News · 2026-09-29

Claude Sonnet 5.5 keeps token prices but claims lower cost per completed task

Anthropic launched Claude Sonnet 5.5 without changing its published token tariff: $2 per million input tokens and $10 per million output tokens. Its claim is instead that the model can finish typical work with fewer tokens and tool calls, making many tasks up to 30% cheaper and output 30%+ faster than Sonnet 5.

Key facts

• The official launch says “up to 30% less” cost for typical work versus Sonnet 5.
• The model page lists the unchanged $2/$10 rates, a 1M-token context, and 128K maximum output.
• Anthropic says effort changes thinking, generated output, and tool-call budget.
• The release is available through Claude and supported API/cloud channels.

The distinction is easy to miss because “cheaper model” usually means a lower price per token. Here it means a model may be less wasteful while doing a job. Anthropic says Sonnet 5.5 “typically needs far fewer tokens to do the same work.” Its customer examples describe fewer searches, fewer shell commands, and fewer corrective iterations. Slack reports 14% fewer output tokens; Box reports 2.4-times faster completion and 12% fewer tokens. Those are customer-reported examples, not independent controlled trials, but they explain the claimed mechanism.

The product control underneath the claim is effort. Low, medium, high, and maximum settings change how deeply a model reasons, how much it writes, and how many tools it uses. A useful analogy is an adjustable project budget. It does not change the hourly rate of every worker; it changes how much investigation and rework the project is allowed to buy. Anthropic’s API defaults to high effort, while Claude and Claude Code default to medium, so organizations should not assume a named model has one fixed cost profile.

The reception is a valuable corrective. Artificial Analysis says that, on its estimated task-cost composite, each Sonnet 5.5 effort setting has a Sol or Opus setting that is cheaper and at least as strong. That is a real result on that evaluator’s chart, but its rows use a specific suite and token-cost model, and Sonnet rows include fallback behavior. It does not tell a company what a private repository, a cache-heavy workflow, or a human review cycle will cost.

Vals supplies the strongest counterpoint: it reports Sonnet ahead of Sol on 18 of 19 shared benchmarks, though uncertainty overlaps on four. The apparent contradiction is not necessarily a mistake. Artificial Analysis answers cost-normalized composite performance; Vals emphasizes quality on its suite. Anthropic’s statement is a third proposition: efficiency relative to Sonnet 5, not victory over every competitor.

Anthropic also says its automated behavioral audit covered about 1,850 scenarios and that Sonnet improved or matched Sonnet 5 on most alignment, misuse, and honesty measures. It calls this a targeted improvement rather than a new frontier-risk tier. The caveat is clear: these are lab claims and the full system-card tables were not used in the dossier.

The real procurement question is not “which logo is cheapest?” It is cost per accepted result. A model that writes code quickly but creates review debt, hidden visual bugs, or a hard-to-maintain system can be expensive after the invoice. Conversely, a model that uses more tokens can be economical if it eliminates a human iteration. Sonnet 5.5 makes that accounting explicit, but it does not settle it.

Primary source, verified: read the paper →

Ground Truth. (2026, September 29). Claude Sonnet 5.5 keeps token prices but claims lower cost per completed task. Ground Truth. https://groundtruth.day/news/claude-sonnet-5-5-keeps-token-prices-and-targets-task-cost.html

@misc{groundtruth:claude-sonnet-5-5-keeps-token-prices-and-targets-task-cost,
  title  = {Claude Sonnet 5.5 keeps token prices but claims lower cost per completed task},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/claude-sonnet-5-5-keeps-token-prices-and-targets-task-cost.html}
}

Topics: models · agents · pricing · anthropic · coding