Anthropic shipped Fable 5.1 and Mythos 5.1 on Tuesday, September 1, billing them as the world's most advanced models for coding and knowledge work. The Hacker News thread hit 1,403 points and over 1,300 comments. Almost none of the top ones were about SWE-Bench.
They were about how the thing writes.
The tone was set early, and by an Anthropic employee. Felix Rieseberg posted that Fable 5.1 is a big improvement in style, sounds less stereotypically like other Claude models, and follows style instructions more reliably. Then he added the part that reads, in hindsight, like a lit match: reading better prose makes him happier.
The replies were a mass exhale from people who have spent months reading Claude output professionally. One developer pasted a real handoff document from a long coding session and said he had no idea what it was telling him. The line everyone seized on described a bug as a "dead drag handle during a booked half-day you do not get back." Another commenter's verdict, and it's the most quoted assessment of Claude's house style I've seen: dense and vacuous at the same time. Full of jargon it invented, saying almost nothing.
A parade of comparisons followed. Corporatese. Purple prose. Dianetics. The Auditors of Reality from Terry Pratchett. Someone described their job as having shifted from writing code to solving the riddle of what Opus 5 is saying, and eventually swearing at it in all caps around the 300,000-token mark. Multiple developers said they'd moved to OpenAI's Codex and GPT-5.6 Sol specifically over readability, not capability. At least one canceled a subscription. Several described a workflow where they pipe Claude's output through a cheaper model just to get it back in English.
The theories about why are more interesting than the complaints. One camp argued the models are writing for themselves, packing signal for their own future context rather than for humans. A better-supported camp argued the opposite: this is what reinforcement learning selects for, a style that scores well with graders and reads like advertising copy. Somebody noticed the models constantly volunteer "honest caveats" because rubrics reward them. That one lands.
So: is 5.1 actually better at this? Cautiously, per the thread, yes, though the users least impressed were the ones on Opus 5, which is the model most people on cheaper plans are actually stuck with. That gap surfaced repeatedly, with users asking whether the style fixes are coming down to the flagship consumer model or staying in the expensive tier.
On the actual capability, the reviewers are polite and unexcited
Anthropic's own numbers show most improvements over Fable 5 landing in the 2 to 4 percent range. The one genuine leap is scientific agent work: 52.6 percent on Terminal-Bench-Science 0.1, up from 24.7 for Fable 5, with Opus 5 at 29.0 and GPT-5.6 Sol at 22.4.
Simon Willison flagged that the announcement spends conspicuous time on science, and that no other benchmark comes close to that jump. He also ran his pelican test, made an animated version for $1.37, and repeated that he's lost faith in the pelican benchmark as a proxy for anything except comparing models within a family at different effort levels. On ARC-AGI, at maximum effort, Fable 5.1 posts 97.5 percent on ARC-AGI-1 semi-private at $1.40 per task and 90.0 percent on ARC-AGI-2 at $4.49.
Frederic Lardinois at The New Stack summarized the competitive picture with a line that says more than any chart: nobody bothers comparing to Gemini 3.1 Pro anymore.
One HN commenter asked the obvious cynical question, which is how every lab manages to win every benchmark on every release day. The answer he got was that there are hundreds of benchmarks and you only need a favorable dozen.
The safeguards story is where the practitioners get sharp
Fable 5's habit of tripping its own guardrails and punting to Opus mid-task was a real workflow problem, and Anthropic has tuned it. Cyber safeguards now intervene about 60 percent less per session, biology safeguards 85 percent less on benign requests, and the model is permitted to find vulnerabilities in source code. Penetration testing, exploit generation, and binary scanning still get redirected.
Developers are split on whether this fixes anything. Enthusiasts of the change note that blocked requests now return an error rather than eating your money, and that you can opt into a fallback model. Detractors note what the fallback implies. One put it as being handed a Ferrari that refuses to drive on certain roads, forcing you back into the Mustang, and the choice isn't yours.
Web developers were the loudest here. The argument, which is hard to dismiss: securing a web app means finding and patching vulnerabilities, and the model cannot tell that apart from the other thing. Others said guardrails almost never fire on their work, which suggests the pain is highly domain-specific.
The enterprise complaints are quieter and probably matter more
Data retention kept Fable 5 out of regulated industries and out of a lot of research, since users had to accept a 30-day retention window. Anthropic's answer is Enterprise Frontier Safeguards, which keeps data in the customer's own cloud with their own keys, phasing in this fall, with zero data retention available to eligible customers in the meantime. Reception in the thread was tepid, mostly on the grounds that phased means not yet.
An academic said plainly that they won't put unpublished research questions into a system where prompts may be human-reviewed. That's a structural adoption problem no benchmark measures.
Then there's the anti-distillation change, which is the sleeper story. New API accounts can no longer edit Claude's prior context while preserving its thinking transcript. Anthropic says this closes a documented distillation route. It will also break agent harnesses that compact conversation history, and the restriction is coming for existing accounts.
Net read
Fable 5.1 is an incremental model with one large jump in scientific agentics, a meaningful price cut on cached tokens, and a materially less trigger-happy safety layer. That's a solid release. But the thing users spent two days actually talking about was whether the machine can write a sentence a human wants to read, which tells you something about where the frontier of frustration has moved. Capability is no longer the complaint. Legibility is.
