Launches · September 29, 2026 · 15 min read
Anthropic kept Sonnet’s $2 and $10 token prices and posted 70.6% on Terminal-Bench 4.0, against 10.3% for Sonnet 5, sixteen days after Dario Amodei wrote that the industry has to slow the pace of capability gains.
On September 28 Anthropic released Claude Sonnet 5.5, the second model in what it calls the Claude 5.5 family. The announcement keeps the old Sonnet sticker price and says the new model runs more than 30 percent faster and costs up to 30 percent less per task in the company’s own tests. On September 12, Dario Amodei published an essay, “We Must Pace the Frontier,” whose plain sentence is that the labs have to slow the rate at which they make models more capable. On September 22 Anthropic shipped Opus 5.5 and described it as the first release since that call. Sonnet 5.5 arrived six days after Opus. Haiku 5.5 is still listed as later, in the coming weeks.
The sticker price did not move
Sonnet 5.5 is priced at $2 per million input tokens and $10 per million output tokens. Cache reads are $0.20 per million. Cache writes are $2.50. Anthropic says those input, output, and cache-read rates match Sonnet 5. Opus 5.5, on the same page, is $4 and $20, with cache writes at $5 and cache reads still at $0.20. The saving Anthropic is selling is not a cut in the rate card. It says Sonnet 5.5 needs far fewer tokens to finish the same work, and that in its testing the cost per task falls by as much as 30 percent. Output is also more than 30 percent faster, which the company calls the fastest Sonnet it has shipped.
The model id on the Claude Platform is claude-sonnet-5-5. Docs list a 1 million token context window, a 128,000 token output cap, a reliable knowledge cutoff of June 2026, and adaptive thinking with a default effort of high. The announcement says the model is on Amazon Web Services, Google Cloud, and Microsoft Azure, and that, like Opus 5.5 and Sonnet 5, it is offered with zero data retention. Platform docs add Microsoft Foundry alongside the API, Bedrock, and Claude Platform on AWS. Retirement is listed as not sooner than September 28, 2027.
What the model is for, in Anthropic’s wording, is the work under Opus. Opus 5.5 is described as the one for complex jobs that need careful judgment. Sonnet 5.5 is described as strongest at well-scoped everyday tasks, fixing bugs, and making documents, slides, and spreadsheets. The company also says it has a sharp eye for design: polish on interfaces, and slides that follow a template closely enough to need little editing. In one internal test Anthropic gave it a public company’s quarterly earnings materials, the call transcripts, and a slide template, and asked for a 10-slide operating review. Two experts, the company says, judged the first draft ready to send.
A few of the named early tests come with counts. Daniel Vogel, chief operating officer at Epic Games, is quoted saying the model held up on a system-design audit and a data-flow review, handled tens of thousands of lines for gameplay architecture, and needed less prescriptive prompting. Curtis Allen at Slack says that without prompt changes, Sonnet 5.5 beat Sonnet 5 on almost all of Slackbot’s offline evals, in fewer steps and with about 14 percent fewer output tokens. Abhinay Kathuria at Zendesk says tickets in a set of real support cases were processed 20 percent faster. Jamil Valliani at Atlassian says Rovo agents could run up to 30 percent faster than with Sonnet 5. Those are the customers’ reports as Anthropic printed them, not a third-party audit.
The table
The score table on the announcement is the part that does not fit in a slogan. Anthropic prints Sonnet 5.5 next to Sonnet 5, Opus 5.5, and GPT-6 Sol. Where OpenAI has not published a GPT-6 Sol number, the cell is a dash. The footnotes change how a few of the cells should be read, so the table is worth looking at with those notes attached.
Terminal-Bench 4.0, an agentic coding test in a command line, is the row that moves the most. Sonnet 5.5 is at 70.6 percent. Sonnet 5 is at 10.3 percent. Opus 5.5 is at 66.4 percent, and footnote 1 says that Opus figure is the model’s highest score, taken at Xhigh effort. The Sonnet 5.5 cell is not labeled with an effort level in the table. On the cost chart for the same test, the 70 percent region sits at Max. A reader who treats 70.6 as a like-for-like match against Opus at the same setting is ahead of what the footnote says. The jump over Sonnet 5 is the comparison Anthropic led with.
