Anthropic shipped Claude Sonnet 5.5 on September 28, with scores that sit within touching distance of Opus 5.5 across several evaluations and output generation billed as more than 30 percent faster. Token pricing holds steady, and the cost of finishing a task still falls by as much as 30 percent because the model burns through fewer steps.
Key Takeaways
- Claude Sonnet 5.5 scores 70.6 percent on Terminal-Bench 4.0 against 10.3 percent for Sonnet 5, and clears Opus 5.5 on that same test.
- Pricing stays at $2 per million input tokens and $10 per million output tokens: the savings come from fewer steps, not from a cheaper rate card.
- Anthropic keeps Opus 5.5 ahead on long, open-ended work that needs sustained judgment.
Have an AI Sum Up This Article
ChatGPTClaude Sonnet 5.5 Goes From 10 to 70 on the Test That Matters
The headline jump lives in two numbers. On Terminal-Bench 4.0, an agentic coding evaluation, Claude Sonnet 5.5 lands at 70.6 percent where Sonnet 5 stalled at 10.3 percent. Opus 5.5, the flagship that shipped six days earlier, stops at 66.4 percent on the same run.
A mid-tier model clearing the flagship on an agentic coding benchmark is the sort of result that usually takes a full generation to arrive. Here it took six days.
The rest of the scorecard moves the same way. CursorBench 4.0 climbs from 34.1 to 55.5 percent, FrontierCode 1.1 from 42.4 to 52.1 percent at maximum effort, and Chartography, which measures chart reading, jumps from 15.6 to 61.6 percent. On GDPval-AA v2.1, the evaluation built around real work across occupations, the model posts 1,844 points against 1,846 for the flagship.
Two points apart on a scale of nearly two thousand is the whole commercial argument of this release. Anthropic sold its flagship on the ability to work out on its own how much thinking a task deserves, and the cheaper model now arrives with almost the same mark.
The model ships under the identifier claude-sonnet-5-5, with zero data retention, served from AWS, Google Cloud and Microsoft Azure. GitHub Copilot opened it to users the same day.
Fewer Steps, Fewer Tool Calls, a Smaller Monthly Bill
The rate card holds: $2 per million input tokens, $10 per million output tokens, $0.20 for a cache read. The drop in cost per task comes from somewhere else. Anthropic explains that the model reaches the same result while spending fewer steps, fewer tokens and fewer tool calls.
That framing has served the company before. Sonnet 5 landed by undercutting GPT-5.5 on agent work, and Opus 5 matched the top of the market at half the price.
For a team running agents around the clock, the saving shows up on the monthly invoice before it shows up on the price list. A job that took fifteen tool calls and now takes ten costs a third less at an identical rate.
Efficiency of that kind also changes what engineering teams can leave running unattended. Budget ceilings on agent loops exist because token spend grows with every retry, and a model that needs fewer retries raises the ceiling without a single line of policy being rewritten.
The cache line deserves attention on its own. At $0.20 per million tokens read from cache against $2 to send fresh input, a workflow that keeps a stable system prompt and a long reference document in place pays a tenth of the rate on the bulk of what it sends. Agent architectures built around a fixed context benefit twice over, once on the cache and once on the shorter run.
Anthropic points Claude Sonnet 5.5 at tightly scoped work: fixing bugs, producing documents, slide decks and spreadsheets. The positioning is stated plainly, with Opus 5.5 holding the long, open-ended jobs that call for judgment over time.
On OSWorld 2.1, which measures driving a computer, Claude Sonnet 5.5 reaches 80.1 percent against 57.0 percent for its predecessor. Sonnet 5 had already shown that a mid-tier model could fix code better than a room full of researchers.
More articles on Horizon
- Bernie Sanders Wants to Ban AI Superintelligence
- ElevenLabs Is Now Valued at $22 Billion
- OpenAI Hacked an Australian Government Website
Anthropic Tightens Its Lineup and Rivals Lose a Talking Point
The gap between the two models now sits at a handful of points across several evaluations. For Anthropic, that moves the model-selection question toward the shape of the task rather more than toward the budget available.
Rivals lose a convenient line of argument. A lab selling a competent mid-tier model against Anthropic’s flagship now finds itself measured against a mid-tier model that beats that flagship on agentic coding.
Distribution pushes in the same direction. Claude Sonnet 5.5 arrived on all three major clouds and inside GitHub Copilot at once, which leaves very little room for a slow pricing response.
There is a longer-term question underneath the scorecard. A flagship whose main defence is judgment on extended work has to keep proving that defence release after release, and each mid-tier launch that closes the gap makes the premium harder to justify to a finance team.
The release cadence itself is part of the message. Opus 5.5 on September 22, Claude Sonnet 5.5 on September 28: six days between a flagship and the cheaper model that nearly matches it. Teams that spent a week benchmarking the first one now have a second set of numbers to run, while procurement cycles run on a scale of months.
What real usage returns outside in-house evaluations is still open. Anthropic itself concedes that Opus 5.5 keeps the edge as soon as a task stretches and turns messy, and that is exactly where the value of production agents is decided today.
Follow the story on Horizon.
