Same price, fewer tokens

$2 per million input tokens and $10 per million output tokens. That’s the list price of Claude Sonnet 5.5, which Anthropic released on September 28. It’s unchanged from Sonnet 5 and half of Opus 5.5, which launched a week earlier at $4 and $20. Cache reads are $0.20 per million on both.

Anthropic says Sonnet 5.5 needs far fewer tokens to do the same work, costs up to 30% less per task than Sonnet 5, and generates output more than 30% faster.

My view: for most recurring operating work, the workhorse model should now be the default and the flagship the exception. I mean coding fixes, document and spreadsheet loops, and ticket triage. The model race that matters to operators is about unit economics, not demos. It’s the same reason enterprise AI harnesses beat model wars.

What Anthropic says it’s for

Anthropic pitches Sonnet 5.5 for “well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets.” It reserves Opus 5.5 for “complex work requiring careful judgment.” It’s available on Anthropic’s Claude Platform and through AWS, Google Cloud, and Microsoft Azure. Haiku 5.5, built for high-volume, cost-sensitive work, is due in the coming weeks.

In Anthropic’s own benchmarks, Sonnet 5.5 comes close to Opus 5.5 on scoped work. It scores 55.5% vs. 57.8% on CursorBench 4.0, which uses coding tasks from real Cursor sessions. On GDPval-AA, a test of real-world tasks across 44 occupations, it scores 1844 vs. 1846. These are vendor-reported numbers. Anthropic also says Opus 5.5 “remains clearly stronger at complex, open-ended work requiring sustained judgment.” That caveat is the useful part: Sonnet is close on scoped work and behind on judgment.

Customer quotes point the same way. Zendesk says tickets were processed 20% faster in its testing. Slack says Sonnet 5.5 beat Sonnet 5 on almost all of its offline Slackbot evals while using about 14% fewer output tokens. These are early-tester claims, not audits. But they measure the right things for operators: cost and speed on repeat work.

The IPO angle: Anthropic sells usage

Reuters says Anthropic is expanding its lineup “ahead of a planned IPO.” In separate exclusives based on a copy of the confidential prospectus, Reuters reports that revenue grew about 12-fold in 2025 to nearly $4.6B. About $3.8B of that came from customers paying by usage, vs. $789M from subscriptions. Anthropic expects usage-based revenue to stay “the substantial majority” for the foreseeable future. Anthropic also told Reuters that enterprise customers are about 80% of its business.

Read that as a CFO. Anthropic’s revenue is your token bill. I read a model that costs less per task at the same list price as a bet that you’ll run more jobs, not that you’ll spend less. Both can happen at once: your cost per job falls while your total bill rises because you automated more. Budget for that on purpose.

Two stacks of blank receipts beside a messy pile of red-marked rework sheets
Budget the finished job, including rework. A cheaper list price only helps if the shorter stack still clears the work.

How to price “good enough” into a workflow

Stop comparing models by leaderboard. Compare them by cost per finished job, including rework. That’s how to evaluate any AI workflow before you scale it.

Here is an illustration using list prices only, with made-up volumes. Say a month-end job reads 1,000 documents at about 20,000 input tokens and 2,000 output tokens each. That’s 20M input tokens and 2M output tokens. On Sonnet 5.5 it costs $40 + $20 = $60. On Opus 5.5 it costs $80 + $40 = $120. That’s before caching, effort settings, or retries. At that volume the $60 gap is small money. The real cost is a wrong answer that someone has to catch.

So run the test this way:

  1. Pick three recurring jobs. Choose work that happens weekly and has an owner, like invoice exceptions, bug triage, or a Monday reporting pack.
  2. Run both models on the same 20 to 50 real cases. For each one, track whether the owner accepts the output, how many tokens it used, and how many minutes of human cleanup it needed.
  3. Price the job, not the token. Cost per accepted output is the model spend plus review time at a fully loaded labor rate.
  4. Default to the workhorse and escalate by rule. Send a case to Opus only when it hits a named condition, like unclear scope, dollars above a threshold, or a task you’ve never run before.
  5. Set effort on purpose. In Anthropic’s apps and Claude Code the default effort is Medium. On the Claude Platform it’s High. Higher effort usually costs more per task, so know which setting your tools use.

Two operational catches

The model you pick isn’t always the model you get. Anthropic says Sonnet 5.5’s cybersecurity capabilities are a large step up from Sonnet 5’s, so it ships with safeguards: higher-risk cybersecurity requests “visibly fall back to Sonnet 5.” Routine bug fixing isn’t affected. If you do security work, say as an MSP or an IT team doing pen-test prep, test for the fallback before you assume you got the upgrade.

Switching isn’t zero-touch either. If you run Sonnet with thinking off, Anthropic says you need to switch to a new between_tools setting before moving to 5.5. It’s a small change, but it’s the kind that breaks a nightly job at 2 a.m.

FAQ

How much does Claude Sonnet 5.5 cost? $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20 per million. That’s unchanged from Sonnet 5 and half of Opus 5.5’s $4 and $20.

When should I pay for Opus 5.5 instead? Anthropic reserves Opus 5.5 for “complex work requiring careful judgment.” Default to Sonnet and escalate to Opus by a written rule, like unclear scope, dollars above a threshold, or a task you’ve never run before.

Is Sonnet 5.5 a drop-in replacement for Sonnet 5? Mostly, with two catches. If you run Sonnet with thinking off, Anthropic says you need to switch to the new between_tools setting first. Higher-risk cybersecurity requests visibly fall back to Sonnet 5.

Will a cheaper cost per task lower my AI bill? Not necessarily. My read is that your cost per job falls while your total bill rises because you automate more. Budget for that on purpose.

Bottom line

For an operator, the useful AI decision isn’t which lab has the smartest model. It’s which model finishes your recurring work at an acceptable error rate for the lowest total cost, plus a rule for when to pay more. Sonnet 5.5 makes that easy to see.

Make the workhorse the default. Put the flagship behind an escalation rule. Track cost per accepted output every month, like any other unit cost, and bring it to your next AI tool stack audit. If your pilots show the flagship paying for itself on routine work, I’d like to see the numbers.

Sources

Written by

Ken Koo

Partner at Force Scaling

Partner at Force Scaling, helping CEOs and leadership teams turn ambition into operating discipline and results.

ADVERTISEMENT