OpenAI halves token prices while courting a $1.2 trillion valuation

Half-price tokens, ninety minutes after Anthropic

OpenAI shipped GPT-6 Sol and Luna on 22 September at half the API price of the GPT-5.6 tiers they replace: Sol at $2 per million input tokens and $10 per million output, Luna at $0.10 and $0.50 ¹. Sam Altman's pitch on X↗: "half the price per token, and even less per task" ². Anthropic shipped Claude Opus 5.5 about ninety minutes earlier at $4 and $20 per million tokens, which it says is 40% less on a typical workload than Opus 5, and still double Sol's price ²⁴. One afternoon repriced the middle of the model market.

The trillion-dollar ask behind the discount

The cuts landed a week after Reuters↗ reported early investor talks on a round valuing OpenAI near $1.2 trillion↗, initiated by the investors ⁵. The deck behind the ask↗ projects $278 billion of negative free cash flow from 2026 through 2030, $856 billion of compute spend against $840 billion of cumulative revenue, and a tenfold revenue climb from $36 billion to $350 billion ⁶. On those projections, the compute bill alone tops all the revenue OpenAI expects to collect.

Agent builders banked the savings

Cached input reads now carry a 90% discount ⁷. In OpenAI's own customer posts↗, Manus lifted cache hit rates from about 85% to above 90%, Wordsmith cut inference costs 36%, and GitHub cut the share of prompt tokens needing fresh processing by more than half across billions of requests ⁷. Parallel halved the time and cost of research runs on GPT-6 Astra ⁸.

Against Anthropic, the ladder is priced per task. Sol scored 33.2% on AutomationBench for $0.27 a task, where OpenAI puts Claude Opus 5 at 26.9% and 11.1 times the cost, on rivals' published scores rather than head-to-head runs ¹². Anthropic's own launch numbers put Opus 5.5 at 40% task completion on the same bench ⁴.

The discount is a volume promise

OpenAI calls the cuts a pass-through, caching and inference savings handed straight to customers ¹. The arithmetic is sterner: a 50% price cut holds revenue only if volume more than doubles, and the deck already assumes a tenfold climb.

Pressure runs from below, where Chinese open-weight models from Alibaba, DeepSeek and Moonshot do much of this work for nothing ², and from Beijing, whose regulator opened a probe the same day into DeepSeek and Moonshot over routing user data to Claude↗ ⁹. Ramp lead economist Ara Kharazian told Fortune↗ the labs are in "a price war that is driving down the price of AI and therefore driving down their ability to profit from it" ³.

Selling tokens at half price, equity at full

OpenAI is selling tokens at half price and equity at full. If cheap-tier usage more than doubles, the discount pays for itself; if not, a $1.2 trillion valuation rests on a revenue line that just got marked down. Anthropic, expected to begin marketing its IPO in mid-October at the earliest, is watching the same dial ⁵.