OpenAI said on August 21 that it will cut developer pricing for its frontier model GPT-5.6 Sol by more than 20% over the next three months. For standard short-context use, input drops from 5 USD to 4 USD per 1 million tokens, and output falls from 30 USD to 20 USD. The change covers the API as well as credits for Codex and ChatGPT Work, while Pro, Plus and Business subscriptions stay the same. Competition from Anthropic and Chinese models is now showing up in per-token rates.

※1 USD = 159 JPY (as of August 21, 2026)

The real cut is on the output side, down by a third

The striking part of this change is that input and output did not move by the same amount. Input fell 20%, from 5 USD (about 795 yen) to 4 USD (about 640 yen) per 1 million tokens. Output went from 30 USD (about 4,770 yen) to 20 USD (about 3,180 yen), roughly a third off. The headline figure of more than 20% is the combined effect of that asymmetry.

Output is where most teams will feel it. Writing long-form text, generating code in bulk, or letting a model reason at length all push output tokens up. Workloads that feed in large documents and ask for short answers will instead lean on the 20% cut to input. The same headline number lands differently depending on which side of the ledger your workload sits on.

These rates apply to the standard short-context tier. Long-context usage sits on a separate table, and whether it moves by the same ratio is worth checking against the pricing page.

API and credits are covered, monthly plans are not

The cut applies to metered API usage and to credits for ChatGPT Work, OpenAI's agentic product, and Codex, its coding tool. On the credit side, the new rates are rolling out across eligible plans.

Pro, Plus and Business subscription fees are unchanged. For anyone who only uses ChatGPT through the interface, this is not a price cut at all. The beneficiaries are developers who embed the API in their own applications and workflows, and teams burning credits on agents and code generation.

The three-month window matters too. This is a time-limited move, not a permanent repricing. Rebuilding your architecture around three months of invoices leaves you exposed when the old rates return. Keeping both the old and new numbers in your cost model is the safer approach for now.

Second move in a month, and it started with the cheaper tiers

This is the second time GPT-5.6 pricing has moved in a month. At the end of July, the mid-tier GPT-5.6 Terra was cut by 20% and the low-cost Luna by 80%. Only those two tiers were affected then, and the frontier Sol model was left alone.

Seen in that order, the wave has now reached the top of the lineup. Cutting cheap models is largely a defensive move in a volume market. Touching the price of the flagship signals that pricing pressure has reached the segment that used to be won on capability alone.

OpenAI frames it as efficiency gains funding the reduction: it keeps pushing capability while improving efficiency, so it can charge less. The company has repeatedly described work on optimizing its own inference stack, so lower serving costs feeding into price is a reasonable reading. Still, a three-month expiry is hard to explain through cost structure alone, and a response to competitors is difficult to rule out.

How far has the gap to Claude narrowed

Anthropic's price list is the cleanest comparison. Its frontier model, Claude Fable 5, is listed at 10 USD (about 1,590 yen) per 1 million input tokens and 50 USD (about 7,950 yen) for output. One step down, Claude Opus 5 is listed at 5 USD (about 795 yen) input and 25 USD (about 3,975 yen) output.

With GPT-5.6 Sol now at 4 USD input and 20 USD output, it sits 60% below Fable 5 on both sides. Against Opus 5, it is 1 USD cheaper on input and 5 USD cheaper on output. A frontier-class model is now priced below a rival's second tier.

Unit price alone does not decide where workloads land, however. Models differ in how many tokens they need to finish the same task, and a verbose model can cost more overall despite a lower rate. Caching behavior, batch discounts, latency and the prompt assets a team has already built all factor in. A rate card is a starting point for comparison, not a conclusion.

Summary

OpenAI has cut developer pricing for GPT-5.6 Sol for three months, bringing input to 4 USD and output to 20 USD per 1 million tokens. The reduction is weighted toward output, and it covers the API plus Codex and ChatGPT Work credits, with monthly plans untouched. Coming after the late-July cuts to the cheaper tiers, it suggests price competition has now reached the frontier model itself. The time-limited nature of the move is worth building into any cost estimate.