Claude Sonnet 5.5 is Anthropic’s new mid-tier AI model, released on September 28, 2026, and it costs the same as Sonnet 5 while running faster and using fewer tokens per task. On the benchmarks Anthropic and press outlets published, it lands within a few points of the pricier Opus 5.5.
If you pay for Claude through the API, that mix matters more than any single score.
Here are the facts to know before you read on:
- Released September 28, 2026, as
claude-sonnet-5-5 - Price: $2 per million input tokens and $10 per million output tokens, unchanged from Sonnet 5
- Context window: 1M tokens, with up to 128K tokens of output
- Available on the Claude Platform, AWS, Google Cloud, and Microsoft Azure
- Closed weights, so you can’t self-host it
The price didn’t change, but the cost of finishing a job did. That’s the real story, and the next section explains why.
Why Claude Sonnet 5.5 can cost less without a price cut
Anthropic kept the per-token rates flat. The savings come from efficiency: the model generates output more than 30% faster and makes fewer tool calls, so it burns fewer tokens to reach an answer.
VentureBeat reports a roughly 30% drop in cost per task from those two changes. Anthropic’s own claim is bolder. On several evaluations, Sonnet 5.5 at Low or Medium effort can beat Sonnet 5’s best score at about one-tenth the cost per task, according to VentureBeat’s launch coverage.
Treat the one-tenth figure as a best case. Your savings depend on how many tool calls your workflow makes today, and agent-style jobs with long tool loops stand to gain the most.
How close is it to Opus 5.5?
Very close on the scores that have been published. The table below compares the two models, using figures from MarkTechPost and VentureBeat.
| Benchmark | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| GDPval-AA | 1844 | 1846 |
| OSWorld 2.1 | 80.1% | 81.8% |
| Chartography | 61.6% | 64.4% |
| FrontierCode | 52.1% | 54.4% |
The gaps run from 2 to under 3 points in every row. MarkTechPost also describes Sonnet 5.5 as delivering this at about half the price of Opus 5.5.
These are vendor-reported and press-reported numbers, not independent tests. Run your own prompts before you move a production workload.
The cybersecurity catch
This is the first Sonnet with built-in cybersecurity safeguards. If a request crosses Anthropic’s risk threshold, it falls back to Sonnet 5 instead of Sonnet 5.5.
Security teams doing legitimate defensive work can apply to Anthropic’s Cyber Verification Program for expanded access. Everyone else may see the fallback now and then on security-flavored prompts.
Sonnet 5.5 also uses classifiers meant to stop people from extracting its reasoning. For normal chat, writing, and coding, you probably won’t notice either feature.
Who should switch to Claude Sonnet 5.5
If you’re on Sonnet 5 through the API, switching is the easy call. Same price, faster output, and fewer tokens per job.
If you’re on Opus 5.5 for cost-sensitive agent work, test Sonnet 5.5 on your own tasks first. A 2-3 point gap on a benchmark may not show up in your use case, and the savings could be large.
Keep Opus 5.5 for the hardest jobs, where that last bit of accuracy pays for itself. For background on how Anthropic’s pricing has shifted lately, see our look at the Claude Fable 5.1 update, and for how a rival’s cheaper tier compares, read our GPT-6.1 Sol price and benchmark breakdown.
What we still don’t know
Most numbers here come from Anthropic’s launch and the outlets covering it. Independent testing usually takes a few weeks, and results on real workloads can look different from lab scores.
One figure deserves a pause. Coverage lists Terminal-Bench 4.0 at 70.6% for Sonnet 5.5 against just 10.3% for Sonnet 5. A jump that large may reflect a changed test setup or heavier tool use, so don’t read it as a simple sevenfold gain in coding skill.
The sources I checked also don’t spell out how the model rolls out in the Claude apps versus the API. If you mainly chat in the app, check your model picker rather than assuming it’s there.
How to test it without wasting money
Take ten real tasks from your own work, the kind you’d actually send to the model, and run them on both Sonnet 5.5 and your current model. Track the answer quality and the tokens used, not just the sticker price.
Then try Sonnet 5.5 at Low or Medium effort, since Anthropic’s biggest savings claim is tied to those settings. If quality holds, you’ve found your cheaper default.
Frequently Asked Questions
How much does Claude Sonnet 5.5 cost?
The API price is $2 per million input tokens and $10 per million output tokens. That’s the same as Sonnet 5.
What is the Claude Sonnet 5.5 API model name?
Use claude-sonnet-5-5. It’s live on the Claude Platform, AWS, Google Cloud, and Microsoft Azure.
Is Sonnet 5.5 better than Opus 5.5?
No, but it’s close. Opus 5.5 scores slightly higher on every benchmark listed above, while Sonnet 5.5 costs about half as much.
Can I run Sonnet 5.5 on my own hardware?
No. The weights are closed, so you can only use it through Anthropic or the listed cloud platforms.
My take: Sonnet 5.5 is the one to try first for most API work. Move to Opus 5.5 only when your own tests show a gap worth paying double for.
Leave a Reply