The new number one on Artificial Analysis costs less per token than the model it replaces: Claude Fable 5.2 scores 65 on the Intelligence Index v5, six points ahead of Claude Opus 5.5, and drops from $10/$50 to $8/$40 per million tokens. The catch is the cost per task: at maximum effort it spends about $26.50 per index task, 2.6 times what Opus 5.5 does.
Anthropic shipped it on November 9, 2026 alongside Claude Mythos 5.2. They are the same underlying model: Fable carries stricter cybersecurity and biology safeguards and is open to everyone, while Mythos 5.2 is limited to organizations approved through Anthropic's verification programs.
What Fable 5.2 brings
| Claude Fable 5.2 | |
|---|---|
| API ID | claude-fable-5-2 |
| Price | $8 input / $40 output per million tokens |
| Fast mode | Up to 2.5x faster, at $16 / $80 |
| Context | 1M tokens, up to 128K output |
| Default API effort | high |
| Cache reads | 2.5% of the input price |
| Retirement | Not before November 9, 2027 |
It is available on the Claude API, Amazon Bedrock (anthropic.claude-fable-5-2), Google Cloud and Microsoft Foundry. Anthropic says it generates output up to 30% faster than Fable 5.1 and that its safeguards trigger about 55% less often than on Fable 5.1. Requests that do get blocked fall back to Opus 4.8 (cybersecurity) or Opus 5 (biology), which matters if you are measuring answer quality in those domains.
How to try it
The minimal call is the usual one with the new ID. Adaptive thinking is always on, so leave headroom in max_tokens:
curl https://api.anthropic.com/v1/messages \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "content-type: application/json" \
-d '{
"model": "claude-fable-5-2",
"max_tokens": 16000,
"messages": [{"role": "user", "content": "Review this runbook and tell me which step is missing."}]
}'On Bedrock the ID becomes anthropic.claude-fable-5-2; Google Cloud and Foundry use the API ID.
Before swapping the ID into an agent that already works with another model, note three incompatibilities from the docs: forcing a tool (tool_choice of type any or tool) returns a 400, temperature, top_p and top_k have to go, and the model requires 30-day data retention, so it isn't available under zero data retention unless Anthropic expressly authorizes it.
Fable 5.2 vs Opus 5.5
The comparison that matters is not with GPT-6 Astra or Gemini but with Anthropic's own model, which the company still recommends as the daily driver. The docs are blunt about it: start with Opus 5.5, and move to Fable 5.2 when your evals on Opus at high effort still fall short.
| Fable 5.2 | Opus 5.5 | |
|---|---|---|
| Price (input / output) | $8 / $40 | $4 / $20 |
| Cache read | $0.20 (2.5%) | $0.20 (5%) |
| Default effort | high | medium |
| Artificial Analysis Index v5 | 65 (Max) | 59 (Max) |
| FrontierSWE v2 | 69.9% | 62.3% |
| OSWorld 2.1 | 89.3% | 87.2% |
| Cost per index task (max effort) | ~$26.50 | ~$10 |
Two numbers here change the math. First, cache reads cost the same on both, $0.20 per million, because Fable applies 2.5% to a rate twice as high. For an agent that rereads a long, stable context turn after turn, the price gap lives in fresh tokens and output, not in the cached context.
Second, Fable 5.2 is more efficient per task than its price suggests. According to The Decoder, at Xhigh effort it uses about 98,000 output tokens per task against 173,000 for Opus 5.5 at Max, and at that level it already scores 62, more than any other model at any level. Anthropic adds that on FrontierCode at High effort it matches Opus 5.5 (45.3% vs 45.5%) at roughly half the cost per task. The jump to $26.50 comes when you let it think as long as it wants.
Against GPT-6 Astra and Gemini 4 Carbon
In Anthropic's table, Fable 5.2 wins almost everywhere: 69.9% on FrontierSWE v2 against 65.5% for GPT-6 Astra and 61.2% for Gemini 4 Carbon, and 89.3% on OSWorld 2.1 against 79.5% and 81.5%. There are two exceptions. On AutomationBench, Gemini 4 Carbon comes out ahead (53.9% vs 51.4%). And on Terminal-Bench 4.0, as measured by Artificial Analysis, Fable 5.2 lands at 60%, level with Opus 5.5 and behind Gemini 4 Carbon (65%) and Kimi K3.1 Pro (67%). For agents that live in a terminal, it is not the clear leader.
Gemini 4 Carbon also isn't publicly available yet, per The Decoder, so at launch the real API race is between Fable 5.2, Opus 5.5 and GPT-6 Astra.
Where it wins and where it loses
It wins on long, hard work: multi-hour agent tasks, reasoning with tools (74.2% on Humanity's Last Exam with tools, against 67.7% for Opus 5.5) and computer use. It also wins on per-token price against Fable 5.1, which cost 25% more.
It loses on latency and on cost per task at maximum effort. Artificial Analysis measures 231 seconds to the first answer token in Max mode and an average of 44 minutes per index task, at 103 output tokens per second. That rules it out for anything interactive unless you lower the effort. And in terminal workflows it doesn't beat Opus 5.5.
Verdict
If you already run Fable 5.1, the switch is easy: more capability, faster output and a 20% lower rate. If you are on Opus 5.5, it is not a default upgrade. It makes sense as an escalation model: the agent tasks Opus can't finish, reviews of large changes, or the orchestrator role Anthropic pitches it for, with effort set to high or xhigh and a per-task budget you actually watch. For everything else, Opus 5.5 is still the sensible pick at half the price.
Sources
- Claude Fable 5.2 and Claude Mythos 5.2, Anthropic, November 9, 2026.
- Claude Fable 5.2, Claude documentation.
- Claude Fable 5.2 and Mythos 5.2 cement Anthropic's AI lead just before its IPO, The Decoder, November 9, 2026.
- Claude Fable 5.2, Artificial Analysis.