Anthropic released Claude Sonnet 5.5 on Monday, and the number worth noting is not the price β it is a benchmark where the cheap model beats the expensive one. On Anthropic's own scorecard, Sonnet 5.5 posts 70.6% on Terminal-Bench 4.0, an agentic coding evaluation, against 66.4% for Claude Opus 5.5. Sonnet 5, three months old, managed 10.3% on the same test.
Key takeaways
- Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0 versus 66.4% for the more expensive Opus 5.5.
- Pricing is unchanged from Sonnet 5 at $2 per million input tokens and $10 per million output tokens; the cost saving comes from using fewer tokens, not a lower rate.
- Balyasny Asset Management reported answers costing 121,000 tokens where Sonnet 5 spent 497,000.
Where the 30% cost cut actually comes from
Anthropic's headline claims are 30%-plus faster generation and up to 30% lower cost per task. The second figure is easy to misread. The sticker price did not move: input stays at $2 per million tokens, output at $10, with cache reads at $0.20 and cache writes at $2.50. Savings come from the model finishing jobs with fewer tokens and fewer tool calls, which is a different kind of win β it only shows up on an invoice if your workload is agentic.
The customer figures Anthropic published make that concrete. Slack reported matching or beating Sonnet 5 across nearly all its evaluations in fewer steps and with roughly 14% fewer output tokens. Box measured results 2.4 times faster with 12% fewer total tokens. Lovable said tasks finished with a third fewer tool calls and about half the shell runs. The sharpest number came from Balyasny Asset Management: 121,000 tokens per answer where Sonnet 5 had burned 497,000.
Reading the benchmark table
Sonnet 5.5 lands close to Opus 5.5 nearly everywhere and ahead of it in one place. On OSWorld 2.1 it scores 80.1% to Opus 5.5's 81.8%, up from 57.0% for Sonnet 5. Humanity's Last Exam gives 64.5% against 67.7%. On GDPval-AA v2.1 the two are effectively tied at 1844 and 1846, and on AA-Briefcase v1.1 the gap is 1811 to 1822. Terminal-Bench is also the only one of the three tests Anthropic files under agentic coding that Sonnet 5.5 wins: Opus 5.5 leads on CursorBench 4.0, 57.8% to 55.5%, and on FrontierCode 1.1, 54.4% to 46.2%. Anthropic notes Sonnet 5.5 scores lower at Max effort than at Xhigh on FrontierCode, because at Max it more often ran a code-review skill that produced out-of-scope edits and timeouts, which the benchmark penalises.
The Sonnet 5 column is where caution is warranted. A jump from 10.3% to 70.6% on Terminal-Bench 4.0, or 15.6% to 61.6% on Chartography, says more about how poorly the previous generation handled those specific harnesses than about a sevenfold gain in capability. The like-for-like comparison to read is Sonnet 5.5 against Opus 5.5, and there the story is that a mid-tier model now sits within a couple of points of the flagship on most work β while spawning more parallel agents inside the same budget.
What Anthropic put behind a fallback
Sonnet 5.5 arrives with a constraint its predecessors did not have. Anthropic says higher-risk cybersecurity tasks will visibly fall back to Sonnet 5, an explicit downgrade rather than a refusal. TechCrunch reported that this makes Sonnet the first model in its tier to sit under the cyber safeguards previously reserved for Opus and Fable, on the basis that its cyber capability is now comparable to Opus 5. It is also the first Sonnet shipped with classifiers meant to stop rivals extracting its reasoning traces for distillation. Biology safeguards carry over from Sonnet 5 unchanged.
Context: the mid-tier is where the fight is
Nobody is competing on flagship pricing anymore. OpenAI halved its mid-tier token prices with Sol and Luna last week, and Meta announced a model to drive a smart-glasses feature. Anthropic's answer is not a cheaper rate card but a model that consumes less of it, backed by up to 90% savings with prompt caching and 50% with batch processing.
Outlook
Sonnet 5.5 is available on Claude.ai and, for developers, on the Claude Platform as well as Amazon Web Services, Google Cloud and Microsoft Azure, under the identifier claude-sonnet-5-5. A refreshed Haiku is expected within weeks, with no date committed. The open question is what Opus 5.5 is still for: if the cheap model takes the terminal-coding benchmark and ties on knowledge work, the flagship's remaining claim is judgment on problems that are hard to benchmark.
FAQ
Is Claude Sonnet 5.5 more expensive than Sonnet 5?
No. It costs the same: $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20 and cache writes at $2.50. Anthropic's claim of up to 30% lower cost per task comes from the model using fewer tokens and tool calls, not from a reduced rate.
Does Sonnet 5.5 really beat Opus 5.5?
On Terminal-Bench 4.0, an agentic coding evaluation, yes β 70.6% to 66.4%, and Anthropic reports that Opus 5.5 figure at its best effort setting. It is the only one of the three agentic coding benchmarks Anthropic published where Sonnet 5.5 comes out ahead; Opus 5.5 leads on CursorBench 4.0 and FrontierCode 1.1. Opus 5.5 also stays slightly ahead on OSWorld 2.1 and Humanity's Last Exam, by one to three points.
Where can developers use Sonnet 5.5?
It is available natively on the Claude Platform and through Amazon Web Services, Google Cloud and Microsoft Azure, using the model identifier claude-sonnet-5-5. Consumer access is via Claude.ai.
