Anthropic released Claude Sonnet 5.5 on September 28, 2026. It is the second model in the Claude 5.5 family, and the news worth reading past the headline is that the price did not move. Input is still $2 per million tokens, output is still $10 per million, exactly as with Sonnet 5.

So what did change? Anthropic says the model generates output more than 30% faster than Sonnet 5, and that it can cut the total cost of a task by as much as 30%. Both of those are company claims. The second one is also the more interesting one, because the saving is not coming from a cheaper token. It is coming from using fewer tokens and fewer tool calls to reach the same finished result.

For the record, the model id is claude-sonnet-5-5, and it is available on AWS, Google Cloud and Microsoft Azure with a zero-data-retention option. Anthropic also says Claude Haiku 5.5, aimed at high-volume, cost-sensitive work, will follow in the coming weeks.

What actually changed

Two numbers, both from Anthropic: more than 30% faster output generation than Sonnet 5, and up to 30% lower total cost per task.

The speed number is the easy one to feel. If you are streaming tokens into a UI, Anthropic’s more than 30% faster output generation is the difference between an assistant that feels attached to your request and one that feels like it is buffering. If that holds on your workload it is a real quality-of-life difference, though a throughput figure from a vendor test bed is not a promise about every prompt.

The cost number is messier, and worth reading carefully. “Up to 30%” is a ceiling, not an average. A model that solves a task in three tool calls instead of seven only saves money if the three-call version produces the same result. On a task that was already solved in one call, there is nothing to save. On a task that needs more calls than the old model used, you pay more. The saving is conditional on your workload, which is exactly why the number cannot be transferred from Anthropic’s test set to your stack.

The second thing that changed is the safety layer, which is genuinely new for a Sonnet. More on that below, because it has a practical consequence for anyone building agentic coding tools.

Where it fits next to Opus 5.5

Anthropic positions Sonnet 5.5 as a faster, lower-cost complement to Claude Opus 5.5, not a replacement for it. The division is roughly: Opus 5.5 is built for complex work that needs careful judgment, and Sonnet 5.5 is strongest at well-scoped everyday tasks — bug fixing, polished documents, slides and spreadsheets. Anthropic also says it has a good eye for design.

Tom’s Guide made the sharper version of the argument: the smaller model may turn out to be the more important release for most people, because Sonnet is the model for everyday work like rewriting an email, summarizing a long document, or building a slide deck. Opus is the one you reach for when the task is genuinely hard.

Read that as a routing decision rather than a ranking. If you have a single model configured for everything, you are probably using Opus for tasks where Opus is not earning its price, and Sonnet for tasks where it is. Sonnet 5.5 does not make Opus worse. It makes the default worth questioning.

The benchmarks, read carefully

The headline benchmark is Terminal-Bench 4.0, where one report puts Sonnet 5.5 at 70.6% against 66.4% for Opus 5.5. The setup matters: Opus is being run at that model’s highest effort setting, and it is being compared at half the per-token price.

That is a real and somewhat surprising result, and it deserves a fair reading rather than either celebration or dismissal. A smaller, cheaper model overtaking a larger one on a coding benchmark is a meaningful signal, and the effort-setting detail in the published comparison works against the vendor’s own case, which counts for something.

It is still a single benchmark, reported in a launch context, with no independent reproduction attached. The Decoder’s coverage of the launch framed it the same way — nearly matching Opus 5.5 on benchmarks — and “nearly matching” is the honest description. Nearly matching, at half the price, at 30% less cost per task, would be a very good outcome. It is not the same as being better at everything, and the two models still sit in different places in the lineup.

Two other claims, both explicitly Anthropic’s:

  • Long-horizon vision. Anthropic says Sonnet 5.5 is the first Sonnet to beat Pokémon Red working only from screenshots. It frames this as an indicator of long-horizon work and image understanding, which is a reasonable reading of why anyone would pick that benchmark.
  • Writing. Anthropic says it writes more clearly than previous-generation models, which is the kind of claim that is very hard to evaluate from the outside because “clearly” is doing a lot of work in the sentence.

None of these are disqualifying. All of them are vendor numbers from launch week, and every one of them will be worth re-reading once someone who is not selling the model has run them.

