Anthropic shipped Claude Sonnet 5.5 on September 28, six days after Opus 5.5. It's the top story on Hacker News right now — 775 points, 500+ comments — and for once the hype is mostly earned. But buried under the benchmark charts is a detail that should bother you more than the marks do: Anthropic reserves the right to silently swap your request to the older Sonnet 5 if it decides your prompt looks "higher-risk." Let's get into both halves of that story.
The numbers, unvarnished
Same price as Sonnet 5. That's the headline nobody's arguing with: $2/M input, $10/M output, unchanged. Everything else got better:
- Terminal-Bench 4.0 (agentic coding): 70.6%, up from Sonnet 5's 10.3%. That's not an iteration, that's a different model class.
- Beats Opus 5.5's 66.4% on the same benchmark — the "cheap" model outscoring the flagship on agentic coding.
- GDPval-AA v2.1 (general knowledge work): 1844, two points off Opus 5.5's 1846, ~400 points above Sonnet 5's 1449.
- Chartography (visual chart recognition): 61.6%, behind Opus 5.5's 64.4% but a massive jump from Sonnet 5's 15.6%.
- Anthropic claims 30%+ faster output and up to 30% lower cost per task, driven by fewer tokens and fewer tool calls to get the same job done.
If you're doing agentic coding — the stuff where a model plans, edits files, runs commands, iterates — Sonnet 5.5 is now the rational default over Opus 5.5 for most workloads. You're paying a fraction of the price for a model that benchmarks ahead of the expensive one on the exact task you're using it for. That's the actual story, and it's a big deal for anyone running agents at volume.
Devs on HN are backing this up with usage data, not vibes. One thread regular said they've "barely used Sonnet" since Opus 5.5 landed but came back because the new one "doesn't consume that many tokens." Another reported their Pro-plan quota lasting all day for the first time in months — no more mid-session rate-limit walls. Someone else burned through a full week's usage allotment and the reset within hours, which tells you people are throwing real workloads at it, not just kicking the tires.
The part that should make you pause
Sonnet 5.5 ships with "frontier-style cyber safeguards" — classifiers that can detect and block reasoning extraction and other misuse patterns, the same category of guardrail previously reserved for Anthropic's top-tier models. On paper that's reasonable: a model this good at agentic tasks is also good at the tasks you don't want it good at.
Here's the catch reported by The New Stack: in what Anthropic calls "higher-risk" situations, your request to Sonnet 5.5 can get silently rerouted to the older Sonnet 5 instead. Not refused. Not flagged. Rerouted. You picked one model, your bill and your prompt say one model, and the response might be coming from a different one with different capabilities.
If you're building anything where consistent model behavior matters — evals, benchmarking your own pipeline, debugging a regression, reproducing a bug report — an invisible model swap is a landmine. "It worked yesterday" stops being a debugging lead and starts being a coin flip. Anthropic hasn't published the trigger criteria, which means you can't design around it, only get surprised by it.
Where this leaves you
If your workload is straightforward coding, docs, or general knowledge work, switch to Sonnet 5.5 today — same price, meaningfully better output, and it now beats Opus on the benchmark that matters most for agent work. Don't reflexively reach for Opus 5.5 anymore; it's the wrong default for cost-sensitive agentic pipelines as of this week.
If you're running anything that depends on deterministic model identity — regression suites, reproducibility, compliance logging — go find out what "higher-risk" actually triggers before you ship on top of it. Log the model field in every response, not just the one you requested. Right now that's the only way you'll catch the swap when it happens to you.
The benchmarks are real. So is the asterisk.
Sources: TechCrunch, VentureBeat, The Decoder, The New Stack, Hacker News discussion
Top comments (0)