Anthropic released Claude Sonnet 5.5 on Monday, 28 September 2026, the second model in its Claude 5.5 family — and the numbers suggest the gap between its mid-tier and flagship models has narrowed to almost nothing on the work developers actually do. The company says the new model runs more than 30 per cent faster than its predecessor, costs up to 30 per cent less per task, and in one headline benchmark outscores the flagship Opus 5.5 that launched the week before.
It is Anthropic's second launch in a month defined by debate over AI safety — and the company used the release to position Sonnet 5.5 as the everyday workhorse for coding, documents, slides and spreadsheets, backed by enterprise customers who now account for roughly 80 per cent of its business.
What changed from Sonnet 5
Sonnet 5.5 replaces Sonnet 5, which launched in June, as Anthropic's medium-tier model. API pricing is unchanged at $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20 per million tokens — exactly half the list price of Opus 5.5, which costs $4 per million input and $20 per million output tokens.
The improvements, according to Anthropic, come from efficiency rather than price cuts: the model generates output more than 30 per cent faster and typically needs fewer tokens and fewer tool calls to complete the same work, cutting cost per task by up to 30 per cent. The company describes it as the fastest Sonnet model it has released, and positions it for well-scoped everyday tasks — fixing bugs, producing polished documents, slides and spreadsheets, and refining user interfaces. Early testers, Anthropic says, found it a clearer writer and a better collaborator for quick iteration.
The technical specifications are substantial for a mid-tier release: a 1-million-token context window, up to 128,000 tokens of maximum output, and a knowledge cutoff through June 2026. Adaptive thinking is switched on by default, with effort settings ranging from low to maximum.
The benchmark that turned heads
The headline number from Anthropic's launch materials is on Terminal-Bench 4.0, an evaluation of models' ability to complete multi-step tasks via the command line. Sonnet 5.5 scored 70.6 per cent — up from just 10.3 per cent for Sonnet 5, and ahead of Opus 5.5's own reported 66.4 per cent on the same settings.
That is an extraordinary jump for a mid-tier model: a more than sixfold improvement on the agentic coding benchmark, landing ahead of the flagship sibling on terminal-heavy work. On Anthropic's other launch benchmarks the picture was more balanced — Sonnet 5.5 trailed Opus 5.5 narrowly on most, including CursorBench 4.0 (55.5 per cent versus 57.8 per cent) and Humanity's Last Exam with tools (64.5 per cent versus 67.7 per cent). Anthropic's own caveat was explicit: "Opus 5.5 remains clearly stronger at complex, open-ended work requiring sustained judgment."
Independent testing broadly confirmed the trend. Artificial Analysis ranked Sonnet 5.5 second on its Intelligence Index, just two points behind Opus 5.5 at maximum effort — an 18-point jump over Sonnet 5. But the same independent analysis added a wrinkle: at maximum effort, Sonnet 5.5 used roughly 193,000 output tokens per task, about 60 per cent more than Opus 5.5. Anthropic's claim of lower cost per task holds for typical use, but pushing the model to its limits can erase the advantage — a trade-off enterprise buyers will need to model carefully.
Safety features move downmarket
Notably, Sonnet 5.5 is the first Sonnet release to ship with the cyber safeguards and fallback protections Anthropic previously reserved for its most capable models, and the first with safety classifiers the company says block reasoning-extraction attempts used in distillation attacks — where rivals try to extract a model's capabilities by probing its reasoning.
The timing is significant. Sonnet 5.5 is Anthropic's second launch amid a growing industry debate over AI risks: chief executive Dario Amodei earlier in September called on the global AI community to slow the pace of releasing new capabilities to address safety concerns. Shipping stronger safeguards on a mid-tier model — the tier most widely deployed — suggests the company is treating safety tooling as a product feature that scales down, not just up.
Enterprise is the audience
Reuters reported that enterprise customers now account for about 80 per cent of Anthropic's business, with clients including Salesforce, Databricks, Goldman Sachs and Novo Nordisk. That explains the product's shape: Sonnet 5.5 is not pitched as a frontier research instrument but as a dependable, fast, cost-controlled engine for the work companies do all day — coding, document creation, spreadsheet work.
Distribution matches the pitch. The model is available on the Claude Platform, AWS, Google Cloud and Microsoft Azure with zero data retention, per Anthropic. GitHub added Sonnet 5.5 to Copilot on the same day, generally available to Pro, Pro+, Max, Business and Enterprise users across VS Code, Visual Studio, the Copilot CLI, JetBrains, Xcode and github.com. A third model, Haiku 5.5 — designed for fast, high-volume work — is due soon, completing the family.
The IPO backdrop
Reuters framed the release as part of Anthropic's product expansion ahead of a planned initial public offering. The strategy is legible: a three-tier lineup — Haiku for volume, Sonnet for everyday work, Opus for the hardest problems — each refreshed on a rapid cadence, each with clearer positioning and pricing than the last.
The cadence itself is the story. Opus 5.5 launched one week before Sonnet 5.5; Sonnet 5 had launched only in June. In a market where model generations used to be measured in years, Anthropic is now iterating its flagship line in weeks — even as its own chief executive calls for the industry to slow down. That tension, between commercial velocity and safety caution, is becoming the defining contradiction of the AI industry in late 2026.
What it means for buyers and builders
For developers choosing models for CI pipelines, bug triage or agentic coding tasks, Sonnet 5.5's Terminal-Bench result is the number that matters: near-flagship performance on terminal-heavy workflows at half the per-token price and meaningfully higher speed. For knowledge work the gap is similarly thin — Artificial Analysis scored it within a couple of Elo points of Opus 5.5 on briefcase-style tasks.
The honest reading, though, is narrower than the hype. These are Anthropic's numbers, run by Anthropic, on benchmarks Anthropic selected — and the company itself says complex, open-ended work still belongs to Opus. Independent tests confirm the direction but add the token-usage caveat. The right conclusion is not that Sonnet replaces Opus, but that the price-to-capability gap between Anthropic's two tiers has collapsed for well-scoped work — which, for most enterprises, is most of the work.
For the industry, Sonnet 5.5 is another data point in a remarkable September: OpenAI shelving a flagship launch on safety grounds on the same week Anthropic ships a mid-tier model that beats that flagship's sibling on coding benchmarks. The race is no longer just about who builds the smartest model. It is about who can ship capability fastest without breaking the safeguards — and who can price it so the rest of the economy can actually use it.
Sources