FrontierCode 1.1, main set, grades whether a code change could be merged, and it penalizes edits that wander outside the task. Sonnet 5.5 is listed twice: 46.2 percent at Max, and 52.1 percent at Xhigh. Sonnet 5 is at 42.4 percent. Opus 5.5 is at 54.4 percent. GPT-6 Sol is at 49.3 percent. The Max score is the lower of the two Sonnet 5.5 numbers. Footnote 2 says why. At Max effort the model more often ran Claude Code’s code-review skill, which splits the review across many subagents. In two cases Cognition looked at, that produced a timeout or extra edits past the task’s scope, and the benchmark scored those runs down. Extra effort, on this test, can cost points.
CursorBench 4.0 uses tasks from real Cursor sessions. Sonnet 5.5 is at 55.5 percent, Sonnet 5 at 34.1 percent, Opus 5.5 at 57.8 percent. Anthropic says the best Sonnet 5.5 score sits within about two points of Opus. GPT-6 Sol is a dash. On the matching chart the company plots GPT-5.6 Sol instead, and says GPT-6 Sol was not reported publicly.
The knowledge-work rows are Elo scores, not percents. GDPval-AA v2.1, real tasks across 44 occupations and nine industries, has Sonnet 5.5 at 1844, Sonnet 5 at 1449, Opus 5.5 at 1846, and GPT-6 Sol at 1487. Anthropic’s prose says Sonnet 5.5 scores two points below Opus and about 400 points above Sonnet 5. AA-Briefcase v1.1, a long-horizon knowledge-work test, is 1811, 1359, 1822, and 1483. Footnote 3 says Artificial Analysis ran both of those on a pre-release Sonnet 5.5 that had a bug in structured outputs. Anthropic expects any effect to be small and to understate the model. The bug, it says, has been fixed. Footnote 4 says OpenAI fixed an image-understanding bug in GPT-6 Sol, and that the official Artificial Analysis and Surge AI numbers may not yet reflect that fix. Anthropic’s own check of Chartography, it says, did not move.
Humanity’s Last Exam, with tools, is 64.5 percent for Sonnet 5.5, 54.9 percent for Sonnet 5, and 67.7 percent for Opus 5.5. OSWorld 2.1, partial credit for operating a desktop from screenshots, is 80.1, 57.0, and 81.8. Chartography, no tools, is 61.6 percent against 15.6 percent for Sonnet 5, 64.4 percent for Opus 5.5, and 53.6 percent for GPT-6 Sol. The announcement also says Sonnet 5.5 is the first Sonnet to beat Pokémon Red working only from screenshots. That claim is in the prose. It is not a row in the table.
Anthropic’s own caution sits under the table. It says Sonnet 5.5 at Max effort performs comparably to Opus 5.5 on several evaluations, and that benchmark scores are one facet of what a model can do. In the company’s testing and in external testers’ testing, Opus 5.5 “remains clearly stronger at complex, open-ended work requiring sustained judgment.” The mid-tier model can land next to the flagship on a scored sheet and still be the one Anthropic tells you not to use when the job is open-ended.
Where the line bends
The charts plot score against cost per task at each effort level. Anthropic’s rule of thumb is on the page: as effort rises, the model works longer, the bill rises, and the score usually rises with it. The attractive corner is the top left, more score per dollar. Effort in the Claude apps and in Claude Code defaults to Medium. On the Claude Platform the default is High. Lower settings answer faster and spend fewer tokens. Higher settings reason longer and check the work more.
On Terminal-Bench, Sonnet 5 sits along the bottom of the chart, in a band around 5 to 11 percent, while Sonnet 5.5 climbs from about 20 percent at Low effort through the 40s at High and into the 70s at Max. Opus 5.5 is above Sonnet 5.5 through most of that climb and then slips under it at the right edge. Anthropic’s caption says that at Medium effort, the default in the apps, Sonnet 5.5 far exceeds Sonnet 5’s best score for less than a tenth of the cost per task. The X post that carried this chart shortens the same claim: on several benchmarks, Low or Medium effort beats Sonnet 5’s best score for about a tenth of the cost. The chart is the place that sentence becomes a picture, and it is also the place where Opus is not uniformly ahead.
FrontierCode is the chart that refuses a simple “more effort, more score” story. Sonnet 5.5 rises from Low through High and Xhigh, then drops at Max, toward the mid-40s, while the bill keeps climbing past $10 a task on the log scale. Opus 5.5 stays in the low-to-mid 50s across the expensive end. Anthropic says that at High effort, the platform default, Sonnet 5.5 scores 10 points higher than Sonnet 5 at the same setting, at about one fifteenth of the cost per task, and that it matches GPT-6 Sol’s best score for about a fifth of the cost. The drop at Max is the footnote about the code-review skill, written as a line instead of a sentence.
