A second launch in the same family, six days apart
Anthropic put out Claude Sonnet 5.5 on September 28, 2026, less than a week after it shipped Claude Opus 5.5 on September 22. That cadence is itself notable: Anthropic is now iterating fast enough that its mid-tier model and its top-tier model launched inside the same month, aimed at different jobs rather than different customers.
Sonnet 5.5 is pitched as the everyday workhorse, not the flagship. Anthropic describes it as suited to well-scoped tasks such as bug fixes, spreadsheets, presentations and interface design, while reserving Opus 5.5 for work that needs deeper judgment. The pricing reflects that split: Sonnet 5.5 costs the same as its predecessor, Sonnet 5, at $2 per million input tokens and $10 per million output tokens, which is half of what Opus 5.5 costs.
The pitch is speed and fewer wasted tokens, not a price cut
Anthropic is not lowering its per-token rate for Sonnet 5.5. Instead, the company's argument is that the model does more with the same rate: it generates output more than 30 percent faster than Sonnet 5, and Anthropic says that speed, combined with needing fewer tool calls and iterations to finish a task, adds up to roughly 30 percent lower cost per completed task in its own testing. Run at low or medium reasoning effort, Anthropic claims Sonnet 5.5 can match or beat Sonnet 5's best score on internal benchmarks at around a tenth of the cost per task.
That is a claim about efficiency, not sticker price, and it is worth keeping the two separate. A model that finishes a coding task in three iterations instead of eight will cost less even without a rate change, but the saving only materializes if the task genuinely needs fewer iterations, which will vary by workload.
What the benchmarks show
On Anthropic's GDPval-AA evaluation, which measures performance on real-world occupational tasks, Sonnet 5.5 scored 1844 against Opus 5.5's 1846 and the older Sonnet 5's 1449, putting the new mid-tier model almost level with the flagship it is meant to sit beneath. On OSWorld 2.1, a computer-use benchmark, Sonnet 5.5 reached 80.1 percent versus Opus 5.5's 81.8 percent and Sonnet 5's 57 percent. On Terminal-Bench 4.0, an agentic coding test, Sonnet 5.5 scored 70.6 percent against Sonnet 5's 10.3 percent, the largest single jump of the group. It scored 55.5 percent on CursorBench 4.0, a few points behind Opus 5.5's 57.8 percent.
Anthropic also says Sonnet 5.5 is the first model in the Sonnet line built with dedicated cybersecurity safeguards, a step up from what Anthropic had previously reserved for its top-tier models, alongside invisible watermarking meant to help distinguish AI-generated text.
What early users are reporting
Anthropic published a handful of customer figures alongside the launch. Zendesk said its support ticket processing got 20 percent faster; Zendesk's Abhinay Kathuria said "tickets were processed 20% faster, getting our customers the help they need without the wait." Collaboration software maker Box reported tasks completing 2.4 times faster using 12 percent fewer tokens. Slack said it saw a 14 percent reduction in output tokens for comparable work. App-building startups Lovable and Base44 both reported needing fewer iterations to reach a finished result, with Base44 citing 3.6 iterations on average versus 7.7 for a prior model.
These are vendor-supplied figures from companies with a commercial relationship with Anthropic, not independent benchmarks, and should be read with that caveat.
Availability
Sonnet 5.5 is live now under the model identifier claude-sonnet-5-5, accessible directly through Anthropic and through Amazon Web Services, Google Cloud and Microsoft Azure, with a zero-data-retention option for customers who need it. Anthropic has said a smaller Haiku 5.5 model is planned for the coming weeks, which would complete the refresh of its entire model lineup within a single month.
What it means in practice
For teams already building on Sonnet 5, the upgrade path looks low-risk: same price, and Anthropic's own numbers suggest fewer wasted tool calls on agentic and coding work specifically, which is where the biggest benchmark gains showed up. For teams currently paying Opus prices for tasks that do not need Opus-level judgment, Sonnet 5.5's near-parity scores on GDPval-AA and OSWorld are the number worth testing against a real workload before committing, since benchmark parity does not always hold on a specific, idiosyncratic task. Anyone evaluating the "10x cheaper per task" claim should measure it on their own workload rather than take the multiplier at face value, since it depends on how many iterations a task took before and how many it takes now.
