Claude Sonnet 5.5 just did something no mid-tier model was supposed to do: it out-coded its own flagship. The cheap one scored higher than Opus 5.5 on the toughest agentic coding test Anthropic runs, and it costs half as much per token.
If you pay for Claude (or you’re deciding whether to start), this changes the math. Here’s what actually shipped, what it costs, and which model deserves your money now.
What Anthropic shipped
Anthropic released Claude Sonnet 5.5 on September 28, six days after Opus 5.5 landed. It replaces Sonnet 5 as the middle child of the lineup, with Haiku 5.5 promised in the coming weeks to round out the family.
On paper, this looks like a routine refresh. In practice, it’s the strongest value proposition Anthropic has ever offered beginners. The model writes 30%+ faster than Sonnet 5, and it needs fewer tokens to finish the same job, so typical tasks come in up to 30% cheaper even though the price list didn’t move. Context window sits at a million tokens, which is enough to swallow an entire codebase or a 700-page PDF without breaking a sweat.
It’s live everywhere you’d expect: the Claude apps, the Claude Platform, Amazon Bedrock, Google Cloud, and Microsoft Azure.
The score that broke the pattern
Terminal-Bench 4.0 measures whether an AI can finish real, multi-step work in a command line: debugging a server, wiring up a deploy, that kind of thing. It maps closely to how people actually use Claude Code.
Claude Sonnet 5.5 scored 70.6%. Opus 5.5, the model that costs twice as much, scored 66.4%. Sonnet 5 managed 10.3%, which tells you how big this generation jump really is.
| Benchmark | Sonnet 5.5 | Sonnet 5 | Opus 5.5 | GPT-6 Sol |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 70.6% | 10.3% | 66.4% | not reported |
| GDPval-AA (Elo) | 1,844 | 1,449 | 1,846 | 1,487 |
| OSWorld 2.1 (computer use) | 80.1% | 57.0% | 81.8% | not reported |
| Chartography (charts) | 61.6% | 15.6% | 64.4% | 53.6% |
| FrontierCode (xhigh effort) | 52.1% | 42.4% | 54.4% | 49.3% |
One benchmark win doesn’t make the cheap model the best model overall. Opus 5.5 still edges it on knowledge work, computer use, and chart reading, usually by a couple of points. Anthropic says Opus stays “clearly stronger” at open-ended work that needs sustained judgment. But the gap is now measured in single points, not leagues.
Independent testers mostly back this up. Artificial Analysis ran its own Terminal-Bench pass and got the same ordering: Sonnet 5.5 at 63.6%, ahead of Opus 5.5 at 59.6% and GPT-6 Astra at 59.1%. Different numbers, same story.
Claude Sonnet 5.5 pricing (no, it didn’t get cheaper per token)
The rate card is identical to Sonnet 5. You pay $2 per million input tokens and $10 per million output tokens. Cache reads run $0.20 per million. Batch jobs take 50% off the whole bill.
So where does “up to 30% cheaper” come from? Token efficiency. The model finishes tasks with fewer tokens, so the same job lands on a smaller invoice even at the same rate. Think of it as a car with the same price per liter but better mileage.
Against Opus 5.5, the gap is brutal:
| Price per 1M tokens | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| Input | $2 | $4 |
| Output | $10 | $20 |
| Cache write (5 min) | $2.50 | $5 |
| Cache read | $0.20 | $0.20 |
Half price, near-flagship output on most everyday work. For a solo founder running automations, that difference compounds fast across a month of API calls.
The effort setting trick
Here’s the part most launch coverage buried, and it’s the part that saves you actual money. Claude models now ship with an effort dial: low, medium, high, xhigh, max. Higher settings make the model think longer, score better on hard tasks, and burn way more tokens.
Anthropic’s own testing shows Claude Sonnet 5.5 at medium effort beats Sonnet 5’s best-ever score for less than a tenth of the cost per task. Medium is the default in the Claude apps and Claude Code, so if you’re a regular user, you’re already on the good setting without touching anything.
API users need to pay attention, though. The Claude Platform defaults to high effort, and independent testing found high roughly doubles output tokens on routine generation for no measurable quality gain. Meanwhile, max effort is a wallet incinerator: one independent run measured about 193,000 output tokens per task at max, roughly $7.60 each, which was 50% MORE than Sonnet 5 costs for the same tasks.
The practical recipe is short. Start at medium. Move up to high or xhigh only for long, open-ended work that’s genuinely hard. Skip max unless you’re running formal evaluations and you like spending money.
Sonnet 5.5 vs Opus 5.5 (when to pay double)
Both models have a place. The split comes down to how well-defined your task is.
Use Claude Sonnet 5.5 when the work has clear boundaries: fix this bug, clean this spreadsheet, build this slide deck, draft this proposal, summarize these documents. It’s tuned for well-scoped everyday tasks, and it produces polished documents and spreadsheets with a decent eye for design. This is 80% of what most people do with AI anyway.
Pay for Opus 5.5 when the problem is fuzzy and judgment-heavy: multi-day research where the questions shift, complex strategy work, or agentic runs that need to make careful calls for hours without supervision. You’re paying double for sustained reasoning quality, and on those tasks it earns it.
Still torn? Run both for a week. Set Sonnet 5.5 as your default, and when a response feels shallow, re-run that one task on Opus. Most people discover Opus earns its premium on maybe one task in ten.
How to switch today
In the Claude apps, open the model picker and select Sonnet 5.5. It’s generally available, not a preview, and it replaces Sonnet 5 everywhere. Developers on the API should point at the model ID claude-sonnet-5-5.
Two migration notes if you build things. Thinking can no longer be fully switched off on this model; you’ll want the new between_tools setting if you relied on thinking-off mode. And forced tool calls now return an error instead of silently degrading, so old workflows using that pattern need a small update.
If you want the full benchmark detail straight from the source, Anthropic’s official Sonnet 5.5 announcement has the complete tables.
The one warning worth reading
The “30% cheaper” headline only holds at lower effort settings. Crank the dial to max and Claude Sonnet 5.5 becomes one of the most token-hungry models ever measured, as the independent cost data above shows. Efficiency here isn’t a property of the model alone; it’s a property of the model plus the setting you pick.
Second warning, smaller but real: Sonnet 5.5 trails Opus on factual recall benchmarks. It hallucinates less than Opus did in the same tests, which is nice, but for fact-dense work you should still verify claims the old-fashioned way.
Which Claude should you actually use?
Default to Claude Sonnet 5.5 and keep Opus 5.5 for the rare, genuinely hard jobs. At $2/$10 with flagship-adjacent scores and a million-token context, it’s the best value in the Claude lineup by a wide margin right now. Set it today, leave the effort dial at medium, and watch what happens to your monthly bill.
If you’re still comparing across companies, our Opus 5.5 vs GPT-6 Sol breakdown covers the flagship matchup, and these 7 underrated Claude features pair nicely with the new model.
