OpenAI Cuts GPT-5.6 Sol API and Codex Credit Pricing by Over 20%

OpenAI's own framing of the cut as "over 20%" undersells what's actually happening to the more expensive side of the bill. Input tokens for GPT-5.6 Sol fell exactly 20%, from $5 to $4 per million tokens, but output tokens - the ones that cost six times more per token to begin with - dropped 33%, from $30 to $20 per million ^[1]. Cached input tokens, the discounted rate OpenAI charges for context a model has already processed once, fell from $0.50 to $0.40 per million ^[1]. Because output tokens are the expensive half of almost every real workload, the products that benefit most are the ones that generate a lot of text rather than just read it: coding agents drafting whole files, customer-support bots writing full responses, and agentic tools like Codex that loop through many turns of model output ^[2]. A chatbot that mostly reads a prompt and returns a short answer barely notices the difference; an agent that writes thousands of tokens of code or tool calls sees close to a third off its heaviest cost line.

This is not an isolated discount - it is the latest step in a rapid-fire repricing of the entire GPT-5.6 lineup. The family went generally available on July 9, 2026, with Sol launching at its original $5/$30 rate ^[3]. Three weeks later, on July 30, OpenAI cut the mid-tier Terra model by 20% and the budget Luna model by a striking 80% on input tokens (from $1 to $0.20 per million) - while explicitly leaving Sol's price untouched at that time ^[3]. Now, another three weeks on, Sol itself gets cut too, marking OpenAI's second GPT-5.6 pricing adjustment in under a month ^[4]. Read together, the sequence looks less like a single generous gesture and more like a company working down its own price list tier by tier as competitive pressure builds, adjusting whichever model is under the most heat at a given moment rather than repricing everything at once.

The timing lines up with a specific competitive target. After the cut, GPT-5.6 Sol's $4/$20 pricing sits at roughly 40% of Anthropic's Claude Fable 5, which charges $10 per million input tokens and $50 per million output tokens, and comes in below Claude Opus 5's $5/$25 rate as well ^[4]. That is a meaningful price/performance gap on paper. Yet the same reporting notes Anthropic reportedly still holds a larger share of enterprise adoption than OpenAI (43.5% versus 39.7%), even as Sol reportedly dominates spend among the highest-performance token usage ^[4]. Pricing for OpenAI's own subscription tiers - Plus, Pro, and Business - was left completely unchanged in this round ^[5], which sharpens the read: this is a developer-and-agent-workload play aimed at usage-based customers, not a broader move to win over ChatGPT's subscriber base. Cutting API rates below a rival while holding your own subscription prices flat is a way to compete on cost where the competition is fiercest - server-side, token-metered workloads - without touching the margin on the product most consumers actually pay for.

OpenAI has been explicit that this is promotional, not permanent: the new $4/$20 pricing is guaranteed to hold only through at least November 21, 2026, matching the stated three-month window from the August 21 announcement ^[6]. That framing matters for anyone building a cost model around the new rate - a lower price with an expiration date is a different planning problem than a lower price with no end in sight. Community reaction reflected some of that skepticism: discussion around the cut skewed mixed rather than celebratory, with developers pointing out that subscription pricing didn't move at all and questioning how sustainable steep API discounts are if OpenAI is still investing heavily in training and infrastructure. The tension between "frontier lab passes along efficiency gains" and "frontier lab defends usage share under competitive pressure" is unresolved by the announcement itself - both readings are consistent with the same set of facts, and which one proves right will depend on whether the discount is renewed, extended, or allowed to lapse when November 21 arrives.