Claude Fable 5.2
Claude Fable 5.2 is Anthropic's top generally available model: ID, 1M context, five effort levels, $8/$40 pricing and what independent tests say about cost.
Claude Fable 5.2 is Anthropic’s most capable generally available model as of November 2026. Anthropic released it on November 9, 2026 with the API ID claude-fable-5-2, a 1M-token context window, 128K tokens of output and a price of $8 per million input tokens and $40 per million output tokens. That is cheaper per token than Fable 5.1, but independent tests show it can cost more per task.
Specs at a glance
| Claude Fable 5.2 | |
|---|---|
| Model ID | claude-fable-5-2 (Anthropic model page) |
| Context window and max output | 1M tokens and 128K tokens (model page) |
| Thinking | Adaptive, always on; it cannot be turned off (model page) |
| Default effort | high (model page) |
| Knowledge cutoff | June 2026 (model page) |
| Earliest retirement | November 9, 2027 (model page) |
Anthropic’s docs pitch it for demanding reasoning and long-horizon agentic work, and recommend starting with Claude Opus 5.5 for most workloads. Move to Fable 5.2 when your own evals on Opus 5.5 at higher effort still fall short.
Price against Fable 5.1
All prices are per million tokens in US dollars, from Anthropic’s pricing page as of November 2026.
| Input | Output | Cache read | |
|---|---|---|---|
| Fable 5.2 | $8 | $40 | $0.20 |
| Fable 5.1 | $10 | $50 | $0.25 |
Anthropic lists Opus 5.5 at $4 input and $20 output. Anthropic’s what’s new page adds that the Batch API is $4 input and $20 output, and that fast mode, which Anthropic’s announcement says is up to 2.5 times faster, costs $16 input and $80 output.
Effort levels
Fable 5.2 accepts five levels through output_config.effort: low, medium, high, xhigh and max, according to Anthropic’s effort guide. The guide says most of the capability gain is already present at high, that xhigh is for the hardest long-running coding and agent work where you have measured a gain, and that at max the model works “without regard to token cost”, running slower and costing more than xhigh “often for only a small gain.”
Two practical notes from the same guide. Set a large max_tokens at high and above, because it caps thinking and reply together. And run a fresh effort sweep instead of carrying settings over from Fable 5.1. Claude apps and Claude Code default to Medium, per the announcement, while the API default is high.
Availability and restrictions
- Platforms. Anthropic’s availability list covers the Claude API, Amazon Bedrock (
bedrock-runtimeendpoint only), Google Cloud, Microsoft Foundry and Claude Platform on AWS. - Mythos 5.2. Anthropic’s model page says Mythos 5.2 offers the same capabilities only to organizations verified through programs such as the Cyber Verification Program. The announcement describes it as the same underlying model without certain cyber and biology safeguards.
- Safeguards. Per the announcement, Fable 5.2 blocks high-risk cyber activity and most research biology tasks, and blocked queries fall back to Opus 4.8 for cyber and Opus 5 for biology. A refused API request returns HTTP 200 with
stop_reason: "refusal", and Anthropic’s docs describe server-side fallback to retry it. - Data retention. Anthropic’s model page says Fable 5.2 requires 30-day data retention and is not available under zero data retention unless Anthropic expressly authorizes it.
- API changes from 5.1. Anthropic’s what’s new page lists four breaking changes: a 400 error when earlier conversation history before a thinking block has been edited, Fable 5.2 thinking blocks that only Fable 5.2 and Mythos 5.2 can read, the older
computer_20251124tool failing everywhere except Bedrock, and a Fable 5.2 executor that accepts only a Fable 5.2 or Mythos 5.2 advisor. Forced tool use (tool_choiceofanyortool) also returns a 400 error, but Anthropic notes this is not new if you come from Fable 5.1. If you build on the Claude Agent SDK, Anthropic says the history change does not affect you.
What Anthropic reports and what others measured
Vendor numbers and independent numbers answer different questions, so keep them apart.
