OpenAI cut GPT-5.6 Sol to exactly 0.8 times Claude Opus 5 on every line it sells. Azure never applied the cut, so identical weights now trade at eight prices, and the cheapest one has an expiry date.

Two numbers moved and nothing else did

OpenAI posted the change to its developer forum on August 21 under a title that gives away the framing: a 20% reduction for API, Codex credits and ChatGPT Work. The body is shorter than the title deserves. Input goes from $5.00 to $4.00 per million tokens, cached input from $0.50 to $0.40, output from $30.00 to $20.00. Pro, Plus and Business subscription usage is explicitly excluded, so this is a developer-side move rather than a consumer one.

Sol has a lot of cells on OpenAI's rate card, and the interesting thing is how few of them are independent. Every derived rate is a multiplier on the two base numbers, and OpenAI left all of the multipliers alone. Cache reads stay at 90% off. Cache writes stay at 1.25 times input. Batch and Flex stay at half. Fast mode stays at a flat 2x. Prompts over 272,000 tokens still reprice the entire request at 2x input and 1.5x output. So a twelve-cell card moved on two degrees of freedom, and you can regenerate the whole thing from $4.00 and $20.00.

Credit where it is due: every one of those twelve cells is printed on OpenAI's pricing page, including the Fast long-context pair at $16.00 and $60.00, which is the dearest thing it sells and the one we expected to have to derive. The pre-cut version of that cell, $20.00 and $90.00, was printed in the same table structure, and we wrote about it on August 9 when OpenAI first let the two rules combine. Publishing the stacked cells rather than leaving readers to multiply is a small thing that several other vendors still do not do.

One more thing did not move: Terra, Luna and Cyber. We checked all three on the pricing page and against Azure's live meters, and they sit where the July 30 repricing left them, at $2.00 and $12.00, $0.20 and $1.20, and $12.50 and $75.00. If you see this week's cut described as a family-wide reduction, it was not one.

0.800000, on every line, against Opus 5

Anthropic sells Claude Opus 5 at $5.00 input, $0.50 cache read, $25.00 output, and $2.50 and $12.50 on batch. Put the new Sol card beside it and every comparable ratio is the same number. Four over five is 0.8. Forty hundredths over fifty is 0.8. Twenty over twenty-five is 0.8. Two over two-fifty and ten over twelve-fifty are both 0.8.

That uniformity makes the comparison independent of your token mix, with one large asterisk we will come back to. Normally "which model is cheaper" depends on how much you cache and how much you generate, and the honest answer is a table rather than a sentence. Here the table is boring on purpose. We priced four workloads with very different shapes and Sol comes out at exactly 0.8000 times Opus 5 on all four, because a common factor applied to every line cancels out of every possible weighting.

The asterisk is that this holds per token and the two models do not count tokens the same way. Anthropic's own pricing page warns that Claude 4.7 and later, which includes Opus 5, use a newer tokenizer emitting roughly 30% more tokens for the same text, and beyond tokenization the two models simply spend different amounts of effort on the same job. The table below feeds both cards identical token counts, which is the right way to isolate what the price change did and the wrong way to predict your invoice. We put numbers on the difference further down.

The four mixes are, in order: 30M fresh input and 10M output; 20M fresh, 180M cached and 15M output; 100M fresh, 20M cached and 4M output; and 8M fresh, 12M cached and 40M output. They are ours, not anyone's published telemetry, and they exist to show that the ratio does not care which one you pick.

Look at the last column, though, because that is where the shape changes. Before August 21 Sol matched Opus 5 exactly on input and was 1.2 times dearer on output, so Opus 5 won every workload that generated a single token. The cut does not just narrow that, it inverts it. And since the two cards diverge in shape rather than by a constant once the promotion ends, the November ratio is no longer mix-independent: it runs from 1.033x on the output-light job to 1.191x on the output-heavy one. Uniform today, uneven in three months.

