The most expensive agent in this vendor's benchmark was its own previous release

The arithmetic that makes this table move is in the payload, not the plan. opencode's cheapest run needed 98,789 tokens across eight requests ^[16], which works out to about 12,349 tokens per request ^[3]. The vendor's own note that six opencode steps at that weight roughly equal its total consumption for the whole task ^[9] puts its sixteen requests somewhere near 4,500 tokens each ^[4]. Same billing rate: credits per prompt token land within a few percent across all four runs, and the vendor's is marginally the highest of the set ^[6]. Nobody got a discount. The invoice moved on volume alone ^[6].

Two things set that volume, and the vendor names both: a tool catalogue of 21,188 bytes against 7,703, roughly triple the standing overhead and charged on every call whether the model touches those tools or not ^[7], and a transcript that keeps resending stale tool output at full length unless something trims it ^[8]. That is why the step-count intuition inverts here. opencode used eight to eleven requests against the vendor's sixteen ^[5], so scoring on turns hands opencode the win while the bill goes the other way. At opencode's highest observed turn count its average run still costs about 196 credits per request, against 81 for the 2.6.6 agent ^[5].

The row worth reading twice is 4395, which is not a constructed strawman but what the vendor shipped one release earlier as 2.6.5 ^[4]. That is about 3.4 times its own successor on the identical task ^[1] and roughly double opencode's three-run average ^[6]. The vendor says the 2.6.5-to-2.6.6 work targeted exactly those two items and cut credit consumption 78.6 percent across its tool-driven runs, 80.4 percent on the longest ^[10]. On this particular bugfix the cut is 70.5 percent ^[2], noticeably less than the headline figure. A vendor building a chart to flatter itself would not usually publish the version of the number that is worse.

The limits are stated rather than buried, and they are real. opencode ran three times, the vendor once, and the spread inside opencode alone is 45 percent, from 1679 to 2433 credits, while the vendor's own variance is unknown ^[12]. opencode ran as it ships, and a trimmed tool set would land elsewhere ^[13]. The scenario is one tiny repo and a one-line bug, with nothing said about multi-file refactors or hour-long sessions ^[11]. opencode is free software, so what is measured is the model bill and not a licence ^[14]. Every run in the table produced correct, committed work under criteria fixed before the runs ^[2][15].

What survives all that is narrower than 40 percent and more useful. On short, well-scoped tasks, opencode can hardly land below the vendor because the fixed per-request overhead sets a floor ^[17]. That mechanism does not care whose agent it is, which is the point of the 2.6.5 row: the same measurement pointed at your own shipped loop will find the same standing block. Anyone running an agent in production can price this without a benchmark harness by measuring the size of the block sent on every call and whether anything at all trims the transcript behind it.