Anthropic shipped Claude Fable 5.2 on November 9. It scores 49.1% on FrontierCode 2.0 — the best agentic coding result any lab has published — and costs 20% less than Fable 5.1. If you’re still running claude-fable-5-1, swap the model string today. There’s nothing to wait for.
The Benchmark That Actually Matters
FrontierCode 2.0 doesn’t measure trivia or toy problems. It measures whether open-source maintainers would merge a model’s code changes into a real repository. That’s a harder bar than “does it compile” — it asks whether the code is correct, idiomatic, and mergeable. Fable 5.2 scores 49.1% at extra-high reasoning effort, clearing Opus 5.5’s 45.5% and leaving GPT-6 Astra (42.8%) and Gemini 4 Carbon (41.1%) behind.
The more interesting number is 45.3%. That’s Fable 5.2 at high reasoning effort — essentially tied with Opus 5.5 at max effort. Opus 5.5 runs at roughly $40/$200 per million tokens. Fable 5.2 runs at $8/$40. The same coding performance for 80% less money, with headroom left before you even hit Fable 5.2’s ceiling.
| Model | FrontierCode 2.0 | Input $/M | Output $/M |
|---|---|---|---|
| Claude Fable 5.2 | 49.1% | $8 | $40 |
| Claude Opus 5.5 | 45.5% | ~$40 | ~$200 |
| GPT-6 Astra | 42.8% | — | — |
| Gemini 4 Carbon | 41.1% | — | — |
| Claude Fable 5.1 | 32.4% | $10 | $50 |
The Price Cut Goes Deeper Than the Headline
The 20% cut isn’t just on base tokens. Cache reads dropped from $0.25 to $0.20 per million tokens. Cache writes at both tiers dropped 20%. Batch API dropped 20%. If you’re running prompt caching on complex workflows — and you should be — the savings compound across every layer of your stack. A team spending $10,000/month on Fable 5.1 drops to roughly $8,000 with a one-line model string change. No refactoring. No testing new prompts. Just claude-fable-5-2.
Adaptive Reasoning That Doesn’t Stall
The headline feature isn’t a toggle — it’s a fix. Fable 5.1’s reasoning could lose the thread on long multi-tool sessions, forcing you to restart or summarize context manually. Fable 5.2 keeps reasoning state across the entire conversation. The model holds a mental model of a large codebase through dozens of tool calls without degrading.
What that enables in practice: Anthropic coordinated more than 1,000 subagents for eight days to build an operating system that boots on a laptop. A Python-to-Rust production migration produced a service that starts 26 times faster. An Uber engineer ran a 10-hour security audit with 170 reviewer agents that returned PRs and a policy proposal. These aren’t demos — they’re the shape of what long-horizon agentic work looks like when reasoning doesn’t fall apart at hour three.
Fast Mode for Interactive Work
Standard Fable 5.2 is already 30% faster than 5.1. Fast mode adds another 2.5x on top of that, at $16/$80 per million tokens. The doubled cost is real, but so is the use case: if you’re building an interactive coding assistant where the user is waiting on the model, every latency cut matters more than marginal token savings. Fast mode is available on the Claude API (Messages), Google Cloud, and Microsoft Foundry.
Migration: One Line, One Gotcha
Upgrading from Fable 5.1 to 5.2 is a model string swap:
# Before
model="claude-fable-5-1"
# After
model="claude-fable-5-2"No prompt changes required. Effort level parameters carry over unchanged.
The one thing to check: thinking block compatibility. Fable 5.2 can read Fable 5.1 thinking blocks. Fable 5.1 cannot read Fable 5.2 thinking blocks. If you’re running a mixed pipeline where a 5.1 instance consumes 5.2 output — that breaks. Pure 5.2 pipelines have no issues. If you’re running mixed models for cost or latency reasons, audit your thinking block handoffs before deploying. The full migration guide covers edge cases worth checking before a production cut.
Upgrade Now
Fable 5.2 is better on the benchmark that matters, 20% cheaper across the board, and faster out of the box. The migration is a one-line change with one known gotcha. There’s no case for staying on 5.1. The competitive benchmarks vs GPT-6 Astra and Gemini 4 Carbon make the positioning clearer: this is currently the best coding model available, at a price that undercuts its nearest competitor by a wide margin.