The date in the small print

OpenAI's model page carries one sentence that most of the coverage dropped: Sol's promotional pricing is available at least through November 21, 2026. Three months from the announcement. The forum post frames it the same way, "for the next 3 months", so this is not a stray line on a docs page that contradicts the press release.

We want to be fair about what that sentence does and does not say. "At least" is a floor, not a ceiling. OpenAI has left itself room to extend, and given that the July 30 cuts to Terra and Luna came with no expiry at all and are still in force, extending is a real possibility rather than a courtesy. What OpenAI has not done is commit to it, and there is no published rate for November 22. Every number in the right-hand column of the table above assumes the simplest thing, which is that the promotion lapses and the standing card returns.

On why now, there are two answers in circulation and only one of them has a source. OpenAI's is efficiency: the entire statement is "as we continue to push the frontier of capabilities while improving efficiency, we're dropping API and credit pricing of GPT-5.6 Sol by over 20% for the next 3 months", which is the same story it told on July 30, when it credited Sol with rewriting its own serving kernels for a 20% reduction in cost to serve. The competitive reading, that this is about Anthropic and cheap Chinese models, comes from the Reuters wire that nearly every outlet ran. It is a plausible reading and we would probably make it ourselves, but it is worth knowing that the wire story contains no quotations at all, from OpenAI, Anthropic, an analyst or a developer. The framing is the reporter's. Nobody on the record has connected this cut to a competitor.

This matters more than a typical promo because of what it is being used to justify. A 20% saving is the kind of number that moves a migration decision, and migrations do not reverse in a day. If you switch a production workload from Opus 5 to Sol this week on the strength of the new card, you are betting on an extension nobody has promised, and the downside is not returning to parity. It is landing somewhere between 3% and 19% worse off than if you had stayed, depending on how much your job writes. We have added the November 21 date to Sol's entry on our pricing page so the reversion shows up next to the rate rather than in a footnote, the same way we handle the Gemini Flash cards that double on January 1.

Eight channels, six prices, 48 hours after one cut

A price cut is a single event at OpenAI and a queue of separate migrations everywhere else, and the interesting reporting is in the lag. We checked every channel that publishes a Sol rate on August 23 and found six distinct prices across eight places you can buy the identical model, spanning 2.75x on input and 3.30x on output.

AWS deserves credit here. Bedrock moved, and moved properly: the Global cross-region profile is at $4.00 and $20.00, matching OpenAI to the cent on input, cached input and output, and the In-Region and Geo rows fell to $4.40 and $22.00, which preserves the 1.1000x multiplier we measured on August 18 exactly. The card's short and long context tables already existed before the cut and both were repriced, so Bedrock Global long context is $8.00 and $30.00, again to the cent. That is a faithful pass-through of a competitor's price change, and it is not the norm. We should be precise about the speed, though: the model card carries no revision date, so all we can say is that it read $5.00 and $30.00 when we looked on August 18 and reads $4.00 and $20.00 today. Somewhere inside that window, not necessarily inside two days.

Azure is the other end. Every Sol meter in Microsoft's public retail price API still carries an effective date of July 1, 2026, and Global Standard still reads $5.00 and $30.00. Azure has not raised anything; it simply stood still while the thing it resells got cheaper, which converts a channel that matched OpenAI into one charging 1.25 times its input rate and 1.50 times its output. On long context the gap is the same shape, $10.00 and $45.00 against OpenAI's and Bedrock's $8.00 and $30.00. Whether that is a deliberate margin decision or a slow meter update, we cannot tell from the outside, and Azure has published nothing either way.