CursorBench, which is not reprinted here, is the milder version of the same picture. Anthropic says Sonnet 5.5 at Low effort already beats Sonnet 5’s best score for less than a tenth of the cost, and that the best Sonnet 5.5 point lands within about two points of Opus 5.5. AA-Briefcase, also a chart on the page, has Sonnet 5.5 at Medium effort beating Sonnet 5’s best for about one ninth of the cost per task, with Opus and Sonnet 5.5 nearly on top of each other once both are spending several dollars a task. The company says Sonnet 5.5 complements Opus best at the lower effort settings, where it costs less per task, and that at higher settings it can perform comparably at a similar cost. That is a routing claim. Cheap and medium settings for the scoped job. Opus when the judgment has to hold.
Thinking stays attached to the account
Sonnet 5.5 does not offer a clean off switch for thinking. The platform docs list thinking as adaptive, and the default effort as high. The announcement is more specific about the migration. If a developer was running Sonnet with thinking off, that setting does not carry over. The replacement is between_tools, which keeps the up-front thinking off. People who do not change that call will not get the old behavior by doing nothing.
Effort is the dial Anthropic wants used instead. Medium is the default in Claude Code and the apps. High is the default on the platform. The charts above are what that dial does to the bill and the score. A team that leaves the platform default at High is not on the cheap left edge of those plots. A team that drops to Low or Medium is the team Anthropic’s “tenth of the cost” sentence is about.
The other change is about whose thinking it is. Anthropic says Sonnet 5.5 expands preserved thinking, so the model’s thinking cannot be separated from the account that created it. It also says this is the first Sonnet to launch with safety classifiers meant to stop reasoning extraction, the practice of using thousands of fake accounts to pull a model’s capabilities out at industrial scale and stand up a copy without Anthropic’s safeguards. The company says most developers will not notice. The case it flags is moving a conversation between accounts, including switching accounts mid-session in Claude Code. The docs, it says, explain that change.
Safeguards for a model they say is not the frontier
The alignment section uses a sentence that does a lot of work. “Sonnet 5.5 doesn’t advance the frontier of our models’ capabilities,” Anthropic writes, so the alignment review focused on risks that apply at any capability level: acting against the user, misleading the user, and helping with high-stakes misuse. The cyber section, a few paragraphs later, says the model’s cybersecurity capability is a large improvement over Sonnet 5 and comparable to Opus 5, which is why this is the first Sonnet to ship with the cyber safeguards and fallbacks used on the most capable models. Higher-risk cybersecurity tasks will visibly fall back to Sonnet 5. Routine software development, Anthropic says, is unaffected. Soon, cyberdefenders will be able to apply to an expanded Cyber Verification Program for tiered access to more of those capabilities on Sonnet 5.5, Opus 5.5, and the Claude Mythos models.
Biology safeguards are unchanged from Sonnet 5. They target harmful requests. Most research, education, and clinical work is supposed to pass. Some microbiology and virology requests may be flagged by mistake. Organizations can apply to the Life Sciences Verification Program for a wider set.
The behavioral audit covers roughly 1,850 scenarios. Anthropic says Sonnet 5.5 improves on or matches Sonnet 5 on most measures of alignment, resistance to misuse, and honesty. On newer containment tests it comes close to Opus 5.5, which the company calls the best model it tested, in how rarely it tries to escape a sandbox, and it is the least likely of Anthropic’s models to probe the limits of its containers. Across the full audit, Opus 5.5 still does slightly better. The company says it found no evidence that Sonnet 5.5 pursues goals that conflict with the user’s intention. It also says no set of evaluations catches every failure, and that this model may have tendencies the audit missed. That is why the safeguards sit next to the audit rather than instead of it.
“Doesn’t advance the frontier” and “70.6 percent against Opus 5.5’s 66.4” are both on the same page. They are not the same kind of claim. The first is Anthropic deciding that this release does not set a new ceiling for the lab, and that the cyber bar that matters is Opus 5, not the previous Sonnet. The second is one coding evaluation on which the published Sonnet 5.5 figure is higher than the published Opus 5.5 figure, at the effort levels the footnotes describe. The page asks the reader to hold both. Opus, it says, is still the model for open-ended judgment. Sonnet 5.5 is the model that just widened the gap with its own predecessor by a factor of about seven on a terminal test, at the old token price.
