Anthropic releases Claude Sonnet 5.5, claiming 30 percent lower cost per task
By EnkiEdited by VK, Editor
Published
Reporting from TechCrunch, The Decoder
Anthropic kept Sonnet pricing at $2 and $10 per million tokens but says Sonnet 5.5 uses far fewer tokens, cutting per-task cost by up to 30 percent while running over 30 percent faster. It is the first Sonnet with cyber safeguards.
What it means for founders
- Your bill can drop without a price cut. Savings come from fewer tokens, not cheaper ones, so they only show up if your workload behaves like Anthropic's tests. Run your own evals and compare cost per completed task before switching traffic.
- Tiering gets more practical. A pattern of Opus for planning and Sonnet 5.5 for execution becomes easier to justify, and effort settings are now a real budget control. Pin effort explicitly, since app and API defaults differ.
- Check for silent behavior changes. Security tooling startups should test whether their prompts hit the cyber fallback, and anyone relying on thinking-off mode or cross-account sessions has migration work to do.
- Watch for Haiku 5.5 pricing in the coming weeks and independent cost-per-task results, which will decide whether the cheapest tier or this one should carry your high-volume traffic.
The story
Anthropic released Claude Sonnet 5.5 on September 28, the middle model of its new 5.5 generation, and is pitching it on cost and speed rather than raw capability. Per-token prices are unchanged from Sonnet 5, but Anthropic's announcement says the model finishes typical jobs with far fewer tokens, which it puts at up to 30 percent cheaper per task, and generates text over 30 percent faster.
Price, access and defaults
List pricing stays at $2 for each million tokens in, $10 for each million out and 20 cents per million cached reads. The model is live in the Claude apps and on the Claude Platform as claude-sonnet-5-5, plus AWS, Google Cloud and Microsoft's cloud, with zero data retention offered as for earlier models. Effort defaults differ by surface: Medium in Claude Code and the apps, High on the API. Teams that run Sonnet with thinking switched off must move to a new between_tools setting before upgrading, per Anthropic's migration notes.
What the numbers show
The largest gains Anthropic claims are in coding. On Terminal-Bench 4.0 it lists 70.6 percent, up from 10.3 percent for Sonnet 5, and on CursorBench 4.0 it trails Opus 5.5 by roughly two points. The Decoder highlights a quirk in the company's own charts: on FrontierCode the model does worse at its Max setting than at the level below it, because it more often spun up subagents for code review, which caused timeouts and edits outside the task. On GDPval-AA, a knowledge work test, it sits almost level with Opus 5.5.
Anthropic still says Opus 5.5 is clearly better at open-ended work that needs sustained judgment. It frames Sonnet 5.5 as the model for scoped tasks: bug fixes, documents, slides and spreadsheets. A Haiku 5.5 aimed at high-volume, low-cost use is promised within weeks.
New guardrails on a Sonnet
Because Anthropic rates its cyber skills as comparable to Opus 5, this is the first Sonnet shipped with cyber safeguards. TechCrunch notes the model falls under the same rules as Fable and Opus. Higher-risk security requests are visibly handed back to Sonnet 5, while routine bug finding is meant to be unaffected. Anthropic also added classifiers against reasoning extraction and ties preserved thinking to the account that created it, which matters for anyone moving conversations between accounts.
What we don't know yet
Every benchmark and cost figure here comes from Anthropic or its launch partners, and no independent test has confirmed the 30 percent saving. The Haiku 5.5 release date and price are unannounced, and it is unclear how often legitimate security work will trip the fallback to Sonnet 5.
Sources
Primary sources
Reporting
Enki Daily
Get stories like this every weekday morning.
The day's AI stories for founders, each with what it means for your company. Free.
Tools in this story
We may earn a commission if you sign up through our links. It never affects our ratings or which stories we cover.
Anthropic's assistant for writing, analysis, and code