slightly more often, or hedges where the previous version committed, changes downstream behavior in ways an aggregate benchmark score will never surface.
Anthropic publishes model documentation and migration notes in its official Claude developer documentation, which is the place to check for deprecation timelines on whatever you are currently running. Launch posts almost never carry that information.
How it compares on cost per task
Rate cards hide how much work a model needs to finish a job. On the Artificial Analysis Intelligence Index, The Batch reports Fable 5.1 scoring 53, tied for first with GPT-6 Astra, with Opus 5 at 51. Per task at max reasoning, Fable 5.1 cost $7.63 and took 12.2 minutes, against $3.26 and 8.2 minutes for GPT-6 Astra and $5.86 and 13.9 minutes for Opus 5.
Those rankings moved within a week, since GPT-6 Astra arrived two days after Fable 5.1 and Artificial Analysis revised its index twice. Treat any single leaderboard snapshot as provisional. The same report notes that Anthropic's published cyber evaluations cover Mythos 5.1 with safeguards off, a configuration most users cannot reach.
Two cost levers are worth pricing in. Anthropic's pricing page lists cache reads at $0.25 per million tokens for Fable 5.1, and says batch processing saves 50%. Also note that the page now lists Sonnet 5.5 and Haiku 5.5 as current models, so the lower rungs in the table above may have changed since publication.
On the 1M context window
The million-token figure is a ceiling on what the API accepts, not a promise about retrieval quality across that span or a statement about what you can afford. Filling a large window on every request is the most reliable way to turn a reasonable rate card into an unreasonable bill.
Rate limits are a separate constraint from window size. A million-token context paired with an account-level tokens-per-minute cap below that number means the request is theoretically expressible and practically unavailable at volume. Those two numbers are published in different places, which is exactly the kind of gap our guide to reading a launch announcement flags.
What Model Drop is watching
Whether independent coding evaluations such as SWE-bench post verified numbers that track Anthropic's own reporting, and how long the day-one GA pattern lasts. If frontier releases keep skipping the preview stage, the gap between announcement and informed judgment gets shorter for everyone.
By Greg Halston, Staff Writer at Model Drop. Reported September 1, 2026. Pricing figures are from Anthropic's published pricing page; capability claims are the vendor's and have not been independently verified by Model Drop.
Model Drop covers AI launches — new models, platforms, features, and tools — for the people who have to decide what to actually ship on.