Anthropic shipped Claude Fable 5.1 on September 1, 2026, and the headline is a pricing trick rather than a price cut. Input and output cost what they did on Claude Fable 5; reading a cached prompt is where the money moved.
The useful part got less coverage: Anthropic tells most teams to stay on Claude Opus 5.
What Claude Fable 5.1 is
It replaces Claude Fable 5, released on June 9, 2026, suspended three days later, and restored on July 1.
One model wears two badges. Claude Mythos 5.1 has the same capabilities and the same prices; the difference is safeguards. Fable 5.1 is generally available, while Mythos 5.1 goes only to vetted organisations in Anthropic’s trusted access programs, aimed at cybersecurity and life sciences work.
Anthropic calls it its most capable generally available model, built for “ambitious, long-running, asynchronous work”. Its output carries the company’s statistical watermark, which Anthropic says adds no tokens, no hidden characters and nothing about you.
The published specifications
| Item | Published value |
|---|---|
| Context window | 1M tokens |
| Max output | 128K tokens |
| Thinking | Adaptive, always on |
| Default effort | high |
| Comparative latency | Slower |
| Input and output | Text and images → text |
| Knowledge cutoff | June 2026 |
| Released | September 1, 2026 |
| Retirement | Not sooner than September 1, 2027 |
Adaptive thinking is the only thinking mode here and it is not something you switch on; depth is steered with effort. Five platforms carry it: the Claude API, Google Cloud, Microsoft Foundry and Claude Platform on AWS use claude-fable-5-1, while Amazon Bedrock uses anthropic.claude-fable-5-1.
What changed from Claude Fable 5
| Billing item | Price |
|---|---|
| Input | $10 / MTok |
| Output | $50 / MTok |
| 5-minute cache write | $12.50 / MTok |
| 1-hour cache write | $20 / MTok |
| Cache read | $0.25 / MTok |
| Batch API | 50% off |
Cache reads are the change worth understanding. They now cost 0.025 times the base input price, against 0.1 across the rest of the line-up, which is where the “75% less” comes from. Anthropic measured the effect over four weeks of real August 2026 usage at default effort: around 25% lower cost on typical work, up to roughly 45% on agentic work where reads dominate the bill.
Two caveats. Workloads that must stay inside the United States pay 1.1 times the listed rate. And retention has not relaxed: both models carry 30-day data retention and are unavailable under zero data retention unless Anthropic expressly authorises it. Enterprise Frontier Safeguards, which keeps data in infrastructure the customer controls, arrives in phases from later in the autumn.
What Anthropic has not published
There is no fixed message count on any plan, and the conversion rate between usage credits and tokens is not published anywhere. Neither is anything about a successor, or about what comes after the retirement floor of September 1, 2027.
Where it sits in the line-up
| Model | Context | Max output | Price / MTok | Latency | Default effort |
|---|---|---|---|---|---|
| Claude Fable 5.1 | 1M | 128K | $10 / $50 | Slower | high |
| Claude Opus 5 | 1M | 128K | $5 / $25 | Moderate | high |
| Claude Sonnet 5 | 1M | 128K | $2 / $10 | Fast | high |
| Claude Haiku 4.5 | 200K | 64K | $1 / $5 | Fastest | Not supported |
On the consumer plans the Fable row is not generous: unavailable on Free, usage credits on Pro, half of weekly limits on Max.
Why Anthropic points to Opus 5
The model page says it plainly: for most workloads, start with Claude Opus 5, and reach for Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your evals on Opus 5 at higher effort fall short.
The price gap explains the advice, since Opus 5 costs half as much per token. Fable 5.1 earns its keep when a task runs for hours and re-reads the same large context, which is the case the cache-read change was built for. Plan details are in the pricing and plan comparison.
Two worked examples you can copy
Neither is a vendor sample. Both are inputs the official documentation recommends, written out so you can paste them.
Case one: moving an existing Claude Fable 5 call across. The model id and the effort level are the only things that change.
{
"model": "claude-fable-5-1",
"max_tokens": 16000,
"output_config": { "effort": "high" },
"messages": [{ "role": "user", "content": "Your task here." }]
}Effort decides what this costs, so sweep it rather than guessing. The model supports low, medium, high, xhigh and max, with high as the default. Anthropic’s guidance is that at low or medium it already matches or beats Fable 5 for far less money, and that effort names do not mean the same amount of thinking across generations. At high and above, set a large max_tokens, because it caps thinking and response together.
Case two: long agent turns that go quiet. This model writes fewer user-facing updates between tool calls than Fable 5 did, and the documented fix is one system-prompt line.
Before you start, say in a line what you're about to do; brief updates while you work help the user follow along. Close with a short recap that stands on its own — what you found, what you did, and what's next — so a reader who only sees the last message has the full picture.That line covers both ends, because those short notes arrive as thinking blocks and come back empty under the default thinking.display of omitted.
Migrating from Claude Fable 5
Swapping the id is the easy part. Anthropic flags three breaking changes, and all three trace back to one decision: thinking blocks are now tied to the model and the conversation that produced them.
Three breaking changes
Forced tool use returns an error. Any tool_choice of type any or tool has to go, and schema enforcement moves to strict tool use with auto, or to structured outputs.
Earlier models cannot read these thinking blocks, so a conversation moved onto a model older than 5.1 will not load them.
Editing an earlier turn invalidates the blocks that followed it, which is what happens whenever a harness rewrites history between requests. The replacement is a turn-scoped system message, holding system authority for one turn and then stopping:
{
"role": "system",
"clear_at": "next_user_message",
"content": "Results have landed in your inbox. Check it before running more code."
}Agent loops have one more quirk. Where Fable 5 batched several independent tool calls into a turn, 5.1 can spend a turn on each, costing a round trip every time. The documented nudge goes at the end of the request:
First privately list what you need next; then request every item that doesn't depend on another's result in this one response.Anthropic recommends re-running your evals regardless, because effort levels do not map cleanly between generations. The API getting-started guide covers authentication if you are starting from nothing.
Frequently asked questions
Is Fable 5.1 a new model or a rename? A new model, shipped September 1, 2026, which Anthropic says sets a new standard for coding, knowledge work and long-running problem-solving tasks.
Why use Opus 5 instead? For most workloads it is enough, and it costs $5 / $25 per million tokens against $10 / $50.
Did the price actually fall? Not the list price. Input and output are unchanged; cache reads fell 75% to $0.25 per million tokens, which Anthropic estimates cuts typical workloads by around 25%.
Fable 5.1 or Mythos 5.1? Safeguards, not capability. Same model, same prices, and Mythos is limited to the trusted access programs.
Can I run it under zero data retention? Not by default. Both models carry 30-day data retention and need express authorisation otherwise.
Where do I go next? The Claude Code terminal guide covers the agent workflow this model was tuned for, and the head-to-head series helps if you have not picked a model family yet.
Start with Opus 5, watch your cache-read share, and let your own evals decide when Fable 5.1 is worth the extra token price.