Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1 on September 1st, and it's the top story on Hacker News right now with over 1,200 points and 1,100+ comments. Not because of the benchmark numbers — though those are real — but because of what the release quietly admits about how frontier labs are handling safety, and because half the thread is arguing about whether the model can still write a sentence without sounding like a compliance memo.
Let's get into it.
The numbers, unvarnished
Fable 5.1 doesn't just nudge the needle. It moves it:
| Benchmark | Fable 5 | Fable 5.1 | Delta |
|---|---|---|---|
| Terminal-Bench-Science 0.1 | 24.7% | 52.6% | +28pp |
| Terminal-Bench 4.0 (agentic coding) | 42.0% | 55.8% | +13.8pp |
| AutomationBench | 17.1% | 31.4% | +14.3pp |
| GDPval-AA v2 (knowledge work) | 1,723 | 1,853 | +130 |
| OSWorld 2.0 (strict) | 36.1% | 41.7% | +5.6pp |
That Terminal-Bench-Science jump is the one to stare at. Doubling a benchmark in a point release is not a "vibes improved" update — that's a real capability shift in a model that's still priced the same on input tokens. For context, Opus 5 scores 29.0% on the same eval. Fable 5.1 isn't just beating its predecessor, it's beating the more expensive model in its own lineup on a task built to simulate messy, real scientific workflows in a terminal.
Browserbase ran their own hardest-tier browser-agent tasks and got 82% completion with Fable 5.1, versus 74% for Opus 5 and 57% for Fable 5. If you're building browser agents and still defaulting to Opus for "quality," that assumption just expired.
The pricing move nobody's competitor wanted
Here's the part that should actually change your architecture decisions: cache reads dropped from $1.00 to $0.25 per million tokens — a 75% cut.
Base rates didn't move ($10 input / $50 output), but if you're running any kind of agentic loop — the kind where the model re-reads a growing context window on every tool call — cache reads are where your bill actually lives, not the input tokens. Anthropic is claiming ~25% savings on typical workloads and up to 45% on highly agentic work. For anyone running Claude Code or a custom agent loop against a large repo context, that's not a rounding error, that's a different cost curve entirely.
Stack it against the competition:
- Fable 5.1: $10 in / $0.25 cache / $50 out
- Opus 5: $5 in / $0.50 cache / $25 out
- Sonnet 5: $2 in / $0.20 cache / $10 out
- GPT-5.6 Sol (promo pricing): $4 in / $0.40 cache / $20 out
- Gemini 3.7 Flash: $0.75 in / $3.75 out (no cache tier disclosed)
Fable 5.1's cache reads now sit at 2.5% of its base input price, versus 10% for every other model in Anthropic's own lineup. That's a deliberate structural choice, not a discount — it's Anthropic betting hard that the future is long-running agents hammering a cached context, not one-shot completions. If your product looks like that, this release was aimed directly at you.
Fable and Mythos are the same model. That's the actual story.
This is the part everyone's dancing around: Fable 5.1 and Mythos 5.1 are not two different models. They're the same weights with different safeguard intensity dialed in. Fable is what you and I get. Mythos is the same brain with the safety rails loosened, gated behind Anthropic's Cyber Verification Program and Life Sciences Verification Program, for vetted cyberdefenders and bio researchers.
The honest part — and credit where due — is that Anthropic published the delta. On Terminal-Bench 4.0, Fable 5.1 hits 55.8%. Mythos 5.1, same model, fewer restrictions, hits 60.9%. That five-point gap is the literal, quantified cost of the safety theater layered on top for general release. Most labs would never publish that number. Anthropic did, probably because they know security researchers would reverse-engineer it anyway.
Practically: cybersecurity safeguards now throw 60% fewer false-positive interventions per Claude Code session, and biology safeguards fire 85% less often on benign queries. If you've been fighting Claude refusing to look at a buffer overflow in your own codebase because it pattern-matched "exploit," that friction just got meaningfully smaller — but it's still there by design, and it's staying there unless you're in the vetted program.
The writing style thing is actually a bigger deal than it sounds
Buried in the HN thread is a fight that matters more than it looks. An Anthropic engineer (Felix Rieseberg) posted that Fable 5.1's writing "sounds a lot less stereotypically like other Claude models." The replies did not agree.
Multiple engineers pulled real examples of Claude output that reads like corporate word salad — dense, hedge-everything, "worth noting that" filler stacked four sentences deep before saying anything. The community consensus forming in that thread: this isn't a prompting problem, it's baked in by RLHF, and no system prompt reliably beats it back. People are now maintaining Claude "skill" files just to force plain language, or routing Opus output through Haiku as a cleanup pass before it reaches a human.
If you're building anything that surfaces Claude's raw text to end users — commit messages, PR descriptions, customer-facing summaries — this is not a cosmetic complaint. It's a production problem people are actively engineering around, and the fix ships as a workaround, not a setting.
What I'd actually do with this
If you're running agentic workflows with large, reused contexts — Claude Code, custom coding agents, anything hitting the same repo or doc set repeatedly — swap to Fable 5.1 today and re-run your cost model before you re-run your benchmarks. The cache pricing alone probably pays for the migration.
If you're doing security research or bio work and think you qualify for Mythos, apply. A five-to-eight point benchmark gap on your actual task is not nothing, and the verification programs exist specifically so people doing legitimate defensive work aren't stuck fighting a model that thinks every SQL query is an attack.
If you're shipping Claude's text straight to users: stop. Put a cleanup pass in front of it, or accept that "Claudish" is now a term of art on Hacker News for a reason.
The model got smarter and cheaper in the same release. That doesn't happen often. Use it.
Sources: Anthropic's announcement, VentureBeat, and the Hacker News discussion.
Top comments (0)