A faster everyday AI model arrived at its predecessor's token price.

On September 28, Anthropic released Claude Sonnet 5.5. Anthropic describes it as a faster, lower-cost model. Small nonprofits would most likely try it first on drafts, documents, slides and spreadsheets. The per-token price is the same as its predecessor. The company says it uses fewer tokens per task.

For a nonprofit with a thin budget, that sounds like good news, and it may be. But for most mission teams, the model’s bill was never the expensive part. This post covers what changed and how to find out what a first project really costs you.

What changed

The facts below come from Anthropic’s announcement. The performance and cost claims are company-reported.

Who it is for. Anthropic calls Sonnet 5.5 “a faster, lower-cost complement to Claude Opus 5.5.” It says the model is strongest at “well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets.” Opus 5.5 remains the model for complex work that needs careful judgment.
Price. $2 per million input tokens and $10 per million output tokens, the same as Sonnet 5. Cache reads are $0.20 per million and cache writes $2.50 per million.
Speed and cost per task. Anthropic says it generates output 30%+ faster than Sonnet 5 and that, in its testing, it costs up to 30% less per task because it needs fewer tokens. “Up to” means that is the best case, not the typical one.
Safeguards. It is the first Sonnet model to launch with cyber safeguards and fallbacks like those Anthropic has developed for its most capable models.
Data. Anthropic says the model is “available with zero data retention,” as with Opus 5.5 and Sonnet 5.

The small worked example

Here is the arithmetic for one task, using the list prices above. The token counts are made up to show the method. They are not measurements.

Say a program manager asks the model to draft a two-page quarterly update for a funder. The prompt includes the grant’s reporting template and last quarter’s notes, about 20,000 tokens in. The draft that comes back is about 4,000 tokens.

Input: 20,000 tokens × $2 per million = $0.04
Output: 4,000 tokens × $10 per million = $0.04
Total: about $0.08 per draft

If the best case of a 30% lower cost per task held, that would drop to roughly $0.056. You would save about two and a half cents per report.

Now count the other cost. Say the program manager spends 40 minutes checking the draft against the grant agreement, fixing numbers and rewriting the parts that sound nothing like your organization. At any staff rate, those 40 minutes cost more than a year of these drafts in tokens.

That leads to the decision rule for this post.

The decision rule

Judge a first AI project by review minutes per finished piece, not by the price per token.

A model that is a little more expensive per task but gives you drafts your staff can approve in 10 minutes is the cheaper choice. A free tool that makes someone rewrite half the page is not cheap. The saving from a model like Sonnet 5.5 matters only if its drafts also need less fixing. You can find that out with your own work in a week.

A test you can run this week

Pick one recurring piece of work that is well defined and low risk. A monthly volunteer newsletter, a board meeting summary or a spreadsheet cleanup all work. Avoid anything containing donor records, client or case information, or personnel files until your data terms are settled.

Write the brief once. Say what the piece is for, who reads it, what it must include and what it must never say.
Run it on each model you are considering. Use the same brief and the same inputs.
Have the usual reviewer check each draft. Time them. Note what they had to fix: facts, tone, structure or rules.
Record three numbers per draft: review minutes, number of fixes, and whether it was approved without being rewritten.
Choose the setup with the fewest review minutes. If there is a tie, choose the cheaper one.

A person still approves every piece before it reaches a funder, a donor or the public. That does not change with a faster model.

Before anything sensitive goes in

This is general information, not legal or privacy advice. A lower cost per task does not make a tool safe for donor or client data. Before any sensitive workflow, find out in writing what the provider keeps, for how long, and whether zero data retention applies to your account. Then check that against your privacy policy and any funder data terms.

What to do this week

Choose one low-risk process that recurs every month.
Write a one-paragraph brief for it.
Run the timed review test above and record review minutes, fixes and approvals.
Ask your AI vendor, in writing, what data retention terms apply to your account.
Keep donor and client records out of the tools until you have that answer.