Last updated: October 07, 2026, 3:20 PM IST
Anthropic put out Claude Sonnet 5.5 on September 28, and the pitch this time is not a bigger brain. It is a faster one that costs less to run. Same token prices as Sonnet 5, answers arriving more than 30% sooner, and a coding score that jumped so sharply that it needs a second look before you believe it.
Sonnet 5.5 is the second model in the Claude 5.5 family, after Opus 5.5. Anthropic splits the two cleanly. Opus is for the hard, open-ended work where judgment matters. Sonnet is for the everyday stuff most people actually do: fixing bugs, writing code, polishing documents, building slides. If you have been defaulting to Sonnet 5 for that kind of work, here is what changed and whether it is worth moving.
What Is Claude Sonnet 5.5?
Sonnet 5.5 is Anthropic's new mid-tier model, called claude-sonnet-5-5 in the API. The context window stays at 1 million tokens, output goes up to 128K tokens, and it is live pretty much everywhere on day one: the Claude apps, Claude Code, the API, AWS, Google Cloud and Microsoft Azure, with GitHub Copilot rolling it out too.
The engine work went into efficiency rather than raw size. The model takes fewer steps to finish the same job, and it batches tool calls better than Sonnet 5 did, which means less back-and-forth and less wasted time waiting on it. Anthropic also tidied up the effort settings. There are five of them now, from low to max. The apps and Claude Code start you at Medium, the API platform starts at High, and that choice matters more than you might expect, as we will get to in the cost section.
How Fast Is Claude Sonnet 5.5?
Anthropic says Sonnet 5.5 generates output more than 30% faster than Sonnet 5, and calls it the fastest Sonnet it has shipped. In chat that is the kind of gain you feel straight away: long answers stream in quicker, and in an agent loop where the model is thinking, calling a tool, thinking again, the whole run finishes sooner.
Speed on its own would be a nice-to-have. What makes this release interesting is that the faster model also scored better, not worse, on the tests that involve actually doing work in a terminal.
The Benchmark Jump: Terminal-Bench 4.0 Tells the Story
The headline number is Terminal-Bench 4.0, a test where models have to complete real multi-step tasks from the command line. Sonnet 5 managed 10.3% on it. Sonnet 5.5 scored 70.6%. That is not a small bump. It also puts the mid-tier model ahead of Opus 5.5's 66.4% on this one test, which is the kind of result that makes people look twice at which model they are paying for.
The rest of the picture is more even. On GDPval-AA, which measures practical work across 44 occupations, Sonnet 5.5 sits about two points behind Opus 5.5 (1844 against 1846) and roughly 400 points ahead of Sonnet 5. On CursorBench it trails Opus slightly, 55.5% to 57.8%. Anthropic is open about the split: Opus stays clearly stronger on complex, open-ended work that needs sustained judgment.
| Measure | Sonnet 5 | Sonnet 5.5 | Opus 5.5 |
|---|---|---|---|
| Terminal-Bench 4.0 | 10.3% | 70.6% | 66.4% |
| CursorBench 4.0 | 34.1% | 55.5% | 57.8% |
| GDPval-AA (work tasks) | 1449 | 1844 | 1846 |
| Price per 1M tokens (in/out) | $2 / $10 | $2 / $10 | $4 / $20 |
| Context window | 1M tokens | 1M tokens | 1M tokens |
If you spend your days inside a terminal agent, our hands-on look at Claude Code mods and the /plugin system shows the setup these models are now competing in.
What Does Claude Sonnet 5.5 Cost?
The rate card did not move. Sonnet 5.5 costs $2 per million input tokens and $10 per million output, exactly like Sonnet 5, with cache reads at $0.20. Opus 5.5 is double that on both input and output.
So where does the "cheaper" claim come from? Per task, not per token. Because Sonnet 5.5 needs fewer tokens and fewer tool calls to reach the same result, Anthropic says the same work costs up to 30% less. Early customer numbers point the same way. Resultsense's launch coverage sums up the shape of it: unchanged list prices, fewer tokens burned per job, and scores that land close to a model costing twice as much per token.