Two smaller notes from the sweep. OpenRouter picked up the new $4.00 and $20.00 list and then stacked its own 50% discount on top, landing at $2.00 and $10.00, and Vercel's gateway shows the same pair; treat that as the least durable number on the board, because it is a promotion on a promotion and neither has a published end date. That router discount is also older than the cut and worth untangling, because it caused a false alarm: a Hacker News thread on August 18 went up titled "GPT-5.6 Sol Pricing Cut by 50%" and was about OpenRouter's non-BYOK promo landing at $2.50 and $15.00 against the then-current list, not about OpenAI moving at all. Sol was unchanged that day. Three days later OpenAI did move, the router's 50% re-based onto the new list, and the two events are still being conflated. And OpenRouter's own comparison table still prices Bedrock at $5.50 and $33.00, the entire pre-cut In-Region card, which Bedrock's own card now contradicts. Against Bedrock In-Region that overstates input by 25% and output by 50%; against the Global row a buyer would more likely use, 37.5% and 65%. Its Azure rows, meanwhile, are correct, because Azure genuinely has not moved. So on one page two rows look equally stale and only one of them is, which is the argument for checking the vendor rather than the aggregator.

Two of our own posts went stale this week

On August 17 we published a piece arguing that the Terminal-Bench 3.0 leaderboard is a dated receipt rather than a live quote, because two of its three Codex rows were billed at a card OpenAI retired on July 30. The one row we singled out as still trustworthy was Sol's: its published $3,950.90 reproduced exactly at the rates you would pay that morning. Five days later it does not. All three Codex rows are now priced at cards that no longer exist, and the post that warned about stale benchmark costs acquired a stale benchmark cost of its own. We enjoyed that less than it deserves.

Repricing it: the drop is bounded between 20.00% and 23.80% without any further assumptions, because cache reads were 71.5% of that bill and fall by exactly 20%, while the remaining 28.5% falls by somewhere between 20% and 33.3% depending on its input and output split. Reconstructing the split from the published total and the cache share puts the run at about $3,068.73, a 22.3% fall, and cost per solved task at roughly $23.97 against the published $30.86. Opus 5 does not move, so the gap between the top of that board and the second place row widens from 1.19x to about 1.54x.

The second casualty is the August 18 Bedrock piece, and that gap widened rather than merely aged. Daybreak Blue is Sol with safeguards recalibrated for defensive security work, and its card has no Global or Geo row, so the cheapest thing a team on it can buy is an In-Region rate. That post was careful about why, and we will keep being careful: the missing Global row tracks the endpoint rather than the approval. Cross-region inference is a bedrock-runtime feature, every OpenAI model reachable only through bedrock-mantle lacks a Global row whether it is gated or not, and the ungated GPT-5.5 is what proves the gating is not the cause.

What changed this week is the size of the consequence. Daybreak Blue did not follow the cut. It sits at $5.50 and $33.00 while ordinary Sol Global fell to $4.00 and $20.00, so a team on the cyber-approved card now pays 1.375x on input and 1.65x on output against 1.10x on both a week ago. The cause is still an endpoint quirk and the effect still lands on the same people, and it has gone from a rounding error to a third more on input and two thirds more on output. It is now the only place on Bedrock where the old $5.50 and $33.00 pairing survives.

What the cut does inside OpenAI's own lineup

Moving the flagship without moving anything below it compresses the family, and Terra is the tier that loses. Sol used to be 2.5x Terra on both lines, which is a clean story: near-flagship quality at 40% of the price. Sol is now 2.0x Terra on input and 1.667x on output. On a three-to-one input-to-output blend Terra was 40% of Sol's cost and is now 56%. Terra did not get worse and did not get dearer, but the argument for choosing it is materially weaker than it was on Thursday, and the tier was already the awkward middle of a three-model lineup.

At the other end, the 25x in-family spread we wrote about on July 31 has closed to 20x on input and 16.7x on output. And Cyber, which nobody outside the Daybreak Red program can buy anyway, drifts further out of reach: it was 2.5x Sol on both lines and is now 3.125x on input and 3.75x on output, at $12.50 and $75.00. OpenAI now sells one model family whose cheapest and dearest output rates differ by 62.5x.

The rate card says 0.8x. The work says 0.39x.

