Failed runs decide whether Claude Sonnet 5.5 undercuts Opus 5.5

Per-task cost is token price multiplied by tokens consumed. Sonnet 5.5 lists at exactly half of Opus 5.5 on both input and output ^[13], so it breaks even when it spends twice the tokens ^[21]. The New Stack ran three tests through the Anthropic API at maximum effort with adaptive thinking. Each model got five runs per test, graded against hidden tests the models never saw ^[4]. On the resolver spec, Sonnet used 15% more output tokens than Opus and came in 42% cheaper ^[18]. On the agentic bug fix it used 2.07 times as many ^[14], and its saving fell to about 7% a run before any failed runs were counted ^[15].

The failed runs trace to one harness setting. Opus ran under the same 32,000-token step cap and never came close to it ^[7]. Both models ended up fixing all four planted bugs on every run ^[5]. Sonnet's truncated first attempts cost about $1.40, and The New Stack left them out of its reported averages ^[8]. Spread across five runs, that adds $0.28 each and lifts Sonnet from $0.70 to $0.98 against Opus's $0.75 ^[16]. A 7% saving becomes a 31% premium ^[22].

A higher cap has its own failure. Twice on the concurrency task, Opus spent 128,000 tokens deliberating over three race conditions and kept its conclusions to itself ^[10]. Those empty runs are inside its $2.24 per-run average ^[11]. Five runs at $2.24 is $11.20, and only three produced a passing fix, so each completed job cost about $3.73 ^[17]. Sonnet's $1.02 a run bought a passing fix every time [c12, c13].

Speed split as well. Opus finished the bug fix in about 35% less time ^[20]. Sonnet was faster on the resolver and on the concurrency task [c11, c13].

The two independent measurements disagree. Artificial Analysis's max-effort figures put Sonnet 28% above Opus per task ^[12]. The New Stack's concurrency runs had Sonnet about 54% cheaper per run ^[19]. The New Stack's article does not describe Artificial Analysis's task set, so the record cannot say which workload each result resembles. The New Stack's own sample is 15 runs per model across three tests ^[23]. For any of these figures to carry over, a team's jobs would need to match the setup: maximum effort, a per-step cap near 32,000 or 128,000 tokens, and pass-or-fail grading against hidden tests [c6, c9].

In my view the comparison belongs at the level of cost per completed job, measured on a sample of a team's own tasks with failed runs left in the total. By that measure The New Stack's tests went to Sonnet twice and to Opus once [d11, d7, d6].