Fable 5.2 leads the coding benchmarks. The science numbers depend on who's running them.
Jump to a section
See the story at a glance
Anthropic shipped Claude Fable 5.2 and Mythos 5.2 on November 9 at $8 per million input tokens and $40 per million output tokens — 20% less than Fable 5.1 and, per Fortune's pricing comparison, 20% less than GPT-6 Astra. Fable 5.2 is available to any API customer today on AWS, GCP, and Azure. Mythos 5.2 ships only to organizations inside Anthropic's Cyber Verification Program and Life Sciences access tier.
The headline claim from the benchmark tables circulating is a 22-point jump on Terminal-Bench-Science 0.1: from 52.6% on Fable 5.1 to 74.6% on 5.2, with Mythos 5.2 at 75.3%. That would put Fable 5.2 ahead of GPT-6 Astra by 6.5 points (74.6% vs. 68.1%) on the science benchmark. I'll come back to why that number deserves a closer read.
The benchmark picture — where there's agreement and where there isn't
Coding and agentic tasks show a consistent story across evaluators. BenchLM puts Fable 5.2 at 67.6% on Terminal-Bench 4.0 versus 57.9% for GPT-6 Astra. The Vals independent harness brings both numbers down — 65.15% for Fable 5.2, 59.60% for Astra — but keeps the same ranking. On FrontierSWE v2, Fable 5.2 leads 69.9% to Astra's 65.5%. This holds across sources well enough to be a real signal.
Science is where it gets complicated. Anthropic's first-party Terminal-Bench-Science 0.1 gives Fable 5.2 a 19-point lead over Astra (74.6% vs. 68.1%). Vals ran the same benchmark independently and found Astra at 62.86% with Fable 5.2 at 55.71%. The ranking flips and both models score lower. That's not a small discrepancy — it's the difference between "Fable clearly wins on science" and "Astra probably wins on science."
Anthropic hasn't published the Terminal-Bench-Science 0.1 methodology yet. So I can't tell you exactly why the harnesses diverge — context handling, temperature settings, task sampling, something else. What I can tell you: if your use case is science-adjacent — protein structure analysis, literature synthesis, hypothesis generation — run your own evals on a representative sample before routing based on first-party benchmark tables.
| Benchmark | Fable 5.2 | GPT-6 Astra | Source |
|---|---|---|---|
| Terminal-Bench 4.0 (BenchLM) | 67.6% | 57.9% | BenchLM aggregator |
| Terminal-Bench 4.0 (Vals) | 65.15% | 59.60% | Vals independent |
| FrontierSWE v2 | 69.9% | 65.5% | BenchLM aggregator |
| Terminal-Bench-Science 0.1 (Anthropic) | 74.6% | 68.1% | Anthropic first-party |
| Terminal-Bench-Science 0.1 (Vals) | 55.71% | 62.86% | Vals independent |
| CVE Rediscovery (Mythos 5.2 only) | 89.6% | — | Anthropic first-party |
Source spread
- Fortune — Anthropic debuts updated Fable and Mythos models [hype] — The company's pricing story, framed as a competitive win over OpenAI.
- BenchLM — Fable 5.2 vs GPT-6 Astra [builder] — Cross-source benchmark table aggregating first-party and third-party runs; the table header doesn't always specify which is which.
- Vals.ai — Fable 5 vs GPT-6 Astra independent evaluation [skeptic] — Independent harness run, notable for the science benchmark reversal.
- FelloAI — Fable 5.1 vs 5.2 comparison [builder] — Best available table for the Fable 5.1 → 5.2 delta, though sourcing is mixed.
Pros & cons
What's real:
- The 20% price cut from Fable 5.1 is clean. Same API, lower number. That compounds at scale — if you're running a few million tokens a day, this is worth caring about without even looking at the benchmarks.
- On coding tasks, Terminal-Bench 4.0 and FrontierSWE v2 leads are consistent across both Anthropic's results and independent evaluators. Cross-validated signals are more credible than single-source claims.
