GPT-5.6 Sol Price Cut Takes Flagship API Rate to $4/$20 Until November

What happened

On August 21, OpenAI cut the price of GPT-5.6 Sol, its flagship model, for developers. Input now costs $4 per million tokens and output costs $20 per million tokens for standard short-context requests. The previous rates were $5 and $30.35 OpenAI described the discount as "over 20%" for three months. That label understates the cut: input fell 20% and output fell about 33%.14 OpenAI says the promotional rate will last at least through November 21, 2026, and it has not said what happens after that.116

The new rates apply to the API and are rolling out to eligible credit plans for ChatGPT Work, OpenAI's agentic product, and Codex, its coding tool. Pro, Plus and Business subscriptions are unaffected.23 OpenAI's developer-forum announcement says the discount also covers Fast mode, long-context requests, and Batch and Flex processing.4 Reuters-syndicated coverage linked the cut to growing competition from Anthropic and from Chinese AI labs.35 OpenAI's own explanation was different. It said efficiency gains made the cut possible, not any rival's pricing.17

Seven weeks later, the Sol cut has turned out to be one step in a broader repricing. OpenAI and Google have both released new models since then, and several of them launched at introductory prices that are scheduled to end.

Agreement on the numbers, with some errors

Most coverage reports the same core figures. Reuters-derived reports, OpenAI's forum post and independent pricing trackers all list $4 input, $0.40 cached input and $20 output.2416 Under the cut, Batch processing costs $2 input and $10 output, and Fast mode costs $8 and $40.815

There are still differences between reports. One business outlet assumed a flat 20% discount and estimated output at about $24. OpenAI's published price is $20.64 Reports also give different start dates, August 21 or August 22, which looks like a time-zone difference in when OpenAI's post went out.89

The disagreement that affects budgets most is long context. One explainer said the cut did not appear to change the long-context multiplier.8 OpenAI's forum post says the reduction applies to long-context requests.4 Current price sheets list Sol at $8 input and $30 output once a request exceeds 272,000 input tokens.1115 The rule is that input doubles and output rises 1.5 times above that threshold.18 Long-context prices did fall. Because the multipliers apply to a lower base rate, though, very long prompts still cost noticeably more than the headline numbers suggest.

There was also confusion with a separate promotion. A few days before OpenAI's cut, the routing service OpenRouter showed Sol at "50% off." That was OpenRouter's own discount for non-BYOK traffic, not a change to OpenAI's prices.8 OpenRouter is still running it, at $2 input and $10 output on eligible routes.17

Why the output cut matters more

The larger cut on output tokens is the most revealing part of the announcement. Agentic and coding workloads, the ones OpenAI named when it extended the discount to Codex and ChatGPT Work credits, spend most of their money on output tokens.1 Cutting output by a third targets the spending that has grown fastest in enterprise budgets, at a time when Anthropic, Google and Chinese labs are all pursuing the same customers.1

The Sol cut also came soon after cuts to smaller models. Late in July, OpenAI permanently reduced prices for GPT-5.6 Terra by 20% and GPT-5.6 Luna by 80%.24 OpenAI said it had used GPT-5.6 to optimize its own serving, reporting 20% lower serving costs from GPU kernel improvements and more than 15% better token-generation efficiency from improved speculative decoding.4 Several commentators saw three cuts within a few weeks as a deliberate repricing of the whole family, not routine adjustment.97

Both explanations can be true. Lower serving costs allowed the cut, and competition determined when it happened and how large it was. The time limit is also a hedge. As one analysis noted, OpenAI can encourage developers to build on Sol without changing its long-term list price.6

Not all developers welcomed the change. Some said rate limits, not price, were their main constraint. One developer called Sol's limits "HORRENDOUS," a problem a lower per-token price does not solve.8

OpenAI's own new models made the discount less relevant

OpenAI then released new models that undercut the discounted Sol. GPT-6 Sol launched on September 23 at $2 input and $10 output, exactly half of GPT-5.6 Sol's discounted rate.19 At least one pricing tracker says GPT-6 Sol matches GPT-5.6 Sol's quality at that price.20 Another reviewer said they could not confirm the two are equal without testing them on real work.18 OpenAI's current lineup also includes GPT-6 Astra at $10 and $50 at the top, and GPT-6 Luna at $0.10 and $0.50 at the bottom.11

As a result, the cut that made news in August now mostly affects teams that are staying on GPT-5.6 Sol. One analysis said older tiers, including Sol at its promotional price, now cost more than GPT-6.1 Sol. Staying on them is a decision to keep behavior that has already been tested, not a way to save money.11

The new generation has also changed what subscribers get. Codex users on Reddit argued over whether the quota under GPT-6 was effectively halved. One reply calculated that, after accounting for the new model's lower per-token price, the dollar value per quota point was about the same.10

Google adopted the same approach

Google has been using temporary prices too. On October 1 it announced Gemini 4 Argon, its new frontier model, at an introductory $2 input and $10 output. Cached input is 95% cheaper than regular input.23 After the introductory period, Argon will cost $4 input and $20 output.2327 Google has not given an end date for the introductory rate.26

Argon's standard price after the introduction is exactly the same as GPT-5.6 Sol's discounted rate. Its introductory price is the same as GPT-6 Sol's. This is analysis, not something either company has said, but the pattern suggests $4/$20 has become the going rate for a top model and $2/$10 the rate for a launch promotion.

Coverage disagrees on whether developers can use Argon yet. Google says Argon is rolling out to a set of trusted cyber defenders through its Fairwind Program.23 Pricing trackers say there is no public API route and no rollout date.26 One outlet said Argon is not yet available even to Google AI Ultra subscribers.27 Another report said developers "can move onto it directly."28 Google's own wording supports the restricted-access reading, and developers should plan on that basis.

Google has been more explicit about temporary pricing for its Flash models. Gemini 3.8 Flash, which Google describes as its most capable Flash model for long-horizon software engineering and agents, costs $0.75 input and $3.75 output through December 31. On January 1, 2027, the price doubles to $1.50 and $7.50.2522 One pricing analysis counts seven Google models across three product lines with introductory rates that Google has already scheduled to end, and calls the January reversion a hidden budget cliff.30

Google is also tightening its consumer plans. From October 9, free Gemini app users are limited to the Flash-Lite model. Paying for Pro at $19.99 a month restores access to the full lineup.2429

What it means

The GPT-5.6 Sol cut looked like a straightforward discount in August. With the releases since then, it looks like the beginning of a period in which OpenAI and Google both set prices that are scheduled to change. OpenAI guarantees Sol's rate only until November 21 and has said nothing about what follows.16 Google has published the date its Flash prices rise but not when Argon's introductory rate ends.2526

For developers, the practical point is that the list price now matters less than when it expires. Any team estimating costs past late November should check four things for each model it uses:

• when the current rate ends
• what the long-context multiplier is
• what the rate limits are
• whether a newer model gives similar quality for half the price

The discounts are real, but they are temporary.