FABLE cheat sheet: Claude Fable 5.2 and Mythos 5.2 (launched 9 Nov 2026)
Source: Anthropic's announcement and model page, 9 Nov 2026 (https://www.anthropic.com/claude-fable-and-mythos-5-2, https://www.anthropic.com/claude/fable). Benchmark numbers are company-reported.
1. Prices
|
Input / M tokens |
Output / M tokens |
| Fable 5.2 (and Mythos 5.2) |
$8 |
$40 |
| Fable 5.1 (before) |
$10 |
$50 |
| Fast mode (up to 2.5× faster) |
$16 |
$80 |
| Cache reads |
$0.20 |
— |
| - US-only inference: 1.1× pricing. |
|
|
| - Outputs up to 30% faster than Fable 5.1 / Mythos 5.1. |
|
|
- API model name: claude-fable-5-2. Apps: Pro, Max, Ultra, Team and Enterprise plans. Also on AWS, Google Cloud and Microsoft Foundry. |
|
|
2. Benchmarks (highest effort unless noted)
| Test |
Fable 5.2 |
Fable 5.1 |
Opus 5.5 |
GPT-6 Astra |
Gemini 4 Carbon |
| FrontierCode 2.0 (agentic coding) |
49.1% |
32.4% |
45.5% |
42.8% |
41.1% |
| FrontierSWE v2 (agentic coding) |
69.9% |
56.3% |
62.3% |
65.5% |
61.2% |
| AutomationBench (business workflows) |
51.4% |
31.4% |
42.5% |
41.4% |
53.9% |
| OSWorld 2.1 offline (computer use) |
89.3% |
83.5% |
87.2% |
79.5% |
81.5% |
| WANDR relaxed (agentic search) |
82.1% |
68.7% |
72.3% |
— |
— |
| RoboDojo Sim (robotics) |
64.9% |
27.6% |
39.4% |
29.8% |
— |
| Mythos 5.2 (security): CVE Rediscovery 89.6% (Mythos 5.1: 78.3%). |
|
|
|
|
|
| Cost notes from Anthropic: at High effort Fable 5.2 ties Opus 5.5's best FrontierCode score (45.3% vs 45.5%) for about half the cost per task. |
|
|
|
|
|
3. Which model?
- Opus 5.5: start here for most coding and knowledge work (Anthropic's own recommendation).
- Fable 5.2: the hardest, open-ended technical work and long agent runs (days, many subagents).
- Mythos 5.2: the same model with fewer cyber and bio safeguards, only for vetted orgs (Cyber and Life Sciences Verification Programs).
- Fable 5.2 blocks some cyber and bio requests and answers them with an older model instead. Anthropic says it blocks about 55% fewer than Fable 5.1.
- Fable needs 30-day data retention for safety monitoring by default.
4. Headline results Anthropic reported
- 1,000+ subagents built a working prototype operating system from an empty folder in 8 days.
- 1,000 agents formalized Fermat's Last Theorem in Lean in 30 hours (the first formalization took 11 days).
- Solved 2 open problems on black holes; designed 3 working electronic devices.
- Testers: Jump Trading (quant research roles), IMC (new best scores on 3 of 4 quant research problems), Fundamental Research Labs (91% vs 86% on a 10-K forecasting eval).
5. Five prompts for long agent runs (my templates, not Anthropic's)
- Plan first: "Before you write code, write PLAN.md: the goal, the steps, how you'll test each step and what 'done' means. Then follow it and update it as you go."
- Verify yourself: "After each change, run the tests. If a result doesn't reproduce, say so and stop. Don't mark anything done that you didn't check."
- Use a team: "Split this into independent parts. Run one subagent per part, then review each result yourself before merging."
- Find the root cause: "Don't trust earlier conclusions in this repo. Find the root cause of this bug and show me the evidence."
- Measure and improve: "Measure the baseline, try one change at a time, keep only changes that improve the metric, and log every attempt in RESULTS.md."
New bot teams every dayFollow @mhtua on Instagram. Comment the keyword on any post and I'll send you the pack.
FOLLOW @MHTUA ← All free packs