Anthropic has released Claude Sonnet 5.5, the second model in its Claude 5.5 family. It generates output more than 30 percent faster, costs up to 30 percent less per task, and nearly matches Opus 5.5 on knowledge-work benchmarks. On Terminal-Bench, a coding benchmark, the model jumps from 10.3 to 70.6 percent.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Anthropic releases Sonnet 5.5, the second model in the Claude 5.5 family. The company says it approaches the flagship Opus 5.5 on coding and knowledge-work benchmarks, cuts cost per task by up to 30%, raises output speed by more than 30%, and adds cybersecurity and anti-distillation safeguards to a Sonnet-tier model for the first time. All these figures come from Anthropic's own statements or from the report's paraphrase, and still need independent testing to confirm.
Anthropic has released Claude Sonnet 5.5, the second model in the Claude 5.5 family, and announced that Claude Haiku 5.5 will arrive in the coming weeks, aimed at high-throughput, low-cost scenarios. According to the positioning relayed by the report, Opus 5.5 targets complex tasks that require careful judgment, while Sonnet 5.5 targets well-defined everyday work such as fixing bugs, writing documentation, and building presentations and spreadsheets.
This positioning means the release narrative centers not on "strongest" but on "good enough and cheap." The report says the performance differences among the models being compared at each tier are small enough that cost could become the deciding factor, but this judgment offers no quantitative basis and can only be treated as the report's speculation.
The question readers really need to judge for themselves is this: when the official benchmark numbers push a mid-tier model up next to the flagship, how much do those numbers translate into real-world gaps on everyday tasks? The report itself notes that all performance and cost figures come from Anthropic's own statements or from the report's paraphrase, and still need independent testing to confirm.
Coding is where Sonnet 5.5 differs most visibly from its predecessor. The report relays that on the agentic coding test Terminal-Bench 4.0, Sonnet 5.5 scores 70.6%, while Sonnet 5 scores 10.3%; on CursorBench 4.0, Sonnet 5.5 scores 55.5% and Sonnet 5 scores 34.1%, only about two percentage points below Opus 5.5's 57.8%.
Anthropic also says that under the High setting of FrontierCode 1.1, Sonnet 5.5 scores ten points above Sonnet 5, at roughly one-fifteenth the cost per task. The report mentions that early testers praised the model's speed at understanding codebases, and says Sonnet 5.5 calls tools in batches more often than its predecessor, reducing the number of steps required.
But there is an anomaly in the effort levels: at the highest effort level, Max, Sonnet 5.5's FrontierCode score is actually lower than at the Xhigh level. Anthropic's explanation is that at maximum effort the model more frequently triggers its code review feature and splits work across multiple sub-agents, leading to timeouts or changes that exceed the task scope, and both of these are penalized by FrontierCode.
This constitutes a clear constraint: the highest effort setting is not always optimal. The report also notes that finding the right effort level for each task is still more art than science, with no clear method.
The comparison basis also deserves attention: Opus 5.5's Terminal-Bench 4.0 result is taken from its highest score at the Xhigh level, so the comparison basis with Sonnet 5.5 is not entirely consistent.
On GDPval-AA, the knowledge-work benchmark developed by OpenAI covering 44 occupations and 9 industries, the report relays that Sonnet 5.5 scores 1,844, close to Opus 5.5's 1,846 and about 400 points above Sonnet 5's 1,449. On the chart-recognition test Chartography, Sonnet 5.5 jumps from 15.6% to 61.6%.
These numbers come with a caveat that must travel with them: the report notes that Sonnet 5.5's GDPval-AA and AA Briefcase figures come from a pre-release version, and a bug that has since been fixed may have affected structured output. In other words, these two results do not reflect the confirmed performance of the final shipped version.
There is another scope limitation: Sonnet 5.5's OSWorld 2.1 result is a partial evaluation. The report also says early testers described Sonnet 5.5 as a more natural conversational partner with a feel for design, while Anthropic claims the model can redo user interfaces and execute slide templates with results that need almost no further polishing — these are vendor claims.
