Guides
Claude Sonnet 5.5 costs up to 30% less. Who sees it?
Harriet Team · · 5 min read
On September 28, 2026, Anthropic released Claude Sonnet 5.5 and said it costs up to 30% less per task than Sonnet 5. The price per token did not change. The saving comes from the model using fewer tokens to do the same work, so it only reaches a bill that is counted in tokens, and only for the people who are running on Sonnet.
What did Anthropic change with Claude Sonnet 5.5?
The figures below are from Anthropic’s announcement of September 28, 2026 and its pricing page as checked on September 30, 2026. Check the source before you budget on them.
- Sonnet 5.5 is $2 per million input tokens and $10 per million output tokens. The announcement says it is “priced the same as Sonnet 5”.
- Cache reads are $0.20 per million tokens and cache writes are $2.50, the same as Sonnet 5 on the pricing page.
- Anthropic says the model typically needs far fewer tokens for the same work, and that in its own testing a task costs up to 30% less than on Sonnet 5.
- It is available on Anthropic’s platform, AWS, Google Cloud and Microsoft Azure, with zero data retention available.
Thirty percent is a ceiling that Anthropic measured on its own tasks. The price list did not change, and your number depends on your work.
Does a Claude Team bill go down?
No. Anthropic’s pricing page on September 30, 2026 lists a standard Team seat at $20 a month on annual billing and $25 on monthly billing, and a premium seat at $100 and $125.
The announcement does not mention usage limits on any plan. Anthropic’s help center says usage depends on which model you use and the effort level you select, so a model that needs fewer tokens should stretch a seat further. How much further is not published.
A company that pays Anthropic by the token sees the saving directly. That is how Harriet bills when you bring your own Anthropic key. You pay Anthropic’s token prices with no model charge from Harriet, so fewer tokens per task is a smaller invoice.
Does a Claude Enterprise bill go down?
It can. Anthropic’s pricing page describes Enterprise as a seat price plus usage at API rates, with the seat at $20 a month billed annually. Fewer tokens per task means a smaller usage line for any work that runs on Sonnet 5.5.
Two conditions decide whether you see it.
The first is which model your people are on. Anthropic’s Claude Code documentation, checked on September 30, 2026, says the default model for Team and Enterprise accounts is Opus 5.5. Opus 5.5 is $4 per million input tokens and $20 per million output tokens on the pricing page, twice the Sonnet rate. Anyone left on that default never touches the Sonnet saving.
The second is who controls the default. Anthropic’s help center says an organization default model is available on the Enterprise plan, and that it sets the starting model for new conversations in chat, Cowork and Claude Code. Users can still pick a different model for any conversation. The same page does not offer the setting on Team. On Harriet the default model is a setting on each team’s profile, and a change applies to everyone in that team on their next request.
Should the default be Sonnet 5.5 or Opus 5.5?
Anthropic’s own guidance points at Sonnet for everyday work. Its Enterprise consumption guide calls Sonnet the daily driver and Opus the model for harder, more complex work.
The announcement reports Sonnet 5.5 scoring nearly level with Opus 5.5 on GDPval-AA, a benchmark of real-world tasks across 44 occupations. At half the input and output price, that is a strong case for well-defined, repeatable tasks.
Cache reads cost $0.20 per million tokens on both models, so on long sessions that reuse a lot of context, Opus 5.5 is closer to Sonnet 5.5 than the headline prices suggest.
What should a company do this week?
- Find out which model your people run on. If nobody has set a default, heavy users of Claude Code are on Opus 5.5.
- On Enterprise, set the organization default for everyday work and keep Opus for the teams that need it. Anthropic’s consumption guide describes spend limits at organization, group and user level. Use them.
- Test before you switch. Take ten real tasks from your own teams, run them on both models, and compare the output and the token count. Anthropic’s 30% is where that test starts. Harriet’s eval mode runs that comparison for you, with output, tokens, cost and time side by side.
As with last week’s Opus 5.5 release, a cheaper model only lowers a bill built on usage, and only when someone decides which work goes to which model.
Harriet’s router carries Claude, GPT, Gemini and open models behind one policy and adds new models as they ship. An admin sets the default model and a budget for each team.
For what routing and caps do to an AI bill, start with cost control. OpenAI shipped a cheaper everyday model a day later at the same list price. See GPT-6.1 Sol. If you are on Team and weighing the move to usage-based pricing, read Claude Team vs Claude Enterprise. For the seat and usage math on Enterprise, see Claude Enterprise pricing.
