OpenAI (pricing)

Corrected 2026-09-24 — OpenAI does cut its rungs, often and fast. This page used to present OpenAI as the clearest example of a lab that never lowers a price point. Its own announcement, changelog and pricing page, ingested 2026-09-24 (close-read), show the opposite for every GPT-5.6 tier:

Date	Move (per million tokens, input/output)	Source
2026-07-09	GPT-5.6 launches: Sol $5/$30 · Terra $2.50/$15 · Luna $1/$6	changelog; rows recorded by the Token Price Index 2026-07-16
2026-07-30	Luna −80% → $0.20/$1.20 · Terra −20% → $2.00/$12 · Sol unchanged	changelog + pricing page
2026-08-21	Sol → $4/$20 (−20% in, −33% out), "promotional ... at least through November 21, 2026"	changelog + pricing page
2026-09-22	GPT-6 Sol $2/$10 · GPT-6 Luna $0.10/$0.50; GPT-6 Astra listed at $10/$50	changelog + pricing page

OpenAI's stated reason for the July cut: GPT-5.6 was set to optimise its own serving, giving 20% lower serving cost from GPU kernel work and 15%+ better token generation from speculative decoding. The launch date is 2026-07-09 (the Index recorded the rows a week later). What OpenAI does hold is a price point at the very top of its list, now $10/$50 with GPT-6 Astra. The text below is kept as the record of the earlier reading, with its claims struck where they no longer hold.

Type: Company — inference-economics lens (pricing behaviour only)

This is the pricing-lensed stub for OpenAI in the inference-economics topic. The canonical OpenAI profile lives on the models rung at models/wiki/entities/openai.md; this page tracks only how OpenAI sets and moves token prices — the observable that feeds the topic's concepts.

OpenAI is the clearest published example of price-rung persistence: it prices by capability tier, and a new model generation steps onto the existing rung rather than lowering it. (Wrong as of 2026-09-24 — see the correction above. It prices by tier, but it cuts the tiers.) The GPT-5.6 release (launched 2026-07-09, recorded by the Index 2026-07-16) was the anchor observation — Sol (frontier) and Terra (strong) shipped at the exact prior-generation per-token price ($5/$30 and $2.50/$15), while Luna introduced a new, higher mid rung ($1/$6) rather than cutting the old one.

Key Contributions (to this topic)

• Held the frontier rung across a generation — until 2026-08-21: GPT-5.6 Sol = $5/$30 = GPT-5.5's exact rung at launch; cut to $4/$20 (promotional) on 2026-08-21. Evidence: strong (own dataset, primary vs OpenAI's pricing page 2026-07-16) (TPI)
• Held the strong rung across a generation — at release only: GPT-5.6 Terra = $2.50/$15 = GPT-5.4's exact rung on 2026-07-16; cut to $2.00/$12 on 2026-07-30. Evidence: strong (own dataset, primary) (TPI)
• Grew a new mid rung rather than cutting the old one — for two weeks: Luna $1/$6 at release, above the prior GPT-5.4-mini rung ($0.75/$4.50); cut 80% to $0.20/$1.20 on 2026-07-30. Evidence: strong (own dataset, primary) (TPI)
• Named a software cause for a price cut, on the record: 20% lower serving cost from GPU kernel improvements and 15%+ better token generation from improved speculative decoding, found by pointing GPT-5.6 at its own runtime. The 20% explains Terra's cut; Luna's 80% goes well beyond it. Evidence: moderate (vendor-stated percentages, no dollar base) (OpenAI)
• Sells speed as its own price line: Fast mode (from 2026-07-30, replacing Priority Processing) is up to 2.5x faster than standard at exactly twice the price on every GPT-5.6 row. Evidence: strong (pricing page, 2026-09-24) (OpenAI)
• OpenAI's rung-holding is, in Matsuoka's model, one of the two conditions the announced buildout's solvency depends on ("sticky premium pricing", alongside ~2x annual token-demand growth for four years). Watching OpenAI's rungs is therefore watching a modelled solvency variable. Evidence: weak-moderate for the corridor model (single-author preprint, modelled); strong for the rung observations themselves (Matsuoka)

Blind spot, stated (2026-07-22): this page tracks published list prices only. Enterprise discounting, committed-use agreements and batch/cached tiers are invisible to it, so "OpenAI held its rung" is a claim about list price, never about realised revenue per token.

Mentioned In

• Price-Rung Persistence — OpenAI's GPT-5.6 rung-hold is the anchor evidence.
• The Price-Decline Distribution — OpenAI's held rungs are one mechanism behind the fast decline in the price of fixed capability.
• The Depreciation Conveyor & Vintage Economics (capital-structure) — premium stickiness as a solvency condition.

Related Entities

• Anthropic (economics) — the other frontier-lab economics stub in this topic (margin claim, attributed).
• vllm-cost-meter — the cost-side counterpart instrument; this page is the price side.

Changelog

• 2026-09-24 (source ingest) — OpenAI's own announcement, changelog and pricing page ingested as one source; removed the "proposed, awaiting approval" note. Corrected the page's central claim (that OpenAI does not cut its rungs): every GPT-5.6 tier has been cut, the frontier one on 2026-08-21 (promotional). Added the stated cause, Fast mode, GPT-6 prices, and settled the launch date as 2026-07-09.
• 2026-09-24 — Added the 2026-07-30 Terra/Luna cuts so the page no longer presents a partly-broken hold as current; noted the pending primary source (proposed, awaiting approval) and the unsettled launch date. Fixed the canonical-profile and depreciation-conveyor links.
• 2026-07-22 — Linked rung-holding to Matsuoka's "sticky premium pricing" solvency condition (2607.07207); stated the list-price-only blind spot explicitly.
• 2026-07-16 — Stub created (pricing lens). Anchored on the GPT-5.6 rung-hold observation from the Token Price Index. Canonical OpenAI profile stays at ai/wiki/entities/openai.md.