Claude Fable 5.2: Benchmarks, the 20% Price Cut and What to Check Before You Switch

Claude Fable 5.2 costs 20% less per token than Fable 5.1, and Anthropic reports large gains on its own coding and agent benchmarks. The API model ID is claude-fable-5-2. Fable 5.1 is not being retired, and four API changes can break code that worked on 5.1.

We covered the launch, the Mythos split and the plan offers in our overview of Claude Fable and Mythos 5.2. This post goes deeper on Fable 5.2 alone. It covers the benchmarks Anthropic reports, what the price cut does to a real bill, what changed since Fable 5 and Fable 5.1, where the model runs, and the checks to run before you change the model ID. All prices and dates are as of November 10, 2026.

The short version: Claude Fable 5.2 is cheaper per token and, by Anthropic’s numbers, stronger than Fable 5.1 on most tests. Your bill only drops if your token counts stay about the same, and some of Anthropic’s own prompt advice raises output. Claude Opus 5.5 still costs half as much. If your code builds its own message history, test it against the prefix check before you switch.

What changed in Claude Fable 5.2 from Fable 5.1

Anthropic’s launch announcement and its What’s new in Claude Fable 5.2 page list these changes:

• Price: $8 per million input tokens and $40 per million output tokens, down from $10 and $50.
• Speed: output arrives up to 30% faster than on Fable 5.1, per the announcement.
• Token use: the announcement says Fable 5.2 finishes hard tasks with fewer tokens than Fable 5.1 at medium effort.
• Fewer blocks: in early testing, about 55% fewer blocked requests than Fable 5.1, with about 60% fewer in cybersecurity and 40% fewer in biology.
• Fast mode: new for Fable 5.2, at $16 input and $80 output, for up to 2.5 times faster output.
• Stricter API rules: the prefix check now applies to every account, and thinking blocks only work in the account that produced them.

Some things did not change. The 1M-token context window, 128k max output, always-on adaptive thinking and the API default effort of high all carry over. The migration guide says the tokenizer matches Fable 5. Blocked cyber requests still get an answer from Claude Opus 4.8, and blocked biology requests from Claude Opus 5.

How Fable 5.1 changed Fable 5

Teams still on the June model should know what 5.1 already changed. Anthropic’s API release notes for September 1 list three items.

Cache reads fell from $1 to $0.25 per million tokens, while input and output stayed at $10 and $50. Forced tool use (tool_choice set to any or tool) started returning a 400 error. The prefix check arrived for accounts created on or after August 31, 2026.

So a jump from Fable 5 straight to 5.2 picks up all of it at once. We covered the 5.1 release in Claude Fable 5.1 and the original in Claude Fable 5.

The benchmarks Anthropic reports for Claude Fable 5.2

Every number below is vendor-reported. They come from Anthropic’s Fable product page. Anthropic says scores are with safeguards on and at each model’s highest effort, unless it marks otherwise.

The same table has two results where Fable 5.2 does not lead. Google’s Gemini 4 Carbon scores 53.9% on AutomationBench. Claude Opus 5.5 has a higher Elo rating on AA-Briefcase, 1800 against 1785.

The system card adds SWE-bench Pro: 93.0% for Fable 5.2, 81.2% for Fable 5.1 and 89.9% for Opus 5.5. It says Fable 5.2 scored highest on 14 of the 16 results in its main capability table.

The announcement also makes cost claims. At high effort, it says Fable 5.2 matches Opus 5.5 at max effort on FrontierCode 2.0 for about half the cost per task. These are Anthropic’s own tests, so treat them as directional. We compare Fable 5.2 with OpenAI’s GPT-6 Astra in a separate post this week.

What the 20% price cut does to a real bill

Here is a worked example using Anthropic’s pricing page. Picture a document-review agent that runs 1,000 times a month. Each run sends 10,000 new input tokens, reads 40,000 tokens from the prompt cache and writes 4,000 output tokens. We left cache writes out to keep the math simple.

For Fable 5.2, the math is $0.08 for new input, $0.008 for cache reads and $0.16 for output. Output is about two thirds of the bill. That is why the output rate matters most.