Everything above this line is arithmetic on two rate cards, and rate cards price tokens rather than jobs. Artificial Analysis runs both models through the same nine-eval suite and publishes what each task cost, which is the number that survives contact with an invoice. Its figures say something the rate cards cannot: Claude Opus 5 emits 40,249 output tokens per index task against Sol's 16,879, so it spends 2.38 times the tokens getting to an answer.

That was already the dominant term before this week. On the old card Opus 5 was cheaper per output token than Sol, $25.00 against $30.00, and still cost 1.90 times as much per finished task, because verbosity beat the rate by a wide margin. The cut does not create the gap, it widens an existing one. Recomputed at $4.00 and $20.00, Sol comes to about $0.92 a task against Opus 5's $2.34, a ratio of 2.55x. Put differently, the rate card moved Sol from 1.0x to 0.8x of Opus 5 while the measured work has it at roughly 0.39x.

Two honest caveats sit under that table. Opus 5 scores higher, 63.05 against 60.93, so the comparison is 2.55x the money for 2.12 index points and whether that is worth it is a judgement rather than a calculation. And the third row is ours: Artificial Analysis has not repriced Sol. Its published figures still run on $5.00, $0.50 and $30.00, which you can confirm without taking our word for it by dividing its published Sol answer cost by its published Sol answer tokens and landing on exactly $30.00 per million. So the most widely cited cost-per-task numbers for this model are, as of today, pre-cut. Add one more name to the list of things that have not caught up.

Check the door before you check the date

Check which door you are buying through before you celebrate. Three of the eight channels above are at or below OpenAI's new rate and five are above it, and if you are on Azure the announcement has changed your bill by nothing at all. That is a strange sentence to write about a price cut, but it is the single most likely way a reader of this post is affected. Of the coverage we read, all of which traces back to one Reuters wire story, none of it checks what the resellers are charging.

Then decide whether you are pricing a quarter or a year. For batch work, evaluation runs and anything you can finish before November, $2.00 and $10.00 on Flex is a genuinely good rate for a frontier model and the expiry does not reach you. For a production migration off Opus 5, the honest comparison is not today's 0.8x but the blend of three months at 0.8x and the rest of the year at somewhere around 1.1x, which for most mixes is not a saving at all. Our cost calculator takes a fresh, cached and output split, and the comparison tool will put Sol and Opus 5 side by side on whatever numbers your own logs give you. Those are the numbers that matter, because the ratio being exactly 0.8 today is a fact about two rate cards and not about your workload.

