Claude Sonnet 5.5 is the faster everyday model

Claude Sonnet 5.5 is Anthropic's model for scoped everyday work, not a new top of the line. The launch page is dated September 28, 2026. CNBC's story on the same release is timestamped 2:00 p.m. Eastern that Monday, which is 9:00 p.m. in Moscow.

TL;DR - Anthropic says the release is more than 30 percent faster than Sonnet 5 and, in company tests, up to 30 percent cheaper per task. - The sticker price matches Sonnet 5: $2 per million input tokens, $10 per million output tokens, and $0.20 per million tokens for cache reads. - The company says this model does not advance the frontier of its capabilities, but it is the first Sonnet shipped with cyber safeguards like those on its strongest models.

What actually happened

Anthropic calls the release a clear upgrade over Claude Sonnet 5 and a faster, lower-cost complement to Claude Opus 5.5. Opus 5.5 is for complex work that needs careful judgment. The new tier is described as strongest at well-scoped tasks: fixing bugs, and producing polished documents, slides, and spreadsheets. Claude Haiku 5.5, for high-volume work, is said to arrive in the coming weeks.

CNBC framed the drop as Anthropic's second launch since chief executive Dario Amodei urged companies to slow how fast they improve their most advanced models. The network says it came less than a week after Opus 5.5. The launch page does not count that gap in days.

Theo Chu, a research product manager, told CNBC that Sonnet is for the cost-conscious customer who may not need as much intelligence. He described the fit as routine execution rather than the judgment Opus is sold for. That interview is the main fact the news story adds. CNBC's price line matches the launch table. Opus 5.5 is listed at $4 per million input tokens and $20 per million output tokens. Cache writes are $5 on Opus 5.5 and $2.50 on the new Sonnet. Cache reads are $0.20 on both. "Half the price of Opus 5.5" is that input and output pair, not every cache line.

Anthropic says outputs come more than 30 percent faster than Sonnet 5, and that this is its fastest Sonnet. The per-token price is unchanged from Sonnet 5, but the model typically needs far fewer tokens. In company testing, that is up to 30 percent less per task. Neither page links an outside audit of those bills.

Both pages say zero data retention is available, as with Opus 5.5 and Sonnet 5. The model is on platforms including Amazon Web Services, Google Cloud, and Microsoft Azure. The identifier is claude-sonnet-5-5. Anyone who ran Sonnet with thinking off must switch to between_tools, which keeps up-front thinking off. CNBC does not mention that setting.

What the score table actually shows

Anthropic says scores are only one facet of a model. In its testing and in outside testing, Opus 5.5 remains clearly stronger at open-ended work that needs sustained judgment. Claude Sonnet 5.5 is not presented as the replacement for that job.

Terminal-Bench 4.0, an agentic coding test, lists 70.6 percent against 10.3 percent for Sonnet 5. Opus 5.5 is at 66.4 percent, footnoted as the Xhigh-effort score, which the page calls that model's highest on the test. FrontierCode 1.1's main set, which asks whether code changes would be merged, lists 46.2 percent at Max effort and 52.1 percent at Xhigh. Sonnet 5 is at 42.4 percent and Opus 5.5 at 54.4 percent. A footnote says Max lands lower than Xhigh because the model more often ran a code-review skill split across subagents. In two cases Cognition examined, that produced a timeout or edits beyond the task.

CursorBench 4.0, drawn from real Cursor sessions, lists 55.5 percent, against 34.1 percent for Sonnet 5 and 57.8 percent for Opus 5.5. The prose says the best score sits within about two points of Opus 5.5. GDPval-AA v2.1 lists 1844, against 1449 for Sonnet 5 and 1846 for Opus 5.5. The page calls that two points below Opus 5.5 and about 400 points above Sonnet 5. The table gap versus Sonnet 5 is 395 points, so "about 400" is rounding, not a second study. AA-Briefcase v1.1 lists 1811, 1359, and 1822. Humanity's Last Exam with tools lists 64.5 percent, 54.9 percent, and 67.7 percent. OSWorld 2.1 partial scores are 80.1 percent, 57.0 percent, and 81.8 percent. Chartography with no tools lists 61.6 percent, 15.6 percent, and 64.4 percent.

A footnote says GDPval-AA and AA-Briefcase were run on a pre-release build with a structured-output bug. Anthropic expects any effect to understate the scores, and says the bug is fixed.

Effort changes the bill. On several benchmarks, low or medium effort beats Sonnet 5's best score for about a tenth of the cost per task. In Claude Code and the apps, default effort is medium. On the Claude Platform, the default is high. At high effort on FrontierCode, the page says the score is 10 points above Sonnet 5 at the same setting, at about one fifteenth of the cost per task.

Named customers on the page are anecdotes. Slack is cited for fewer Slackbot steps and about 14 percent fewer output tokens. A private set of 2,441 finance tasks is said to have used about 121,000 tokens per answer versus 497,000 for Sonnet 5. Unity is cited at 90 percent task completion on an editor benchmark. CNBC does not recheck those notes.

Safety claims, and what they do not cover

CNBC says Anthropic described Claude Sonnet 5.5 as not advancing the frontier, so most alignment testing looked at risks that apply at any capability level. An automated audit of roughly 1,850 scenarios is said to match or beat Sonnet 5 on most measures of alignment, misuse resistance, and honesty. On newer containment tests, the page says the model comes close to Opus 5.5 in how rarely it tries to leave a sandbox, and that it was the least likely of Anthropic's models to probe container limits. Opus 5.5 still does slightly better across the full audit. The company says it found no evidence of goals that conflict with the user, and that no test set catches every failure.

Cyber skills are called a large step up from Sonnet 5 and comparable to Opus 5. That is why this is the first Sonnet with cyber safeguards and fallbacks like those on the strongest models. Routine bug fixing is supposed to keep working. Higher-risk cybersecurity tasks will visibly fall back to Sonnet 5. A Cyber Verification Program is described as the later route to tiered access. Biology safeguards match Sonnet 5 and target a narrow set of requests. Some microbiology and virology prompts may be flagged in error.

The page also says this is the first Sonnet with classifiers meant to block reasoning extraction through many fake accounts. Preserved thinking stays tied to the account that created it. Chu's line to CNBC, that alignment has been a priority from the start, is a mission statement rather than a new score.

What to do as a reader (not a trade)

None of this is a reason to buy or sell a subscription, a stock, or a token because a benchmark moved. This is not investment advice. Replay one weekly task at the effort setting you will actually leave on. A medium default in the apps and a high default on the platform can share a model name and not share a bill. Do not treat "up to 30 percent less" as your invoice until the token counts are yours.

If the work is open-ended judgment, Anthropic is telling you Opus 5.5 is still stronger. If the work is scoped coding, documents, or support macros, Claude Sonnet 5.5 is the tier it wants tried. Security teams should test the fallback, because higher-risk requests are supposed to drop to Sonnet 5 even when the config name stays put. If thinking was off, read the migration note before you switch. between_tools is a breaking change the CNBC summary skips.

Other posts from the same news window, including how Gemini 4 Argon was rolled out and what Cohere shipped with Embed 5, are separate products. They do not confirm this table. The primary page is Anthropic's Claude Sonnet 5.5 announcement. CNBC's September 28 story tracks that page and adds Chu's interview. Overlapping figures above come from the launch page, not from the news recap.