Claude Sonnet 5.5 Gets Within a Hair of Opus 5.5, and the Free Tier Just Got Much Better
Quick verdict
Anthropic released Claude Sonnet 5.5, the second model in the 5.5 family, one week after Opus 5.5 and the day before OpenAI DevDay. Early independent evals put it at or near Opus 5.5 while keeping Sonnet 5's price of $2 per million input tokens and $12 per million output. Anthropic says it runs more than 30% faster and costs up to 30% less per task, and it now powers the free tier on claude.ai. If you pay for AI by the task, this is the more interesting launch of the two, because a mid-tier model landing this close to the flagship changes the default answer to "which model should I use." The one thing to know before you switch: leave it on medium effort, because the savings evaporate at the top setting.
What actually shipped
Sonnet 5.5 is live on the Claude Platform, in Claude Code, and across a long list of third-party tools on day one. Anthropic frames it for well-scoped everyday work, the bug fixes and quick feature iterations that make up most of what people actually do, and points you to Opus 5.5 for the harder, open-ended problems.
- Priced at $2/$12 per million input/output tokens, unchanged from Sonnet 5
- 1M-token context window and 128K max output
- About 30% faster output and up to 30% lower cost per task, because it needs fewer tokens for the same work
- Now the model behind the free tier on claude.ai
- Shipped with a usage reset good through Oct 22 for existing subscribers
Availability was wide from the start: GitHub Copilot in VS Code, Cursor, Factory, Devin, Cline, and Arena's agent modes all had it on launch day. Anthropic also extended its "preserved thinking" work to counter distillation through account-switching, so reasoning traces stay with the org that generated them rather than leaking to another account. Mike Krieger said Haiku 5.5 will round out the family in the coming weeks.
Where it lands on benchmarks
The headline numbers are strong. On the Vals Index it debuted at #2, 0.47 points behind Opus 5.5, giving Anthropic the top four spots. Artificial Analysis put Sonnet 5.5 at max effort at #2 on its Intelligence Index, ahead of GPT-6 Astra Max and just below Opus 5.5. On agentic coding the story is the same shape: Cline reported it beats Opus 5.5 on Terminal-Bench 4.0 at half the cost, and Cognition's FrontierCode 1.1 put it at 64.4%, up from Sonnet 5's 56.2%.
The vision numbers are the cleanest illustration of the value pitch. On Roboflow's eval, Sonnet 5.5 did comparable accuracy work for roughly half the cost of Opus 5.5.
| Model and effort | Cost per 1,000 images | Latency |
|---|---|---|
| Sonnet 5.5, high | $8.82 | 8.9s |
| Opus 5.5, high | $18.12 | 12.2s |
| Sonnet 5.5, low | $6.53 | 10.8s |
| Opus 5.5, low | $13.80 | 12.8s |
The catch is max effort
Those top leaderboard placements are max-effort numbers, and that setting comes with a token bill Anthropic staff are openly telling people to avoid. Artificial Analysis measured Sonnet 5.5 at max using about 193,000 output tokens per task, the most they have ever recorded and roughly 7x GPT-6 Astra at max. At low, medium, and high effort it sits behind GPT-6 Sol's higher settings on the intelligence-versus-tokens curve, meaning Sol delivers more performance for fewer tokens in that region.
The advice from Anthropic was blunt. Edwin Arbus said not to run Sonnet at max effort, and that if you are reaching for max you should probably be using Opus instead. Claude Code defaults to medium, and several testers who dropped it a level found it stopped overthinking. So the cheap, fast Sonnet everyone is excited about is the medium-effort Sonnet. The chart-topping Sonnet is an expensive way to approximate a model you already own.
The free tier is the quiet win
Simon Willison flagged the change that will reach the most people: Sonnet 5.5 now powers the free tier on claude.ai, while ChatGPT's free tier still runs GPT-5.6 Luna, which he calls a lot less capable. For anyone who has only tried the free version of either assistant, the gap just widened in Anthropic's favor. That matters more than a leaderboard row, because most people never touch the API and judge these tools entirely by the free experience.
Why it matters
Step back and the pattern is hard to miss: the model itself is becoming a commodity faster than anyone expected. A mid-tier model matching last week's flagship on several evals, at a quarter to half the cost on real tasks, is the whole story of 2026 pricing in one launch. It landed hours before OpenAI DevDay, which is not a coincidence, and it arrived against a backdrop where Anthropic's own reported numbers show compute eating most of its costs, so per-task savings are strategic, not just a marketing line. Users are blunt about the caveat, though: Sonnet still is not Opus. It handles well-defined tasks well but lacks the general intuition Opus 5.5 brings to messy, open-ended problems. So the right move is not to pick one model and commit. It is to route each task to the cheapest model that clears the bar, which is exactly what an AI aggregator is built to do, and it is why the future of AI pricing looks like per-task cost rather than a flat monthly seat.
Video: Sonnet 5.5 vs Opus 5.5 on real tasks
A hands-on comparison of the two models on actual work, which is a useful counterweight to launch-day benchmark noise.
FAQ
Is Claude Sonnet 5.5 as good as Opus 5.5?
On benchmarks it is close. It debuted at #2 on the Vals Index, 0.47 points behind Opus 5.5, and it beats Opus on Terminal-Bench 4.0 at half the cost. In daily use, testers say it is excellent on well-defined tasks but still lacks the general intuition Opus 5.5 shows on open-ended problems. For most everyday work it is close enough that the price difference decides it.
How much does Sonnet 5.5 cost?
It is $2 per million input tokens and $12 per million output, the same as Sonnet 5. That happens to match OpenAI's GPT-6 Sol pricing. Because it uses fewer tokens per task, Anthropic says most work comes out up to 30% cheaper than Sonnet 5 even though the per-token rate did not change.
Should I run Sonnet 5.5 at max effort?
No. Anthropic staff explicitly told people not to. At max effort it uses about 193,000 tokens per task, which erases the speed and cost advantage, and its own team suggests you should be using Opus at that point. Medium is the default in Claude Code and the setting where the value shows up.
Is Sonnet 5.5 on the free plan?
Yes. It now powers the free tier on claude.ai, which is a meaningful upgrade over what free users had before and more capable than ChatGPT's current free model. If you have only used the free version of these tools, this is the change you will notice.
Sources
- @claudeai - Sonnet 5.5 launch, "a clear upgrade over Sonnet 5"
- @AnthropicAI - more than 30% faster and up to 30% less for most work
- @ValsAI - specs: 1M context, 128K output, unchanged $2/$12 pricing
- @ValsAI - Vals Index debut at #2, 0.47 behind Opus 5.5
- @kimmonismus - same price as OpenAI's GPT-6 Sol
- @_catwu - ~30% more tasks done in Claude Code on fewer tokens
- @ArtificialAnlys - ~193K output tokens per task at max effort, ~7x GPT-6 Astra
- @edwinarbus - "do not use Sonnet with max effort"
- @simonw - Sonnet 5.5 now powers the free tier on claude.ai
- @cline - beats Opus 5.5 on Terminal-Bench 4.0 at half the cost
- @cognition - FrontierCode 1.1 at 64.4%, up from Sonnet 5's 56.2%
- @skalskip92 - Roboflow vision cost per 1,000 images
- @rishdotblog - useful for defined tasks, "nowhere near opus" on intuition
- @mikeyk - Haiku 5.5 to round out the family in the coming weeks
Further reading
Try all the models mentioned in this article
Admix gives you GPT-5, Claude, Gemini, and 350+ AI models in one app. Compare responses side by side. Free to start.
Start free on Admix