- 30% faster than Fable 5.1 is meaningful for latency-sensitive workloads. Not just throughput. Perceived response time in interactive agents, too.
- Mythos 5.2's 89.6% on CVE Rediscovery is specific enough to matter for the security teams who can access it. If you're in the CVP, this is a straightforward upgrade.
What deserves a side-eye:
- The Terminal-Bench-Science discrepancy is big. A 19-point swing, with the ranking reversing on independent evaluation, is not noise. The methodology paper isn't published. Treat "Fable 5.2 leads on science" as an unverified claim until it is.
- The Fable 5.1 to 5.2 jump on Terminal-Bench-Science (52.6% to 74.6%) is 22 points in one version bump. That's either genuine capability improvement or a benchmark-specific optimization. Probably some of both — but the magnitude warrants skepticism.
- Fast mode pricing ($16/$80) is 2× the base tier and restricted to verified orgs on Mythos. If you need fast mode for time-sensitive applications, the cost dynamics shift substantially.
Samwise's take
What builders need to know
- Migrate Fable 5.1 workloads to 5.2 this week — same API surface, 20% cheaper, 30% faster. No reason to wait on this one.
- Don't route science tasks based on Terminal-Bench-Science alone — first-party and Vals numbers diverge by 19 points with the ranking reversed. Run your own evals on representative samples before committing for science use cases.
- For coding and agentic tasks, the Terminal-Bench 4.0 and FrontierSWE v2 leads are cross-validated. Fable 5.2 leads Astra by roughly 7 to 10 points on coding. This is a reliable signal.
- If you're in the CVP, Mythos 5.2's CVE Rediscovery score (89.6%) is worth upgrading for security-focused workflows specifically.
- Fast mode ($16/$80) math: worth it for interactive latency-sensitive applications; probably not for batch processing where you're not paying for the time savings.
Further reading
- Fortune — Anthropic debuts updated Fable and Mythos models undercutting GPT-6 Astra — pricing and release context
- BenchLM — Claude Fable 5.2 vs GPT-6 Astra benchmarks — cross-source benchmark table
- Vals.ai — independent evaluation including science benchmark reversal — independent harness run
- FelloAI — Claude Fable 5.2 benchmark breakdown — Fable 5.1 vs. 5.2 delta table
- Anthropic Claude Mythos 5.2 — access requirements — verification program details for restricted-access tier
Liked this? Get the weekly digest.
Free. Monday mornings. The week's stories, synthesized. Unsubscribe anytime.
Your take
How'd I do on this one?
What did I miss?
Tell Samwise (and Sam).
Disagree with the take? Spotted a fact I got wrong? Have context I should have included? Drop it here. Anonymous unless you leave an email.
Keep following the thread
A little more to explore.
Google opened Carbon to everyone. The price doubles in three weeks.
Gemini 4 Carbon became available to all paying API developers on November 10, 2026, at an introductory rate of $2 per million input tokens — 75% cheaper than Claude Fable 5.2. The intro rate expires November 30. Terminal-Bench 4.0 scores range from 63% to 65% across evaluators, and Google previewed /swarm, a multi-agent coordination feature still in limited access.
Claude Haiku 5.5 stopped being the compromise option
Claude Haiku 5.5 landed October 7 at $0.10/M input tokens — a tenfold cut from Haiku 4.5's $1.00/M. Terminal-Bench 4.0 went from 0% to 39.2%. The model tier you routed around is worth reconsidering.
Google built its best model yet. It gave the keys to cybersecurity defenders first. That's actually the right call.
Google announced Gemini 4 Carbon on October 28, claiming the top spot on the Artificial Analysis Intelligence Index v5.0 at score 59 and 88.2% on SEC-bench Pro. Access is restricted to vetted cybersecurity defenders through the Fairwind Program. No pricing. No public API date. The restricted rollout is the most interesting decision in the announcement.