Where every number came from, and the three pages still on the old card

• OpenAI: API pricing - Read August 23, 2026. GPT-5.6 Sol at $4.00 input, $0.40 cached and $20.00 output on short context, $8.00 and $30.00 above 272K, Fast at $8.00 and $40.00, Batch and Flex at half. Terra at $2.00 and $12.00 and Luna at $0.20 and $1.20, both unchanged since July 30, and Cyber at $12.50 and $75.00. This is also the page carrying the sentence about promotional pricing running at least through November 21, 2026. Note that openai.com/api/pricing returns 403 to automated fetches; the developers subdomain is the one to cite
• OpenAI: 20% price reduction for GPT-5.6 Sol - The announcement itself, read August 23. Effective August 21, framed as "for the next 3 months", covering the pay-as-you-go API plus eligible ChatGPT Work and Codex credit usage, and stating that Pro, Plus and Business subscription usage is unchanged. It gives all three moved rates but does not restate any of the derived cells, which is why the twelve-row table above is built from the pricing page rather than from this post
• AWS: Bedrock model card for GPT-5.6 Sol - Read August 23. Global cross-region at $4.00, $5.00 cache write, $0.40 cache read and $20.00 output; In-Region and Geo at $4.40, $5.50, $0.44 and $22.00; and a separate long-context table at exactly 2x input and 1.5x output of each. Standard tier only, with no Priority or Flex. The consumer-facing aws.amazon.com/bedrock/pricing page renders its tables in JavaScript and returns no numbers to a fetch, so the docs model cards are the only citable Bedrock source
• AWS: Bedrock model card for Daybreak Blue GPT-5.6 Sol - Read August 23. Still $5.50 input, $6.875 cache write, $0.55 cache read and $33.00 output, In-Region only, us-east-2 only, reachable only through bedrock-mantle. This is the evidence that the gated card did not follow the cut, and therefore that the approval penalty grew from 1.10x to 1.375x and 1.65x
• Microsoft: Azure Retail Prices API - Queried August 23 for GPT-5.6 Sol meters. All 24 return an effective start date of July 1, 2026: Global Standard at $5.00 and $30.00, Data Zone at $5.50 and $33.00, long context at $10.00 and $45.00, Priority at $10.00 and $60.00. We used the API because the Azure AI Foundry pricing page is JavaScript-rendered and yields nothing to a fetch. The unchanged effective date is the actual proof that Azure did not act, and it is stronger evidence than a price table would be
• OpenRouter: GPT-5.6 Sol - Read August 23. Shows the new $4.00 and $20.00 list with a further 50% off, landing at $2.00 and $10.00, cache read $0.20 and a documented above-272K override of $4.00 and $15.00. The same page prices Amazon Bedrock at $5.50 and $33.00, which Bedrock's own card now contradicts, and prices Azure at $5.00 and $30.00, which Azure's meters confirm. So one of the two stale-looking rows is stale and the other is correct, which is a good reminder to check the vendor rather than the aggregator
• Vercel: AI Gateway, GPT-5.6 Sol - Read August 23. $2.00 input, $0.20 cached and $10.00 output, described as mirroring provider pricing with no markup. Vercel's changelog dates its half off from August 17 to September 18, 2026, which is the only end date any of the stacked promotions on this page actually publishes, and it lands two months before OpenAI's own November 21
• Artificial Analysis: model comparisons - Read August 23. Intelligence Index v4.1.1 across nine evals: Opus 5 at 63.05 and $2.3369 a task on 40,249 output tokens, Sol at 60.93 and $1.2312 on 16,879. Its cost figures still run on Sol's pre-cut card, which you can verify from the page's own numbers by dividing its published Sol answer cost by its published Sol answer tokens and landing on exactly $30.00 per million. The repriced $0.92 is ours, not theirs
• Anthropic: model pricing - Claude Opus 5 at $5.00 input, $0.50 cache read, $25.00 output and $2.50 and $12.50 on batch, unchanged. Every 0.8 in this post is Sol's new card divided by one of these figures
• TokenCost: Bedrock sells GPT-5.6 Sol at OpenAI's list price and also at 1.10x it - Our August 18 post, and the source of the pre-cut Bedrock card that this week's numbers are measured against. Its finding that the Daybreak Blue penalty was 1.10x is what makes the new 1.375x and 1.65x figures a change rather than a level
• TokenCost: Terminal-Bench 3.0 cost per solved task - Our August 17 post, now partly overtaken. It is the source of the $3,950.90 Sol run, the $30.86 per solved task, the 71.5% cache-read share used to bound the repricing, and the $36.83 Opus 5 figure that does not move
• Four limits, stated plainly. The 0.8000 ratio is a per-token fact about two rate cards and does not survive translation into cost per job: Anthropic says Claude 4.7 and later emit roughly 30% more tokens for the same text, and Artificial Analysis measures Opus 5 spending 2.38 times Sol's output tokens per task, so the four workload rows deliberately feed both cards identical token counts and that is an assumption rather than a measurement. The repriced Terminal-Bench figure of about $3,068.73 rests on a token split reconstructed from a rounded 5.76B total and a rounded 71.5% cache share, so treat the 20.00% to 23.80% bound as the solid claim and the point estimate as an estimate. Our recomputed $0.92 per Artificial Analysis task applies the new card to AA's published cost components and assumes cache writes at 1.25x input, which AA does not itself break out. And every November 22 number assumes the promotion simply lapses back to $5.00 and $30.00, which is the plain reading of "at least through November 21" but is not something OpenAI has published a rate for