Fast mode doubles the base rates, and the pricing page says caching multipliers stack on top. US-only inference adds 10% to every token category. Batch jobs cost half, at $4 input and $20 output.

Two more points from the table. Opus 5.5 is about half the price of Fable 5.2 on the same tokens. And the cut only holds if your token counts hold. For a refresher on cache math, see our prompt caching explainer.

Why your token count can cancel the cut

In the example above, Fable 5.2 stays cheaper than Fable 5.1 unless it writes about 39% more output tokens per run. That sounds like a lot. Anthropic’s own Fable 5.2 prompting guide shows how you could get there.

The guide reports that a suggested "finishing work" prompt block raised output tokens about 30% in Anthropic’s replay. A code ownership sentence raised output to about 1.5 times. A progress update block added about 20%. Each one fixes a real behavior, so the trade may be worth it. Measure it first.

Effort is the other lever. The guide says the model thinks before almost every reply at high and above. It also says lines like "answer directly without deliberating" do not stop thinking at medium and above. Lower the effort level instead.

We saw the same pattern with Fable 5.1. Benchmarks disagreed on cost because cache hit rates and output volume differed. Our Fable 5.1 cost post walks through it.

API model ID, lifecycle and where Claude Fable 5.2 runs

The model ID is claude-fable-5-2 on every platform except Amazon Bedrock, where it is anthropic.claude-fable-5-2. The What’s new page lists these platforms:

• The Claude API, for all customers.
• Amazon Bedrock and Claude Platform on AWS.
• Claude Models on Google Cloud, both Anthropic managed and Google managed.
• Claude in Microsoft Foundry, on Anthropic infrastructure.

Fast mode runs on the Claude API, both Google Cloud offerings and Anthropic-hosted Foundry deployments. On Foundry it uses a separate model, claude-fable-5-2-fast. On the Claude API you set speed: "fast" on the request, per the release notes.

In the Claude apps, the Fable page lists the Pro, Max, Ultra, Team and Enterprise plans. Our earlier post on Fable usage limits explains how Fable is metered by plan.

Nothing was deprecated. Anthropic’s model deprecations page lists Fable 5.1 as active, with retirement not sooner than September 1, 2027. Fable 5 is not sooner than June 9, 2027, and Fable 5.2 not sooner than November 9, 2027. The models overview table now shows Fable 5.2 in place of Fable 5.1, but the deprecations page is the one that sets dates.

The migration guide notes two capacity points. Fable 5.2 has its own rate limit, separate from the one Fable 5.1 and Fable 5 share. It also does not support Priority Tier.

Migration checks before you switch to Claude Fable 5.2

Work through these in order:

1. Prefix check. If your code edits the system prompt, the tools or an earlier message, then sends back a thinking block, the request returns a 400 error. Claude Code, claude.ai and the Agent SDK are not affected. The migration guide suggests a test run with the thinking-binding-controls-2026-08-01 beta header and prefix_mismatch_behavior: "drop_block". Fix each mismatch it logs, even if 5.1 never showed the error.

2. Model switches. If a router moves a conversation from Fable 5.2 to another model, that model loses Fable 5.2’s reasoning. The guide warns this can raise cost and latency on the next turn.

3. Thinking and prompts. Remove thinking: disabled, budget_tokens, "think less" lines and hard token budgets. Rerun your effort sweep, starting at high.

4. Caching. Keep history append-only and keep the system prompt and tools fixed for the session. The guide says this keeps the cache warm. The minimum cacheable prompt is 512 tokens.

5. Tools. Move computer use to computer_toolset_20260801, and set the advisor model to claude-fable-5-2.

6. Retention. Fable 5.2 still requires 30-day data retention unless Anthropic expressly authorizes zero data retention.

One note from the system card matters for agent builders. In prompt injection tests, Fable 5.2 falls back to Opus 4.8 more often than recent models, which raises its attack success rate there.

Similar moves at OpenAI and Google

Anthropic’s Fable page scores Fable 5.2 against OpenAI’s GPT-6 Astra and Google’s Gemini 4 Carbon. We covered OpenAI’s lineup in which OpenAI model to use and Google’s model in Gemini 4 Argon. Check each vendor’s own pricing page before you compare.

Frequently Asked Questions