One honest caveat before you rewrite your budget. Independent testers running the model at its highest effort setting found it thinking for a long time and using a lot of output tokens, which can push the per-task cost back up. Both things can be true at once: cheaper at the settings most people use, expensive if you leave it on max and walk away. Pick your effort level on purpose. Medium is plenty for well-scoped coding work.
Sonnet 5.5 or Opus 5.5: Who Should Switch?
Here is the simple way to think about it.
- Everyday coding, bug fixes, docs, slides, spreadsheets: Sonnet 5.5. It is faster, half the token price of Opus, and on terminal-heavy work it is level with or ahead of its bigger sibling.
- Hard, open-ended projects where the plan itself is unclear: Opus 5.5. Anthropic itself says sustained judgment is still Opus territory, and the benchmark gaps on that kind of work back it up.
- Still on Sonnet 5: switch. Same price, faster answers, better scores. Developers should check the migration notes first, though. Unite.AI's pricing and migration rundown flags the practical bits: adaptive thinking is now on by default, and the old thinking-off setting needs the new between_tools option instead.
A workflow plenty of developers are landing on: plan the big pieces with Opus, then hand the build to Sonnet. You get the judgment where it counts and the speed everywhere else. And if you want a reminder of why unchecked agent speed needs a human watching, our report on GPT-6 Astra getting caught cheating at StarCraft is the cautionary tale of the month.
What About Claude Haiku 5.5?
The family is not finished. Anthropic says Claude Haiku 5.5, the small and cheapest tier, follows in the coming weeks. No date and no prices yet. If Sonnet 5.5's job is to make the mid-tier cheaper per task, Haiku's job will be the high-volume stuff: classification, quick summaries, the calls you make thousands of times a day. We will cover it when it lands.
Frequently Asked Questions
When was Claude Sonnet 5.5 released?
Anthropic released it on September 28, 2026. It is the second model in the Claude 5.5 family, after Opus 5.5, and it is available in the Claude apps, Claude Code, the API, AWS, Google Cloud and Microsoft Azure.
How much does Claude Sonnet 5.5 cost?
The same as Sonnet 5: $2 per million input tokens and $10 per million output, with cache reads at $0.20. The saving comes per task, because the model uses fewer tokens and tool calls to finish the same work, which Anthropic puts at up to 30% less for most jobs.
Is Claude Sonnet 5.5 faster than Sonnet 5?
Yes. Anthropic says output generates more than 30% faster than Sonnet 5, making it the fastest Sonnet model so far. It also works in fewer steps, so agent runs tend to finish sooner end to end.
Should I use Sonnet 5.5 or Opus 5.5?
For everyday coding, bug fixes and documents, Sonnet 5.5 is the better value: half the token price, faster, and within a couple of points of Opus on practical work tests. For complex, open-ended projects where judgment matters most, Opus 5.5 is still the stronger pick.
When is Claude Haiku 5.5 coming?
Anthropic has said Haiku 5.5 arrives in the coming weeks. It has not given an exact date or pricing yet.
Sources
- Claude — Anthropic
- Claude models overview — Anthropic docs
- Claude Sonnet 5.5 nears Opus 5.5 at half the price (Dev.to)
My Personal Take
Anthropic finally selling speed instead of IQ points is the smartest move they've made in a while. For real daily work — coding, drafting, thinking through problems — latency shapes your experience far more than another few points on a reasoning benchmark. Same token prices plus noticeably faster answers is a quiet upgrade that most users will feel more than any 'smartest model ever' launch. This is the workhorse release, and workhorses win.
The honest caveat: if your work needs maximum reasoning depth, this isn't that model — it's the fast, affordable default, not the deep thinker. That's a fine trade for 90% of tasks. My take: make Sonnet 5.5 your first try for everyday work and only reach for heavier models when a task actually defeats it. Most of the time, it won't.
0 Comments