OpenAI Cuts GPT-5.6 Sol API Prices by 20%, for Three Months

When the market leader cuts prices, is that generosity or pressure? OpenAI announced on Friday that developer pricing for its frontier model GPT-5.6 Sol is dropping by more than 20 percent: input falls from $5 to $4 per million tokens, output from $30 to $20. The cut lasts three months and covers the API, Codex credits, and eligible ChatGPT Work plans. Consumer subscriptions, Pro, Plus, and Business, stay where they were.

So the discount is aimed at one audience only: the developers building products on top of OpenAI's models. Which answers the opening question.

Second cut in a month

This is the second price move in the GPT-5.6 family within weeks. In late July, the mid-tier Terra model dropped 20 percent and the budget Luna model a full 80 percent. Reuters names the reason without hedging: competition from Anthropic and from Chinese models. The specifics are easy to list. xAI markets Grok 4.6 as matching Sol's benchmarks at 60 percent less. DeepSeek sells input tokens at $0.14 per million. Anthropic, meanwhile, reported that businesses have been tilting toward Claude. Frontier-tier quality no longer excuses frontier-tier pricing on its own.

Note the expiry date, though. A permanent cut would reset the market's reference price forever; a three-month one buys a quarter of developer loyalty while keeping the old rates available for a quiet return. That asymmetry tells you how OpenAI wants this read: a retention offer, priced like one.

Does a temporary discount deserve a migration?

Here's a quick sizing. A product burning 50 million input and 10 million output tokens a month on Sol paid $550 before Friday and pays $400 now, roughly 27 percent less, with output-heavy workloads saving more. Real money, but the strategic question isn't whether to enjoy the discount. It's whether to make decisions you'll regret when it lapses.

Our advice runs in two directions. Budget against the old prices, not the promotional ones; a plan that only works at $4 input is a plan with a countdown timer attached. And use the window for the experiment that pays off regardless: routing. Most workloads mix routine tasks (classification, extraction, summarizing) with genuinely hard reasoning, and only the second tier needs a frontier model. Teams that split traffic between a cheap model and Sol routinely cut bills by half or more, before any vendor discount. A price war is a fine time to build the architecture that makes you indifferent to who wins it.

Sources: Reuters, OpenAI Developer Community