OpenAI Cut Sol's Output Price 33%. Past 272K Tokens, It Didn't.

TL;DR

On August 21 OpenAI cut developer pricing on GPT-5.6 Sol. Short-context input went from $5 to $4 per million tokens and output from $30 to $20, a 20 percent and 33 percent cut. Reuters reports the cut runs three months; OpenAI's own pricing page words it as promotional pricing available at least through November 21, 2026. The detail that did not make any headline: the discount only exists below 272,000 input tokens. Cross that line and Sol bills at 2x input and 1.5x output, which lands at $8 and $30. Thirty dollars is exactly what output cost last week.

What actually changed

The published rates on Sol now read like this, per million tokens:

• Short context (up to 272,000 input tokens): $4.00 input, $0.40 cached input, $20.00 output.
• Long context (above 272,000 input tokens): $8.00 input, $0.80 cached input, $30.00 output.
• Previous short-context list: $5.00 input, $30.00 output.

Cached input stays at a tenth of the input rate, so the cache discount got cheaper in absolute terms along with everything else. Sol's siblings are unchanged in this round: Terra sits at $2 and $12, Luna at $0.20 and $1.20, both short context. Those two took their cuts last month, Terra by 20 percent and Luna by 80 percent.

Reuters puts the scope at the API plus credits on ChatGPT Work and the Codex coding tool. Pro, Plus and Business subscription pricing is untouched, so if you pay a flat monthly fee for a seat, absolutely nothing happened to you today.

The 272,000-token asterisk

Sol carries a 1,050,000-token context window, with a documented ceiling of 922,000 input tokens and 128,000 output tokens per request. The billing does not treat that window as one flat rate. Requests that exceed 272,000 input tokens are charged at 2x input and 1.5x output for the whole request.

Run the multiplier against the new base and the result is the joke that writes itself. Output above the line is 1.5 times $20, which is $30, which is the number you were paying on short requests seven days ago. The cut is real, it is just a cut on the cheap half of the pricing table.

Think of it as a phone plan with a data cap. The advertised rate is genuinely lower, and it applies right up until the moment you actually use the thing hard, at which point you are quietly back on the old tariff.

This matters more than it sounds, because 272,000 tokens is not an exotic threshold for anyone running agents. A long coding session with a repo map, a few thousand lines of context, tool schemas, and forty turns of accumulated history walks past it without trying. The workloads most likely to blow through the boundary are exactly the workloads that generate the most output tokens, which is the axis where the discount was supposed to be biggest.

It undercuts Opus 5 on both axes now

Last week Sol matched Claude Opus 5 on input at $5 and sat 20 percent above it on output, $30 against $25. At $4 and $20, Sol is now below Opus 5 on both. Claude Fable 5 lists at $10 and $50, and Claude Sonnet 5 at $2 and $10.

Put a concrete month against it. Take a workload that burns 50 million input tokens and 10 million output tokens, all of it under the long-context line, no caching, no batching:

That is a 27 percent cut on the total bill for that shape of workload, and it flips the ordering against Anthropic's Opus tier. It does not touch Sonnet 5, which remains half the price of the new Sol and is a different argument entirely.

Two caveats before you rewrite your router config. Anthropic prices its full million-token context at the standard rate with no long-context step, so the comparison inverts once your requests get big. And Claude 4.7 and later models use a tokenizer that Anthropic says produces roughly 30 percent more tokens for the same text, which is a per-token price comparison and a per-request cost comparison quietly disagreeing with each other. Price the workload, not the rate card.

What "at least through November 21" actually buys you

The wording on OpenAI's pricing page is doing real work. It is not a permanent reduction and it is not a hard cliff either. It is a floor on the duration with no stated ceiling, which means your November budget has a fork in it that you cannot resolve today.

There is precedent in both directions. Anthropic announced Sonnet 5 at $2 and $10 as introductory pricing through August 31, then made it permanent and cancelled the scheduled increase to $3 and $15. OpenAI has also gone the other way on other lines. Neither history tells you what happens on November 22.

The practical read: treat $4 and $20 as the rate for the next three months and treat $5 and $30 as the number you model against for anything that has to survive Q4. If your margin only works at the promotional rate, you do not have a margin, you have a coupon.

What to do about it

• Instrument the 272K boundary. If you are not already logging input token counts per request, you cannot tell which slice of your bill is getting the discount and which is quietly paying 2x. Alert on requests crossing the line.
• Compact before you cross. Trimming a 280,000-token request down to 260,000 halves the input rate and cuts a third off output for that request. That is the largest single lever in this pricing table and it costs you a summarization pass.
• Recheck your cheap-model routing. Sol at $20 output is now closer to the mid tier than it was. Some of the work you pushed to Terra to save money may no longer be worth the quality tradeoff.
• Do not rip out Anthropic over this. The gap is $100 a month on the example workload above, and it reverses on large-context requests. Switching a production router for a promotional rate that expires in November is how you end up doing the migration twice.

Key Takeaways

• GPT-5.6 Sol short-context pricing dropped on August 21 from $5 and $30 to $4 and $20 per million tokens, a 20 percent input cut and a 33 percent output cut.
• The discount applies only below 272,000 input tokens. Above that, billing is 2x input and 1.5x output, or $8 and $30, so long-context output costs exactly what short-context output cost before the cut.
• OpenAI describes the rate as promotional and available at least through November 21, 2026. Reuters reports it as a three-month cut.
• Sol now lists below Claude Opus 5 on both input and output, where it previously matched on input and sat 20 percent above on output.
• The cut covers the API plus credits on ChatGPT Work and Codex. Pro, Plus and Business subscriptions are unchanged.
• Anthropic charges standard rates across its full million-token window, so the cheaper option flips depending on how large your requests actually are.

Sources: OpenAI API pricing, OpenAI GPT-5.6 Sol model page, Reuters, Reuters via Investing.com, Anthropic pricing documentation