Sonnet 5.5's extra tokens shrink its half-price edge over Opus to roughly a fifth at max effort

The independent coding win and the heaviest bill come from the same setting. Artificial Analysis's Terminal-Bench score is for Sonnet 5.5 at max effort ^[14]. At max effort the model wrote about 193,000 tokens per test task, the most the firm has measured and roughly 60% more than Opus 5.5 ^[8]. Half the per-token price times 1.6 times the tokens is 0.8. On that basis a max-effort Sonnet task costs about 20% less than the same task on Opus ^[24]. Opus would run near $9.50 a task against Sonnet's $7.60 ^[25].

That estimate assumes the token gap runs through the whole bill, and the figures leave a puzzle. At the $10 output rate ^[1], 193,000 written tokens come to about $1.93, roughly a quarter of $7.60 ^[19]. The rest is presumably input the model reads back on each step at $2 per million, though the source does not break the bill down. If Sonnet's longer runs also re-read more context, the saving stays near a fifth. If its input volume matches Opus's, the saving widens, since each input token costs half as much.

Both published savings figures can hold at once. Anthropic says Sonnet 5.5 costs up to 30% less per task than Sonnet 5 because it needs fewer tokens ^[22], while Artificial Analysis's max-effort bill points the other way ^[9]. The difference is the effort setting. Artificial Analysis calls High effort the best value ^[10], and Anthropic says that at High, Sonnet 5.5 matches GPT-6 Sol on FrontierCode for about a fifth of the cost per task ^[11].

The coding lead holds up under outside testing. Sonnet's margin over Opus on Terminal-Bench 4.0 is 4.2 points on Anthropic's self-reported table ^[3] and 4.0 points on Artificial Analysis's run ^[20], even though the independent scores sit 7 points lower ^[21]. Overall, the firm ranks Sonnet second behind Opus ^[7]. "Sonnet 5.5 is strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets. It's also got a sharp eye for design," Anthropic wrote ^[4]. The company says Opus 5.5 remains clearly stronger at complex work needing sustained judgment ^[5].

Anthropic held Sonnet's list price flat from Sonnet 5 ^[1] and left Opus at twice that, about $4 and $20 per million tokens ^[12]. It published no benchmarks against GPT-5.6 Terra, OpenAI's mid-tier model ^[17].

The result turns on where a team sets the dial. It can run Sonnet at Medium or High and keep most of the price gap. It can find it needs max effort to match Opus on its own tickets and save about a fifth. Or a retest could revise Artificial Analysis's figures, which came from a pre-release build with a bug ^[18]. In my view the first case fits bug fixes and scoped tickets, with Opus kept for the open-ended work Anthropic itself assigns it. The counter-case is that real tickets are messier than benchmark tasks, and teams drift to max effort anyway. The view is wrong if a team's own logs show Sonnet reaching Opus-level pass rates only at max effort, with a per-task bill within a fifth of Opus's.