Anthropic ships Claude Opus 5.5: Fable 5.1-level performance at 40% lower serving cost
- Anthropic released Claude Opus 5.5, the first model of the Claude 5.5 family, which matches Claude Fable 5.1 on most work and costs 40% less to run than Opus 5; one early tester finished a 680,000-line code migration in under a day.
- On Terminal-Bench 4.0 it scores 66.4% against GPT-6 Astra's 57.9% and Opus 5's 52.3%, but GPT-6 Astra still leads on AutomationBench (41.4% vs 40.0%) and Terminal-Bench-Science 0.1 (64.6% vs 58.7%).
- Token pricing is $4 input and $20 output per million (20% under Opus 5), cache reads are $0.20 per million (60% under Opus 5), and output generation is more than 30% faster than Opus 5.
- Opus 5.5 posts the best results to date on Anthropic's automated behavioral audit, was checked pre-release by Frontier Design and METR, and ships with Fable 5.1-style safeguards because its biology and cybersecurity capability matches Claude Mythos 5.1.
- Anthropic raised five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans and gave subscribers a storable rate limit reset, with Sonnet 5.5 and Haiku 5.5 due in the coming weeks.
Hacker News opinions
After yesterday's outage is the new Opus 5.5 load-bearing? I'm genuinely asking.
Worth stating plainly: requests to Claude Mythos 5.1, Claude Fable 5.1 and Claude Opus 5 all hit elevated error rates. Those models are the rungs my whole dev tooling stack sits on, so the outage really bit.
It was difficult to not notice the complaints. Opus 5 was unusable, most of my team went back to Opus 4.6 for most of their work.
I'm confused about the link between model improvements and incoherence. The vocabulary shifts on the same task and the same codebase, and it's not just the verbosity. Are there studies on this?
Opus 5 was just incoherent. I'd love a postmortem on how writing styles change from model to model; I wouldn't be surprised if Opus 5 was trained on content written by other LLMs.
I don't think it's substantially different. I pasted a random chunk of code at 5.5 and got the same annoying cadence and writing style, just with slightly less prominent claudisms.
Just post the bloody content. This UI and scrolling thing is horrific, and it hijacks back and forward.
Turn on "reduce motion" in your accessibility settings and you get served a sane version. It only looked broken because I didn't have prefers-reduced-motion enabled.
The performance in Firefox is terrible too. I couldn't make it past the hero.
This makes Fable not really make any sense?
Didn't they say Opus 5 was Fable-level too? Benchmarks don't tell us much anymore at this point. I'd love it to be as strong as Fable, but I'm skeptical about how that looks in practice.
What happened to "slowing down"?
Slowing down only makes sense if you can coordinate a slow-down for everyone.
The test they ran porting HAProxy from C to Rust is crazy. That's not a toy workload.
At this point I'm convinced they skip numbers so fast they'll end up at or ahead of OpenAI's numbering scheme. It's the Xbox 360 versus PS3 thing all over again.
Wow, though I need to see its token efficiency to better assess. I keep hearing rumors it generates much more output tokens per task.
My projection is they're still pretty far behind on token efficiency, but they'll sew it up in the next few releases. OpenAI has put real work there, though there's no magical secret sauce Anthropic can't catch up with.
They changed their approach to usage resets for this release. Instead of an instant reset I now have a 'Reset for free' button expiring Oct 22, so I can trigger a whole new usage window whenever's convenient.
That's copied from Codex, since that's exactly how Codex works.