On 28 September 2026 Anthropic published Claude Sonnet 5.5 as a faster, lower-cost complement to Claude Opus 5.5. Where Opus targets complex work that needs sustained judgment, Sonnet 5.5 is positioned for well-scoped everyday tasks: fixing bugs, polishing documents, slides and spreadsheets, with an eye for design that early testers praised. Haiku 5.5, aimed at high volume, will join the family in the coming weeks per the company.
Per-token pricing stays the same as Sonnet 5—$2 per million input, $10 output and $0.20 for cache reads—but Anthropic says it typically needs far fewer tokens for the same work, translating to up to 30% lower cost per task. It also generates outputs more than 30% faster, making it the fastest Sonnet to date.
On evaluations it publishes, Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0 versus Sonnet 5’s 10.3%; on GDPval-AA it sits two points below Opus 5.5; on CursorBench 4.0 it reaches 55.5% versus the predecessor’s 34.1%. It also stands out on long-horizon work and image understanding: Anthropic says it is the first Sonnet to beat Pokémon Red working only from screenshots.
Customers quoted in the announcement—Epic Games, Slack, Zendesk, Unity, Atlassian, SpaceXAI among others—stress speed, fewer tool calls and near-Opus quality on scoped flows. Epic cites system-design audits and gameplay architecture across tens of thousands of lines; Slack reports better offline evals with ~14% fewer output tokens; Zendesk cites tickets processed 20% faster in internal tests.
Anthropic insists benchmark scores do not replace judgment: in its own testing and that of external testers, Opus 5.5 remains clearly stronger at complex, open-ended work. The implicit product recommendation is a fleet: Opus for architecture and judgment, Sonnet 5.5 for fast iteration and volume, Haiku soon for extreme cost.
On alignment, the automated behavioural audit (~1,850 scenarios) shows Sonnet 5.5 improves on or matches Sonnet 5 on most measures. Because its cybersecurity capabilities are comparable to Opus 5’s, it is the first Sonnet to launch with cyber safeguards and fallbacks like those on the most capable models; biology safeguards match Sonnet 5’s.
The security point most aligned with the 1 October news cycle is distillation. Anthropic says Sonnet 5.5 is the first Sonnet with classifiers that prevent reasoning extraction and with expanded “preserved thinking” so Claude’s thoughts cannot be decoupled from the account that created them. It arrives hours before OpenAI details its own adversarial distillation campaign.
Availability: Claude Platform under id `claude-sonnet-5-5`, plus Amazon Web Services, Google Cloud and Microsoft Azure, with zero data retention like Opus 5.5 and Sonnet 5. Teams that ran Sonnet with thinking off must migrate to the `between_tools` setting. Default effort is Medium in apps and Claude Code, High on the platform.
For product teams, the draw is the cost-quality curve: on several Anthropic charts, Sonnet 5.5 at Low or Medium effort beats Sonnet 5’s best score at a fraction of the cost per task. That pushes reallocation of “always Opus” workloads toward difficulty-based routing, with a direct hit on agent budgets.
Timing relative to Anthropic’s potential IPO—covered days earlier in the 29 September batch—adds a financial reading: a Sonnet that compresses cost and lifts agentic coding strengthens the enterprise-revenue narrative just as the prospectus discusses existential risks. They are pieces of the same board, not separate stories.
Residual risks: Anthropic acknowledges that no evaluation suite catches every failure, and that Sonnet 5.5 may have tendencies not yet found. Cyber and Life Sciences verification programmes offer tiered access to dual-use capabilities; without them, certain high-risk tasks visibly fall back to Sonnet 5.
Neural Pulse sees Sonnet 5.5 as the classic “middle model” move that shifts real production volume. If tokens-per-task figures hold in customer deployments, migration from Sonnet 5 will be swift; if not, it will remain a polished release note with generous benchmarks. The market will measure it on October invoices.
