OpenAI: GPT-5.5

For three releases running, "Pro" listed at twelve times the price. openai/gpt-5-pro heads at $15.00per million tokensopenrouter-models, last checked 2026-09-14 against plain openai/gpt-5's $1.25per million tokensopenrouter-models, last checked 2026-09-14; GPT-5.2 Pro repeated the ratio ($21.00per million tokensopenrouter-models, last checked 2026-09-14 against $1.75per million tokensopenrouter-models, last checked 2026-09-14), and GPT-5.4 Pro repeated it again ($30.00per million tokensopenrouter-models, last checked 2026-09-14 against $2.50per million tokensopenrouter-models, last checked 2026-09-14). This row is where the multiplier broke: openai/gpt-5.5-pro lists at $30.00per million tokensopenrouter-models, last checked 2026-09-14 against this row's own $5.00per million tokensopenrouter-models, last checked 2026-09-14 — half the premium of the three releases before it. Each of those eight figures is the top listed provider's rate for its row rather than necessarily OpenAI's own, and two rows are not obliged to be headed by the same provider, so every ratio here is a gap between two listings.

The slide didn't stop there. One generation later, each of the three GPT-5.6 variants — Sol, Terra, Luna — has its Pro sibling listing at exactly the same rate as the plain row: openai/gpt-5.6-sol-pro lists at $2.00per million tokensopenrouter-models, last checked 2026-09-14, identical to $2.00per million tokensopenrouter-models, last checked 2026-09-14 for Sol itself, and Terra and Luna match their own base rows the same way. Across the five releases that shipped a Pro row — GPT-5, 5.2, 5.4, 5.5 and 5.6 — the premium attached to the "Pro" label in these listings went from twelve times the base rate to nothing, as observed on 28 August 2026 — again a comparison between top-listed-provider rates rather than between two rates one company is known to have set. These are not launch prices: OpenAI cut Luna 80% and Terra 20% on 30 July 2026, though base and Pro moved together and the ratios are undisturbed.

Reasoning followed a similar on/off pattern. The reasoning.default_enabled flag reads trueopenrouter-models, last checked 2026-09-14 on this row, where GPT-5.4 carried the opposite value in the same snapshot — a bare request to 5.5 spends reasoning tokens that the same request to 5.4 would not have, on top of the price move the Pro tier already made.