OpenAI published a joint price-performance post with AWS on Monday about GPT-5.6 running in Kiro, Amazon's spec-driven coding IDE, headlining a roughly 82% reduction in the cost of completed tasks on Terminal-Bench 2.1. Trade coverage read it as a launch — "available in Kiro starting August 24, 2026." Kiro's own model changelog dates the arrival of GPT-5.6 to 14 July 2026, and carries no entry at all for 24 or 25 August.

What the conventional framing gets wrong

Three errors, stacked. First, this is not a launch. GPT-5.6 has been in Kiro for six weeks; Monday's artefact is a co-marketing post. Second, the 82% is a vendor-run benchmark cost delta, not a customer bill and not an accuracy result. No accuracy figure for the Kiro configuration is given, and nothing separates what the model saved from what the harness saved. Third — and this is the one that should give buyers pause — the pricing lever was pulled three weeks earlier. On 31 July Kiro cut Luna's credit multiplier from 0.6x to 0.1x and Terra's from 1.2x to 1.0x. A reader encountering "82% cheaper" alongside no mention of that change could easily credit the model for a billing decision.

The dates, from both vendors' own records

Kiro's 14 July entry announced GPT-5.6 Sol, Terra and Luna with a 272K context window and multipliers of 2.4x, 1.2x and 0.6x, as "experimental support ... rolling out to Pro, Pro+, Pro Max, and Power customers" in two regions — us-east-1 and eu-central-1. Terminal-Bench 2.1 scores published then: Sol 88.8%, Terra 87.4%, Luna 84.7%. The status is still experimental, still on named paid tiers, still two regions. Nothing about that changed on Monday.

Why this pattern is going to recur

Model vendors are now publishing cost-per-completed-task claims inside a partner's harness, where the model, the scaffolding and the credit multiplier all move independently and none of them are held constant. A buyer comparing coding tools on "82% cheaper" has no way to attribute the saving to any of the three. The attribution problem is not incidental to this kind of announcement; it is what makes the announcement possible.

What would make the number checkable

An accuracy figure alongside the cost figure, a fixed harness version, and a note on which credit multipliers were in force during the run. None of the three is present.