Vendor-reported. On Anthropic’s product page, Fable 5.2 scores 49.1% on FrontierCode 2.0 at Xhigh effort against 45.5% for Opus 5.5, and 69.9% on FrontierSWE v2 against 62.3%. The page notes that “all scores show each model’s highest-available effort unless otherwise noted”, so the comparisons are not matched on cost.
Independent results, as of November 2026:
| Source | Result for Fable 5.2 |
|---|---|
| Artificial Analysis (model page) | Intelligence Index 65 at Max effort, ranked first of 95 models at the time read |
| Vals AI (model page) | Vals Index 71.31%, ranked first, at $17.94 per test |
| SemiAnalysis (GDPx) | 16.4% score, first on the leaderboard, at $88.15 per task |
Benchmarks differ in tasks and scoring, so read the Artificial Analysis Intelligence Index as one measure among several. SemiAnalysis’s GDPx page lists Opus 5.5 at 11.9% and Fable 5.1 at 8.5%.
The cost-per-task caveat
A lower token price does not mean a lower bill. SemiAnalysis wrote that the average cost per task for Fable 5.2 is 31% higher than for Fable 5.1, $88.15 against $67.16, despite a per-token price 20% lower. Its GDPx page shows why: 1.23M output tokens per task for Fable 5.2, against 418K for Fable 5.1 and 549K for Opus 5.5. SemiAnalysis wrote that Fable 5.2 is “very token inefficient for certain tasks” because of its persistence and its readiness to spawn subagents, and said that on its materials science tasks Fable 5.2 used 4.3 times more output tokens than Astra for effectively the same score.
Artificial Analysis shows the same pattern across effort levels. Its model page lists output tokens used to run the Intelligence Index:
| Effort | Index score | Output tokens | Minutes per task |
|---|---|---|---|
| Max | 65 | 275M | 41.5 |
| Xhigh | 62 | 144M | 24.5 |
| High | 62 | 85.3M | 15.2 |
| Medium | 56 | 41M | 6.5 |
| Low | 54 | 34.9M | 5.5 |
Going from high to max scores three points higher and uses about 3.2 times the output tokens, which matches Anthropic’s own advice to reserve max. Anthropic’s what’s new page also warns that subagents “can account for a large part of what a session costs.”
Anthropic’s announcement makes the opposite case for lower effort: Fable 5.2 at High effort scores about 45.3% on FrontierCode against 45.5% for Opus 5.5 at roughly half the cost per task, and at Low effort on Humanity’s Last Exam it matches Opus 5.5’s top score for under a tenth of the cost. Both can be true, since the result depends on the task and the effort level.
OpenAI’s GPT-6.2 and Google’s Gemini 4 are the other current flagship families. To compare models for your own workload, measure tokens and cost per completed task on your own evals, trace each run and set max_turns or a budget so a persistent agent cannot run unbounded.
FAQ
What is the Claude Fable 5.2 model ID?
The Claude API ID is claude-fable-5-2, and Anthropic's model page lists the same ID on Google Cloud, Microsoft Foundry and Claude Platform on AWS. On Amazon Bedrock it is global.anthropic.claude-fable-5-2.
How much does Claude Fable 5.2 cost?
As of November 20, 2026, Anthropic's pricing page lists $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1. Batch requests are 50% off.
Is Claude Fable 5.2 cheaper than Fable 5.1 in practice?
Not always. SemiAnalysis's GDPx leaderboard shows $88.15 per task for Fable 5.2 against $67.16 for Fable 5.1, because it spends far more output tokens. Cost per task depends on the effort level and the workload.
What is Claude Mythos 5.2?
It is a version of Fable 5.2 with the same capabilities, offered only to organizations verified through Anthropic's programs such as the Cyber Verification Program, per Anthropic's model page. It shares Fable 5.2's specifications and pricing.
Which effort level should I start with on Fable 5.2?
Anthropic's effort guide says to start with high, the default, and step up to xhigh only where you have measured a gain. It reserves max for finding the model's ceiling.