Claude Fable 5.1 Ships With Identical Input and Output Prices, Only Cache Reads Got Cheaper

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. Fable 5.1 is open to every Claude API customer under the model ID claude-fable-5-1. Mythos 5.1 is limited to participants in Project Glasswing, a program for United States organizations. The Hacker News announcement thread drew 648 points and 604 comments.

Several outlets led with "up to 45% cheaper." Open the pricing table in the platform docs, though, and base input sits at $10 per million tokens with output at $50, numbers identical to Fable 5, released in June. Exactly one line item went down: cache reads.

Cache reads are the only thing that got cheaper

Start with what a cache read is. A long conversation or agent session resends its entire opening every time it makes a request: the system prompt, the tool definitions, and the transcript so far. Prompt caching stores that opening on the server and reuses it on the next request. The rate charged on those reused tokens is the cache read price.

Across other Claude models, cache reads cost 0.1x the base input price. Fable 5.1 and Mythos 5.1 charge 0.025x. With base input at $10, a million cached tokens that used to cost $1 now costs $0.25.

The announcement quotes two savings figures: roughly 25% against Fable 5 on typical work, and up to roughly 45% on complex coding and agent workloads. The arithmetic explains the gap. Savings scale directly with the share of your input tokens that come from cache. A long agent session that rereads the same opening dozens of times approaches 45%. A one-shot call that sends a fresh prompt every time saves nothing.

Cache write prices did not move. A five-minute cache write still costs $12.50 per million tokens and a one-hour write $20. The minimum cacheable prompt length is unchanged at 512 tokens. If your code does not use caching yet, that means you pay the write cost first before the read discount starts applying at all.

Three things that break on the way over

This is not a swap-the-model-ID update. The docs name three behaviors that break when you move from Fable 5.

First, forcing the model to call a tool is now rejected. Sending tool_choice as {"type":"any"} or {"type":"tool","name":"..."} returns a 400.

tool_choice: type "tool" and "any" are not supported for this model.

The token counting endpoint applies the same validation. The default {"type":"auto"} and {"type":"none"} still work. The docs give the reason: this model always has its reasoning turned on, and forcing a tool call makes it skip that reasoning and write the content into the tool arguments instead, which degrades argument quality. Two replacements are offered. If you were forcing a tool to pin the output to a JSON schema, move to strict tool use with strict: true or to structured outputs. If you were forcing it so the tool actually gets called, state in the prompt when the model should reach for that tool.

Second, earlier models cannot read thinking blocks that Fable 5.1 produced. Preservation runs one direction only. A conversation that started on Claude Opus 5 or Fable 5 carries its reasoning intact into Fable 5.1, but going the other way, from Fable 5.1 down to another model, drops that turn's reasoning. If a router or fallback swaps models mid-conversation, the API silently discards blocks it cannot read. Dropped blocks do not count toward input_tokens and are not billed, but without the thinking-binding-controls-2026-08-01 beta header the response never tells you the drop happened.

Third, editing an earlier turn invalidates every thinking block after it. Change the system prompt, the tools array, or any prior message, and the next request is rejected with a 400 carrying The block is bound to a different conversation.

This check has a per-account grace period. It is enforced only on accounts created after August 31, 2026; older accounts merely log mismatches unless you set thinking.block_binding.prefix_mismatch_behavior yourself. Mythos 5.1 does not run this check at all.

All three apply only to code that builds the messages array itself. Claude Code, claude.ai, Claude Managed Agents, and the Claude Agent SDK keep the prefix stable on their own. If you only use the official tooling, there is nothing to change here; this section is for developers calling the API directly.

The official migration guide walks all three through step by step. To find out whether your own code mutates conversation history, the docs suggest setting prefix_mismatch_behavior to "drop_block", running one session, and inspecting input_transformations.

Behavior that changes even if you touch nothing

A separate set of changes alters the output without a single line of code edited. Of the seven items the docs group under "Changed from Claude Fable 5," two hit cost directly.

• Fewer parallel tool calls. Where Fable 5 would batch several independent read operations into one call, Fable 5.1 sometimes issues them one per turn. Answer quality holds, but round trips, tokens, and wall-clock time all go up. The docs say this shows up most in hand-built coding agents and in setups that move between a command line and an editor.
• Full file rewrites for small edits. When editing a text file, the model now more often rewrites the whole thing instead of changing just the affected region. The result is usually the same, but it spends more output tokens, and output runs $50 per million.

The other five change the character of what comes back. Intermediate progress reports during long tool runs are less frequent. At low effort the model calls search tools less often and answers from memory more often. Prose is denser, with longer sentences and fewer paragraph breaks. In chat it uses bold, headers, and lists less than earlier models. And when summarizing a document it more often lifts phrasing from the source without quotation marks. That last one matters if you pipe summaries straight to an external audience: quotation marking now needs its own check.

On benchmarks, per the announcement, Fable 5.1 scores 55.8% on the agentic coding benchmark Terminal-Bench 4.0, while Mythos 5.1, with safeguards removed, scores 60.9%. OSWorld 2.0, a computer-use benchmark, comes in at 77.9% partial and 41.7% strict, and CursorBench 3.2.0 at 73.4%. Anthropic says cybersecurity-related false positives in Claude Code dropped roughly 60%.

Can you use it today

Fable 5.1 is available worldwide, Korea included. Neither the Claude API, claude.ai, AWS, Google Cloud, nor Microsoft Foundry lists a regional restriction. Mythos 5.1 is open only inside Project Glasswing, which admits United States organizations only.

The Fable 5 promotional credits that expired on July 19 do not carry over. The support docs state plainly that Fable 5.1 was never in scope for them.

Data retention terms need checking too. Both models retain data for 30 days and cannot run under a zero data retention configuration unless Anthropic grants a specific exception. For teams whose internal policy requires no retention, that single line decides adoption before anything else does.

Outputs carry the statistical text watermark applied to models released after August 2. Images and video retrieved through the Files API carry C2PA Content Credentials signatures. The detection API that reads those watermarks opened in private preview with this release, but access is restricted to regulators, law enforcement, journalists, fact-checkers, independent researchers, educational institutions, and EU civil society organizations, so ordinary developers still cannot verify the watermark on their own output.

One more thing: the docs do not recommend Fable 5.1 as a default. They say most workloads should start on Claude Opus 5, and to move up to Fable 5.1 only when Opus 5 at xhigh or max effort still falls short on your own evaluations.

If your team runs long agent sessions on the Claude API, pull last month's bill and calculate what percentage of your input tokens were cache reads. That ratio is the size of the discount you get, and it tells you whether you land nearer 25% or 45%. If the number comes back low, turning on prompt caching will save you more than switching models will. And if you do decide to move, grep the codebase for tool_choice and clear out any and tool first, or your very first request comes back a 400.