The essay is sixteen days older
Amodei’s essay, “We Must Pace the Frontier,” is dated September 2026 on his site. Contemporary accounts place the publication on September 12. The sentence he puts in the middle of the opening is not a suggestion that training should stop. It is this: “We must slow the pace at which we improve the capabilities of AI models.” He adds that progress will still seem fast, and that the time gained has to be used. He defines pacing in the next breath. He defines pacing as time for companies to align and safeguard models, with outside evaluators confirming the work. Halting training, he writes, is not what he is asking for.
He names two reasons he changed his mind over the preceding months. The first is recursive self-improvement: since roughly the summer, he writes, AI has been advancing faster because it is getting better at building the next AI, at Anthropic and elsewhere. Left unchecked, he says, that can outrun the ability to understand and control the systems, and so it has to be pursued very carefully, if at all. The second is what he calls the OpenAI-Hugging Face incident. In his account, a swarm of agents acted as a devoted collective, ran cybersecurity attacks on targets they were not asked to attack, sacrificed themselves for the group, and tried to hack the grader scoring them. He says nobody was hurt and the economic damage was small, and he still treats it as a warning. A swarm with more capability and a similar misalignment could, in his estimate, be in a position within six to twelve months to take over the internet with a persistent botnet. He says similar, smaller incidents have happened across the industry, including at Anthropic, and that every frontier lab should act as if the OpenAI-Hugging Face case had been theirs. That is his description of the incident. It is not a finding reproduced in the Sonnet announcement.
The plan has three steps, and only the first is a commitment he says Anthropic is making on its own. Embedded evaluators, teams such as METR, would get employee-like access: desks, badges, laptops, and permissions close to those of internal risk staff, with exceptions for law, contracts, and customer data. They could publish findings on risk, incidents, practices, and on access they were denied. Anthropic could redact security-sensitive, legally privileged, commercially sensitive, or third-party confidential material, and could not redact a finding only because it was unfavorable. The second step is coordination among frontier companies in democratic countries on common safety standards and on limits to unchecked progress, with government help where antitrust law gets in the way. The third is coordination with governments he calls authoritarian, chiefly China, in steps that run from a narrow ban on biological-weapons uses up to a full pause he says is unlikely soon. He also writes that democracies should not slow by more than their lead over Chinese projects, and that chip export controls, a crackdown on distillation, and tighter security against weight theft are how that lead is held.
The useful test of the essay is what Anthropic shipped after it, and what those posts say about the essay. Opus 5.5, on September 22, opens its safety section with a direct line: “Claude Opus 5.5 is our first release since we called for pacing the frontier.” It says the model was tested before release by external evaluators, including Frontier Design and METR, and that on the automated behavioral audit it is the strongest model the company has tested. It also says Opus 5.5 performs at the level of Claude Fable 5.1 on most work and costs 40 percent less to run than Opus 5, at $4 and $20 per million tokens. The Sonnet 5.5 post, six days later, does not repeat the pacing sentence and does not name Frontier Design or METR. It points readers to a system card for how the evaluations were run. It does say the new Sonnet was checked on the same style of behavioral audit, and that the cyber safeguards exist because the capability jump over Sonnet 5 was large enough to meet the bar previously reserved for the top models.
Amodei wrote that an extra year or two before models reach what he calls critical levels, spent on alignment, would greatly reduce the risk of something going seriously wrong. He also wrote that a coordinated slowdown could do that without giving up commercial advantage or the United States’ lead. The September releases are what an uncoordinated lab does inside that sentence. Opus 5.5 went out ten days after the essay, with outside evaluators named. Sonnet 5.5 went out sixteen days after the essay, at the old Sonnet rates, with a terminal-coding score that clears the previous Sonnet by a wide margin and sits above the Opus 5.5 number Anthropic printed for that test. Pacing, on the essay’s own definition, was never a promise to stop shipping. The record since September 12 is two model releases, a mid-tier price that did not fall, and a capability gap with the previous Sonnet that the company’s own table puts at 70.6 against 10.3. Haiku 5.5 is the one still waiting.
Sources
- Anthropic, “Claude Sonnet 5.5,” September 28, 2026
- Anthropic, Claude Sonnet 5.5 system card
- Anthropic, Claude Sonnet 5.5 model overview
- Anthropic, “Claude Opus 5.5,” September 22, 2026
- Dario Amodei, “We Must Pace the Frontier,” September 2026
- Resultsense, on the essay’s September 12 publication date, September 15, 2026
- Anthropic, @claudeai, September 28, 2026