Claude Fable 5.2: A Developer's Field Guide to the New #1 Model

Shivam Kumar | VisionQuantech | November 10, 2026

Anthropic shipped Claude Fable 5.2 on November 9. It scores 64.6 on the Artificial Analysis Intelligence Index v5 — first place, ahead of Gemini 4 Carbon at 60.0 — and it costs 20% less than Fable 5.1: $8 input / $40 output per million tokens, cache reads at $0.20. The API model ID is claude-fable-5-2, available on the Claude Platform, AWS, Google Cloud and Azure, with the same 1M-token context and 128K max output as before.

Here is what to actually do with it, in the order that saves you money.

1. Set effort to "high", not "max"

This is the single most important decision. On AA's cost-per-task measure, Fable 5.2 at maximum effort costs $26.54 per task — the most expensive entry in the entire table. At "high" it costs $9.08. The score gap between the two settings is 64.6 vs 61.8: 2.8 points, 4.5%.

Do the division: you pay 2.9× for 4.5% more capability. Reserve max for one-shot, high-stakes reasoning where a few points decide correctness. Default every agent, every CI job and every coding assistant to high. If your harness does not expose an effort parameter, check the docs — Anthropic's model docs list the effort options, and "high" is where the value lives.

2. Price a real session before you migrate

A heavy coding session — 1M input tokens of context, 200K output tokens of code — costs $16.00 at list prices ($8 + $8). Twenty sessions a month per developer: $320.

Now add caching. If 80% of that input is cached project context and system prompts, input drops to $1.76 and the session costs $9.76 — 39% cheaper. Cache reads are $0.20 per million tokens, 40× cheaper than fresh input. The practical moves: put your system prompt, repo map and coding standards in a cached prefix; keep conversation history in cache blocks; never re-send a static context block as fresh tokens. This is not a micro-optimization — it is the difference between $320 and ~$195 per developer per month.

3. Know what the coding numbers actually say

On FrontierCode 2.0, Fable 5.2 scores 49.1% at xhigh, against 32.4% for Fable 5.1 and 45.5% for Opus 5.5. That is a genuine jump: +16.7 points over its predecessor, and ahead of Opus 5.5, which was the coding reference until yesterday. But two caveats from Anthropic's own disclosure: the numbers are self-reported, and the cyber and protein-design rows in its benchmark table use Mythos 5.2's scores, not Fable's.

Also note what Fable 5.2 does not lead: Gemini 4 Carbon wins the AutomationBench workflow test (53.9%), and Opus 5.5 leads AA-Briefcase Elo (1800). If your workload is long multi-step workflows rather than code generation, benchmark Carbon on your own tasks before switching. Re-run your own eval suite — the vendor's 49.1% is a starting hypothesis, not a guarantee.

4. Route the cheap work away from the flagship

Meta launched Muse Ember the same day: $0.10/$0.50 per million tokens, AA index 36.0, $0.14 per task. That is roughly 190× cheaper per task than Fable 5.2 at max. Ember is built for tool calls, routing and repeated agent tasks — exactly the work that makes up most of an agent's token burn.

The pattern: a small model classifies, routes and makes tool calls; Fable 5.2 handles the reasoning steps that actually need it. If your agent currently sends every step to a flagship model, you are paying flagship prices for classification work. Measure the split on your own traces — most teams find 60–80% of their agent calls are routable.

5. Mind the two dates in the fine print

Anthropic's new usage policy takes effect November 12. And Sonnet 4.5 retires from the API on November 30. If any service, script or CI job still calls Sonnet 4.5, migrate it this week — Fable 5.2 at high effort is both cheaper per task than the old Sonnet pricing in most workloads and substantially stronger. Check your model IDs with a grep across your repos before the 30th, not after.

6. Do not confuse the closed race with the open one

The top downloadable model is still Kimi K3.1 Pro (46.5 on AA). Tencent's Hy4 and Thinking Machines' Inkling 2 Small have only promised their weights. If your deployment is self-hosted or on-prem, Fable 5.2 changes nothing for you today — it is API-only. The open-weight race is still waiting on delivery; the closed race just got a new leader.

Checklist

• [ ] Switch coding agents to claude-fable-5-2, effort "high"
• [ ] Enable prompt caching on system prompts and repo context
• [ ] Re-run your own evals; do not trust the 49.1% blindly
• [ ] Route tool calls and classification to a small model (Ember-class)
• [ ] Grep for Sonnet 4.5 calls; migrate before Nov 30
• [ ] Read the new usage policy before Nov 12

The leaderboard will flip again. The economics — high over max, caching, routing — are the part that stays valuable after it does.

Sources: Anthropic's Fable 5.2 announcement and model docs; Artificial Analysis Intelligence Index v5 (Nov 9, 22:15 UTC); Meta's Muse Ember announcement. Cost arithmetic is the author's own, from published list prices.