Sonnet 5.5 comes within two points of Opus 5.5 at half the per-token price

Someone asks Claude to summarize a long document, reads the answer, asks a follow-up and waits again. The Tom's Guide reviewer says that loop is most of their own Claude use, and hard problems are the exception ^[5]. The reviewer notes that a few extra seconds per response add up over a long conversation ^[6]. The speed gain Anthropic claims over Sonnet 5 applies to each of those waits ^[2].

The argument for defaulting to Opus has been quality. On the 44-occupation benchmark Anthropic cites, Sonnet 5 trailed Opus 5.5 by 397 points ^[3][4][18]. Sonnet 5.5 made up 395 of them ^[19].

Cost points the same way. Opus 5.5 charges twice Sonnet's per-token rate, which works out to $4 per million input tokens and $20 per million output ^[9][20]. Anthropic also says Sonnet 5.5 finishes a typical task in fewer tokens than Sonnet 5, cutting the cost of that task by up to 30% ^[8].

I'd make Sonnet 5.5 the default for API traffic and send work up to Opus by rule. The tradeoff is that the default would rest on one test, and every score so far is Anthropic's, as reported by Tom's Guide and BusinessToday ^[3][15]. Neither report identifies the benchmark or says which effort setting produced the 1,844 score. Anthropic pitches Sonnet for well-defined tasks that do not need continuous judgment ^[13]. Tom's Guide says Opus is still built for ambiguous problems and long jobs with many moving pieces ^[14].

Those two descriptions supply the axes of a 2x2. One axis is whether the task is well defined or ambiguous. The other is whether it wraps up in a few turns or runs long. An email rewrite or a document summary is well defined and short, and goes to Sonnet at Medium effort, the app default ^[11]. A scoped coding job is well defined but long, so it goes to Sonnet at High, the Platform default ^[11]. Anthropic's evidence for that case is a Terminal-Bench 4.0 score of 70.6%, up from Sonnet 5's 10.3% ^[15]. Ambiguous but short work starts on Sonnet and moves up when an answer fails, the habit the Tom's Guide reviewer says they now follow ^[16]. Ambiguous and long goes to Opus ^[14].

Rework is the test of the default. Messages per session will not settle it. A team can count how often Sonnet output gets redone on Opus, and how long each task takes to reach an answer someone accepts.