Anthropic
Blogcast

Alex: Anthropic just dropped Claude Sonnet five point five, and the numbers are kind of wild. We're talking about the everyday workhorse model here, not the flagship, and it just posted a seventy point six percent score on Terminal Bench, which is an agentic coding evaluation. For context, the old Sonnet five scored ten point three percent. That is not an incremental bump. That is a different animal.

Jordan: Yeah, and here's the part that got my attention. That seventy point six actually beats Opus five point five, the flagship, which scored sixty six point four on the same test. Now, Anthropic is quick to say Opus is still the stronger model for complex, open-ended work that needs sustained judgment. But on this particular coding eval, the mid-tier model is out in front. That is a strange and interesting world to be living in.

Alex: It really is. And it is not just coding. On GDPval, which tests real-world knowledge work across forty-four occupations and nine major industries, Sonnet five point five scored eighteen forty-four, basically neck and neck with Opus five point five at eighteen forty-six, and about four hundred points above the old Sonnet five. So for everyday knowledge work, you are getting flagship-adjacent performance at a fraction of the price.

Jordan: Let's talk about that price piece, because this is where the story gets really interesting for anyone actually running this stuff in production. The list price is the same as Sonnet five: two dollars per million input tokens, ten dollars per million output tokens, and twenty cents per million for cache reads. But Sonnet five point five needs far fewer tokens to do the same work. Anthropic says it costs up to thirty percent less per task. And on some benchmarks, at low or medium effort, it beats Sonnet five's best score for about a tenth of the cost per task.

Alex: A tenth. And it also generates output thirty percent faster, which makes it the fastest Sonnet they have ever shipped. There is an effort-level system, low, medium, high, extra high, and max, that lets you balance cost and speed against quality. Medium is the default in the Claude apps. And the sweet spot seems to be running Sonnet five point five at lower effort, where it complements Opus beautifully. Opus handles the gnarly judgment calls, Sonnet handles everything else, cheap and fast.

Jordan: The coding numbers deserve a second look too. On FrontierCode, at high effort, it scores ten points higher than Sonnet five at the same setting, at about one fifteenth of the cost per task. On CursorBench, which is built from real Cursor coding sessions, its best score is within about two points of Opus five point five. And early testers noticed it batches tool calls together more aggressively than Sonnet five, so fewer steps and lower costs.

Alex: There is a great quote from Epic Games in the announcement. Their chief operating officer, Daniel Vogel, said Sonnet five point five cleared the same quality bar you would expect from a higher-tier model. It held up on a system design audit and a data flow review, managed tens of thousands of lines of code for gameplay system architecture, kept responses snappy, and handled multi-hour tasks with less prescriptive prompting. Tens of thousands of lines of game code. That is not a toy demo.

Jordan: And Slack's principal engineer, Curtis Allen, said that without changing any of their prompts, Sonnet five point five did better than Sonnet five on almost all of their offline Slackbot evals, in fewer steps, with about fourteen percent fewer output tokens. When someone gives Slackbot a task, quality and speed are what matter most, and the new model delivers better outcomes for users, faster. That is the dream sentence for anyone running AI in production. Better, faster, cheaper, and no prompt changes required.

Alex: There is also a fun detail I love. It is the first Sonnet model to beat Pokemon Red working only from screenshots. Long-horizon work plus image understanding. It is a benchmark, sure, but it tells you something about how far the mid-tier models have come.

Jordan: On the knowledge-work side, testers kept mentioning two things that do not show up in benchmarks. It is a more natural conversational partner, and it has a real knack for design. Anthropic ran an internal test where they gave it a public company's quarterly earnings materials and call transcripts, plus a slide template, and asked for a ten-slide operating review. Two experts judged the first draft ready to send as is. If you have ever wrestled a deck into shape at midnight, you know how absurd that sounds.

Alex: Okay, so what about safety? Because this model is a big jump in capability, especially in cybersecurity, Anthropic is shipping it with safeguards similar to the ones on Opus five point five. It is the first Sonnet to launch with that level of cyber safeguard. Routine software development is unaffected, you can still find and fix bugs in your own code, but higher-risk cybersecurity tasks will visibly fall back to Sonnet five. Biology safeguards are the same as Sonnet five's, targeting a narrow set of high-risk requests, with most research, education, and clinical work unaffected.

Jordan: And there is a new one I found interesting: distillation protection. Sonnet five point five is the first Sonnet to launch with safety classifiers that prevent reasoning extraction. That is basically stopping attackers from using thousands of fake accounts to siphon off the model's capabilities at industrial scale. It also expands something called preserved thinking, so the model's thinking cannot be decoupled from the account that created it. Most developers will not notice a change, but it is a meaningful hardening move.

Alex: Alignment-wise, on their automated behavioral audit across roughly one thousand eight hundred fifty scenarios, Sonnet five point five improves on or matches Sonnet five on most measures. And on containment evaluations, it came close to Opus five point five in how rarely it tries to escape its sandbox. It was actually the least likely of any of their models to probe the limits of its containers. Opus still performs slightly better overall, but they found no evidence of the model pursuing goals that conflict with the user's intent.

Jordan: So where does this leave us? The framing Anthropic wants is pretty clear. Opus five point five is for complex work requiring careful judgment. Sonnet five point five is the fast, cheap, sharp complement for well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets, with a sharp eye for design. And Haiku five point five, built for high-volume and cost-sensitive applications, joins the family in the coming weeks.

Alex: The way I would put it is this. The middle of the lineup just got scary good. When your mid-tier model is beating the flagship on coding evals, matching the flagship on knowledge work, and doing it at a tenth of the cost per task, the question for builders is not whether to upgrade. It is what you do with all the budget you just got back.

Jordan: It is available now on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure, and on the Claude platform as claude-sonnet-five-five. One heads-up for developers: if you run Sonnet with thinking off, you will need to switch to the new between-tools setting before migrating. There is a migration guide in the platform docs.

Alex: That is the rundown on Claude Sonnet five point five. Faster, cheaper, sharper, and the mid-tier just became the most interesting tier. Thanks for listening, and we will catch you on the next one.