What early customers reported

Four companies, all self-reported, none independently verified:

  • Base44 reports an average of 3.6 iterations per build across 118 real app builds, against 7.7 for Opus 5, with the fewest failed tool calls of any model it compared.
  • Zendesk says tickets are processed 20% faster than with the Claude models it runs in production today.
  • Slack reports roughly 14% fewer output tokens on offline Slackbot evaluations, with no prompt changes.
  • CodeRabbit plans to move simple and moderate code reviews to the model.

Base44’s number is the most interesting of the four, because fewer iterations is the metric that actually matters to someone shipping software. It is also the number most likely to be workload-specific: an app-building loop that fails a few times less is a very different result from a summarisation task that never loops at all.

Slack’s figure is the cleanest, because it was measured on offline evaluations with no prompt changes. That is closer to an apples-to-apples comparison than most vendor case studies ever get to be. It is also the smallest effect, at roughly 14%.

The pattern across all four is the same, and it is consistent with the pricing story: the model is not cheaper per token. It is more efficient per task. Whether that shows up as dollars depends entirely on how many retries your workload was doing before.

The safety layer is new for a Sonnet

Sonnet 5.5 is the first Sonnet to ship with frontier-style protections, and this is the part of the release that could change behavior in your pipeline.

Anthropic says Sonnet 5.5’s cybersecurity capabilities are comparable to Opus 5, so higher-risk cyber requests visibly fall back to Sonnet 5. Routine software development is unaffected.

Read that carefully. A visible fallback means that for a subset of requests you are not getting 5.5 at all — you are silently getting an older model, and the only signal is the downgrade itself. If you are building anything that leans on frontier-adjacent security reasoning, that is a behavior change worth knowing about before you find it in a log. For ordinary development work, it should not register at all.

It is also the first Sonnet with classifiers that block reasoning extraction. Reasoning extraction is the practice of pulling a model’s chain of thought out of it to train or to distil a competitor, and the classifiers exist to stop that. The related mechanism is preserved thinking, which ties a model’s reasoning to the account that produced it.

This matters more than it first appears. If your workflow treats extended thinking as the product — long autonomous runs where the reasoning trace is what you are inspecting — then the reasoning is now bound to the account that generated it, and extraction is blocked at the classifier layer. That is a deliberate change in the economics of long-running agent work, and it is worth a look at your own assumptions about what you can do with a long trace.

Biology safeguards are unchanged.

What this means for your bill

The practical advice is short: run your own evaluation before migrating anything.

The evaluation should measure cost per finished task, not cost per token. Those are different numbers and only one of them appears on the invoice. Take a fixed set of real tasks from your own work, run them on Sonnet 5 and on Sonnet 5.5 with identical prompts, and record two things: total tokens consumed, and total tool calls made. Latency is worth logging too, since the speed claim is the one most likely to reproduce on your own stack.

Then compare against your current default. If your current default is Opus, the interesting column is not “which score is higher” — it is how many tasks fall inside the gap between them, and what those tasks cost you when routed to the smaller model. If your current default is already Sonnet 5, the calculation is simpler: same token price, fewer tokens per task, and you keep everything else about your integration.

On when to stay on Opus: keep it for the work that is genuinely open-ended, where the scope has to be defined as the model works, and where a misstep costs more to reverse than to prevent. The cost argument does not apply cleanly there, because a task that needs careful judgment is often a task where the expensive part is the thinking, and fewer tool calls will not compress the thinking. The token savings are worth having on well-scoped work. They are not a reason to reroute the work you built Opus for in the first place.

On the benchmarks, the same logic applies. Treat the model as promising and unproven in roughly equal measure, and let your own traces over the next few weeks be the thing you trust.

Sources

  • Unite.AI, “Anthropic Releases Claude Sonnet 5.5 at Unchanged Sonnet 5 Pricing”
  • Thurrott.com, “Anthropic Releases Claude Sonnet 5.5”
  • VentureBeat, “Anthropic launches Claude Sonnet 5.5 with 30% cost reduction per-task…”
  • The Next Web, “Anthropic releases Claude Sonnet 5.5 with the cyber limits it reserved for its best models”
  • Tom’s Guide, “Claude Sonnet 5.5 just launched…”
  • The Decoder, “Anthropic’s Claude Sonnet 5.5 nearly matches Opus 5.5 on benchmarks…”