Claude Sonnet 5.5 Is Faster and Cheaper to Run. Here's What the Numbers Do and Don't Prove

Anthropic's new mid-tier model beats its own flagship on one coding benchmark and runs 30% faster than its predecessor. But the evidence so far comes mostly from Anthropic itself.

Anthropic released Claude Sonnet 5.5 on September 28, 2026. It is the second model in the Claude 5.5 family and the successor to Sonnet 5, available today on the Claude Platform (claude-sonnet-5-5) and through Amazon Web Services, Google Cloud, and Microsoft Azure. The pitch is efficiency: Anthropic says it generates output more than 30% faster than Sonnet 5 and costs up to 30% less per task, even though the list price is unchanged at $2 per million input tokens and $10 per million output tokens. The savings come from the model needing far fewer tokens to finish the same work.

Beyond speed and cost, there are four changes worth knowing about. Coding is the biggest jump, with the model scoring 70.6% on Terminal-Bench 4.0, up from Sonnet 5's 10.3%. Knowledge work, like building slide decks, spreadsheets, and documents, now lands close to Anthropic's flagship Opus 5.5. Writing is clearer, and early testers describe it as a better collaborator. And it is the first Sonnet model to ship with cybersecurity safeguards, which means some higher-risk security requests will automatically fall back to the older Sonnet 5.

The benchmarks that matter

Anthropic reports a long list of results. These are the ones that map to real work, compared against its previous model, its flagship, and OpenAI's GPT-6 Sol where a score exists:

A few things stand out:

• The leap over Sonnet 5 is enormous, especially on Terminal-Bench and chart reading. That says more about how weak Sonnet 5 was on those tests than about a miracle in Sonnet 5.5.
• It nearly matches Opus 5.5 at much lower cost. On CursorBench it sits within about two points of the flagship, and Anthropic says it reaches comparable scores on several tests at a fraction of the price per task.
• It even tops Opus 5.5 on Terminal-Bench, though Anthropic's footnote notes Opus's score there is at one particular effort level, so this isn't a clean head-to-head.
• Not everything is a win. On FrontierCode, Sonnet 5.5 stays behind Opus 5.5, and its score at the highest "Max" effort setting is lower than at "Xhigh." Anthropic explains that more thorough reviewing led to extra edits that the benchmark penalizes.

Reading these numbers carefully

• The scores are mostly self-reported. GDPval-AA was run by the outside firm Artificial Analysis, but on a pre-release version that had a bug, which Anthropic says has since been fixed and likely understates the model.
• In some comparisons, GPT-5.6 Sol is shown instead of GPT-6 Sol because the newer model's results weren't public. Not every row compares the same rivals.
• Anthropic itself says benchmarks capture only one side of a model, and that Opus 5.5 is still clearly stronger at complex, open-ended work that needs sustained judgment.
• Some GPT-6 Sol numbers may be out of date, since OpenAI recently fixed an image-understanding bug.

What the numbers don't show

Early testers quoted by Anthropic mostly describe efficiency, not raw intelligence. Base44 said its app builds needed about 3.6 iterations on average versus 7.7 for Opus 5. Balyasny Asset Management reported roughly 121,000 tokens per answer on finance tasks, versus 497,000 for Sonnet 5. Box reported it ran 2.4 times faster than the previous model, and Zendesk said support tickets were processed 20% faster. These are company-supplied testimonials, so treat them as signals rather than proof.

Anthropic also describes an internal test where the model turned a company's earnings materials and a slide template into a 10-slide operating review, and two experts judged the first draft ready to send.

[YOUR SECTION: Test it yourself. Try three tasks you actually care about, such as fixing a bug in a React Native project, summarizing a long document, or building a small UI. Report what worked and at least one thing that didn't. This is the part of your article no press release can give readers.]

Limitations and open questions

• Independent verification is thin. Most results come from Anthropic, and outside testing is still catching up.
• Safeguards will affect some users. Higher-risk cybersecurity tasks fall back to Sonnet 5, and Anthropic says its biology filters can flag some microbiology and virology requests by mistake. Approved security teams and life-science organizations can apply for broader access.
• Alignment testing has limits. Anthropic's behavioral audit covered roughly 1,850 scenarios, and Sonnet 5.5 matched or improved on Sonnet 5 on most measures, but Opus 5.5 still scored slightly better overall. Anthropic also acknowledges that no set of tests catches every failure.
• There are migration details. Developers running Sonnet with thinking turned off need to switch to a new setting called between_tools, and a new anti-distillation measure ties the model's reasoning to the account that created it, which could affect people who move conversations between accounts.
• The family isn't complete. A smaller Haiku 5.5 is promised "in the coming weeks."

The takeaway

Sonnet 5.5 looks like a strong upgrade for anyone doing everyday coding, document work, or high-volume tasks where cost and speed matter. For the hardest, most open-ended problems, Anthropic's own data says the pricier Opus 5.5 is still the better pick. The most interesting claim, that a cheaper model can rival a flagship at a fraction of the cost, deserves independent confirmation before you rebuild anything around it.

---

Source: Anthropic, "Introducing Claude Sonnet 5.5" (anthropic.com/claude-sonnet-5-5), September 28, 2026.