Opus 5.5 cost less per unit of coding work than Sonnet 5, with fewer review rounds needed

An agent's bill follows two numbers: how many API calls it makes, and how much context each call re-sends ^[5]. Every step sends the whole conversation again, including the brief, every file read and every test log, and most of it comes back from the prompt cache ^[5]. Across 57 agents in the post, cost tracked calls multiplied by average context size with a correlation of 0.99 ^[6]. In one six-hour window on September 24, the agents used about $593 at list prices, and 68% of that was cached reads ^[7]. Re-reading context alone came to roughly $403 ^[27].

Per-token list price touches only part of that bill. For the median Sonnet implementer lane, cached reads were about three quarters of spend ^[10]. Only the remaining quarter doubles on Opus. The same call therefore costs 0.75 + 0.25 x 2 = 1.25 times as much, about 25% more ^[12]. The rest of the gap has to come from making fewer, smaller calls. Per 100 changed lines, Opus 5.5 made six times fewer calls than Sonnet, read ten times less context and wrote three times less output ^[14]. Median PR size was close, at 504 lines for Opus 5.5 and 484 for Sonnet ^[13].

The first draft is close to even. Up to the first review verdict, Opus 5.5 cost about 0.77 times Sonnet, and that range includes equal cost ^[17]. It got there in a median 31 calls to Sonnet's 97 ^[17]. Fix rounds took about 71% of implementer spend in the September 24 window ^[18]. Sonnet implementers ran 400 to 600 calls while their context grew to between 600K and 970K tokens, and nothing compacted it before about 967K ^[8]. One of them, holding a 900K context, paid $5.60 to $6.40 for a small last-round fix, because it re-read everything to make it ^[19].

The reviewer is well built. It runs the tests, breaks the code on purpose to check that a test goes red, and checks the claims in the PR description before anything merges ^[25]. Its strictness sets how often fix rounds happen. Sonnet 5 and Opus 5 passed first review 12% to 20% of the time, and all 20 PRs that changed production code in the September 24 window failed it ^[21].

Opus 5 shows efficiency and price pulling apart. Its prices are exactly 2.5 times Sonnet's on every token type, and at Sonnet's prices it would have cost $0.98 per 100 changed lines against Sonnet's $1.57 ^[22]. "Opus 5 was not wasteful; its price was," the author wrote ^[23]. Opus 5 ran only on September 20 and 21, and Opus 5.5 only on September 23 to 25, each compared with Sonnet on the same days ^[26].

The post states its own limits. The raw table put Opus 5.5 at a fifth of Sonnet's cost, but the author chose which tasks went to Opus and gave it mostly authentication and crypto work ^[15]. The 0.45 figure is the author's correction for that selection, and the review-round comparison rests on five Opus lanes ^[16][20]. The dollars are API list prices, used to measure how fast a flat subscription's usage limit drained ^[4].

For the number to hold on another team's workload, its agents would need cached-read-heavy bills, large uncompacted contexts and a reviewer that sends most PRs back ^[6][8][21]. A team whose PRs mostly pass first review is paying mainly for first drafts, and there this data shows no clear gap between the two models ^[17].