OpenAI reduces GPT-5.6 Sol API pricing by more than 20% for three months

OpenAI announced a price cut of over 20% for its GPT-5.6 Sol API and credit usage, effective for the next three months, citing efficiency gains. The move comes as Anthropic prepares an IPO and OpenAI reported 18% revenue growth, while Anthropic saw near‑double quarterly growth. Analysts view the discount as a competitive tactic in the fast‑evolving generative AI market.

OpenAI cut developer pricing for its frontier GPT-5.6 Sol model by more than 20% on Friday, a promotional rate the company says will hold for the next three months. The reduction is concrete rather than headline-only: input tokens drop from $5 to $4 per million and output tokens from $30 to $20 per million, which works out to a 20% cut on input and roughly 33% on output, the side of the ledger that dominates costs for agentic and long-generation workloads.

The scope matters as much as the size. The new rate applies to pay-as-you-go API calls, Codex credits, and eligible ChatGPT Work plans, while Pro, Plus, and Business subscriptions stay at their existing prices. OpenAI has listed the discount as promotional and available through at least November 21, 2026, so this is a defensive move on developer volume rather than a permanent repricing of the model.

The competitive backdrop explains the timing. OpenAI is facing simultaneous pressure from Anthropic, which is preparing an initial public offering, and from lower-cost Chinese models, while Meta has been pushing a cheaper coding agent into the same developer market. The growth comparison is unflattering in relative terms: OpenAI posted an 18% quarter-over-quarter revenue increase against Anthropic's near-doubling over the same period , which suggests the discount is aimed at defending share in the API and coding-agent segment specifically.

For the broader market, frontier-model inference pricing is now visibly deflating, and that cuts two ways. Cheaper tokens lower the cost of building AI products, which supports the application layer and the software names that resell inference. They also compress the revenue per unit of compute that funds the data-center buildout, so the useful thing to watch is whether the three-month window produces enough volume growth to offset the lower rate, and whether competitors match it before November.