Anthropic also says Sonnet 5.5 is the first Sonnet model to beat Pokémon Red using screenshots alone. The report places this alongside similar progress on game benchmarks as an illustration of mid-tier model capabilities moving upward.
To understand the cost narrative of this release, the key is to distinguish two levels. Sonnet 5.5's per-million-token pricing is the same as Sonnet 5: $2 input, $10 output, $0.20 cache read. The unit price has not changed.
Anthropic says that because the model uses fewer tokens per task, actual cost per task drops by up to 30%; output generation speed also rises by more than 30%. The report states explicitly that these claims still need independent testing to confirm.
The report also says that at low or medium effort settings, Sonnet 5.5 already beats Sonnet 5's best result at roughly one-tenth the cost per task, but this is Anthropic's unilateral claim. Combined with the phenomenon that the Max level is actually worse, the relationship between effort level, cost and quality is not linear.
So "up to 30% lower" is a ceiling statement, not a universal value; it depends on the specific task, effort level and token usage, and has not yet been verified by third parties.
Because Sonnet 5.5's cybersecurity capabilities far exceed its predecessor's, Anthropic adds safeguards to a Sonnet model for the first time: high-risk cybersecurity requests are clearly rerouted to Sonnet 5, and qualified professionals can apply for tiered access through an expanded Cyber Verification Program.
Anthropic also adds anti-distillation attack safety classifiers of the same kind used on its strongest models to Sonnet 5.5. The report says whether these measures are truly effective will only become clear over the coming months.
The report further speculates that if the anti-distillation measures do close off the relevant attack paths, they could widen the gap between Western and Chinese labs again. This is the report's speculation, not a confirmed result.
Placing the security provisions alongside the earlier capability narrative, the logic is self-consistent: a mid-tier model that leaps forward in coding and cybersecurity simultaneously amplifies misuse risk, so the vendor chooses to set up rerouting and tiered access while keeping it openly available.
Sonnet 5.5 is already live on AWS, Google Cloud and Microsoft Azure, callable through the Claude Platform under the model ID "claude-sonnet-5-5", and offers zero data retention. This is the most certain part of the release.
The list of items that need to be flagged as unverified is equally clear: the speed increase of more than 30% and the cost-per-task reduction of up to 30% have not yet been verified by independent testing; whether the anti-distillation safety classifiers and cybersecurity safeguards are truly effective, the report says, will only become clear over the coming months; Sonnet 5.5's GDPval-AA and AA Briefcase figures come from a pre-release version; OSWorld 2.1 is a partial evaluation; and Opus 5.5's Terminal-Bench result is taken from its highest score at the Xhigh level, so the comparison basis is not entirely consistent.
Stacking these limitations together, the safest reading of this release is: Anthropic has given self-reported figures showing a large leap in coding and knowledge work, and has simultaneously lowered cost-per-task expectations, but "close to Opus 5.5" remains for now a conclusion under the vendor's and a single report's framing.
All facts in this article come from a single source: THE DECODER's report on Anthropic's release (article_id 19048). Among these, the model release, the Haiku 5.5 timeline, token unit pricing, cloud platform availability and model ID, and the GDPval-AA pre-release note are confirmed facts; the benchmark scores, the speed and cost reductions, the cybersecurity rerouting and anti-distillation classifiers, and the Pokémon completion are all Anthropic's own statements or the report's paraphrase, constituting source claims that have not been confirmed by independent testing. The report's judgments that cost could become the deciding factor and that anti-distillation measures could widen the lab gap are the report's speculation, and this article has labeled them as speculation.
Sonnet 5.5's core selling point is not that it is stronger, but that it pushes near-flagship coding and knowledge-work performance down to a lower cost per task; however, the price has not changed, the cost reduction comes from token usage, and all the key numbers remain at the stage of Anthropic's own statements and a single report's paraphrase, so they should not be treated as settled before independent verification.
Primary report
Primary source