On August 1 we wrote that nobody was going to bill you $20 per million input tokens. Four days later OpenAI let Fast mode accept long prompts, and now somebody will.

We got this wrong eight days ago

In our piece on what speed costs, published August 1, we wrote that Fast mode and the long-context surcharge were mutually exclusive, and finished the thought with a flat assurance that nobody was going to bill you $20 per million input tokens. Accurate that morning. Wrong by Wednesday. We have dated a correction onto that post rather than quietly editing the sentence away, because a pricing blog that silently rewrites its old numbers is worth nothing. What follows is the figure we said would not exist, and what it does to a job you are probably already running.

One sentence, no price attached

Here is the entry, tagged Update and carrying model tags for all three tiers:

"Fast mode now supports long-context requests for GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna. As of today, long-context prompts exceeding 272K tokens can run in Fast mode, delivering speeds up to 2.5x faster than the Standard tier."

Read it as a capability note and it is good news. The workload with the worst wall-clock problem, a giant agent context that takes minutes to chew through, was until Tuesday the one workload forbidden from paying for speed. That was a genuinely odd hole and closing it is right.

Read it as a pricing note and it is not a pricing note at all. The word cost does not appear. Neither does any number except 272K and 2.5x. To find out what the newly permitted request costs you have to go to the pricing page, switch the view to Fast, and read a row that did not previously describe anything you could buy.

One honest caveat about what we can and cannot establish. We can show that the capability arrived on August 5, because the changelog is dated. We cannot show when those Fast long-context cells were added to the pricing table, and we are not going to claim the price is new when what we can prove is that the request is new. Before August 5 the cells described a purchase the API would refuse.

Every price a GPT-5.6 request can carry

Two independent switches now sit on every GPT-5.6 request. Is the prompt over 272,000 input tokens, and did you ask for Fast. Each has two settings, so each tier has four prices. Everything here is published on OpenAI's pricing page; we have arranged it rather than derived it.

Notice the middle two rows of each block. Standard over 272K and Fast under 272K charge the identical input price, every time. A 300,000 token prompt on standard Sol pays $10.00 per million on input, which is the Fast rate, and receives standard speed and no latency SLA for it. We flagged that cell as the worst in the grid a week ago and it is still the worst in the grid. It has just been joined by a fourth row that is twice as expensive again.

Take the full span now. Batch on Sol is $2.50 and $15.00. Fast above the threshold is $20.00 and $90.00. Same weights, same model, an eightfold range on input and a sixfold range on output depending only on how you asked. Our August 1 piece put that span at 4x. It is 8x now, and it took four days.

The number is published. The rule is not.

We went looking for a sentence in OpenAI's documentation saying that Fast pricing and long-context pricing combine. There is not one. The word stack does not appear on the pricing page, on any of the three model pages, or in the Fast mode guide, and the service-tiers URL that might have carried such a rule returns a 404.

What OpenAI does publish is the answer itself, as a figure in a table. Sol Fast above the threshold is $20.00 and $90.00. That happens to equal the standard card multiplied by two for Fast and then by the long-context rule, which the model pages do state plainly: prompts with over 272K input tokens are priced at 2x input and 1.5x output for the full request. Both roads reach the same place. Only one of them is signposted.

This matters more than a documentation nitpick, because the two halves live in different places and neither points at the other. The 272,000 threshold is defined only on the individual model pages. The pricing table itself labels its columns Short context and Long context and never says where the boundary falls. The Fast guide offers no number at all, just the phrase that Fast charges a per-token premium over standard. Anyone reasoning about their bill from any single one of those pages will get it wrong.

Worth noting that Anthropic, whose pricing page we criticised on a different point yesterday, writes the sentence OpenAI does not: that Fast mode pricing stacks with other pricing modifiers. It uses the actual word. It has less to stack, which we will get to, but it says so.

What one token costs at the line

Because the surcharge reprices the full request rather than the excess, the cost curve has a step in it, and the step is large. At exactly 272,000 input tokens, Sol charges $1.36 for the prompt. One token later it charges $2.72. That single token costs $1.36. In Fast mode the same token costs $2.72.

Now a realistic job instead of a knife edge. Take an agent turn with a 300,000 token context and 20,000 tokens of output on Sol. Standard processing bills $3.90. Trim the same job to 270,000 tokens of input and it bills $1.95. Those 30,000 extra tokens, about a long file, cost $1.95 between them, which prices them at $65 per million against a sticker rate of $5. Some of that is the output on the same request being repriced at 1.5x alongside the input, which is easy to forget when you are thinking about a prompt getting longer.

One more symmetry that is either elegant or annoying depending on your mood. That 300,000 token job on standard costs $3.90. The 270,000 token version of it in Fast mode also costs $3.90. Identical money, and one of them arrives up to 2.5x sooner. If your context is hovering near the line, the cheapest thing you can do is often not to slow down but to cut 30,000 tokens and buy the speed with what you saved.

Nobody decides to cross it

A threshold you opt into is a pricing tier. A threshold your tooling wanders across while you watch is something else. On July 11 a Codex user opened issue 32486, titled "Default GPT-5.6 context can cross the 272K higher-usage threshold", and it is still open four weeks later.

Their account: the CLI reports a 372,000 token window with a 95% effective-context multiplier, giving roughly 353,400 usable tokens. The billing threshold sits at 272,000. That leaves about 81,400 tokens of room above the line that the default configuration will happily use without asking. The reporter notes it is easiest to hit when Sol runs subagents to audit its own work, which is precisely the kind of long-running task somebody would want Fast mode for.

Treat those figures with the care they deserve. They are one user's reading of CLI metadata, not an OpenAI specification, and the issue carries an enhancement label with no commitment attached. But the shape of the complaint is checkable against OpenAI's own documented rule, and the arithmetic is not kind. A session that settles at 353,400 input tokens pays $3.53 per turn on Sol input. Held under the line at 272,000 it would pay $1.36. That is 2.6 times the money for 30% more context.

The workaround in the thread is one config line, model_auto_compact_token_limit = 270000, and if you run Codex on Sol we would set it today. The reporter also asks OpenAI to charge the multiplier only on tokens above the threshold instead of the whole request. That would be the humane design. It is not the one in force.

Anthropic shipped this rule, then deleted it

The interesting thing about OpenAI's 272K rule is that it is not unusual. It is the industry's standard first attempt at pricing long context, and one competitor has already been through the whole arc.

When Anthropic launched a 1M window for Claude Sonnet 4 in August 2025, prompts over 200K cost $6 and $22.50 against a $3 and $15 card. Look at the shape of that: 2x input, 1.5x output, triggered by a context threshold. It is OpenAI's current rule with a different number on the door. Anthropic ran it for about seven months and then removed it on March 13, 2026, which we covered when the surcharge came off. The pricing page now says that Claude 4.6 and later include the full 1M window at standard pricing, and spells out that a 900k-token request bills at the same per-token rate as a 9k one.

Google sits in between and is worth a look because it is the closest structural match to what OpenAI just did. Exactly one Gemini 3.x model still has context tiers, Gemini 3.1 Pro Preview, at $2 and $12 up to 200k and $4 and $18 above it. Same 2x and 1.5x again. And Google publishes its Priority tier with its own long-context columns, which is the same table-cell approach OpenAI uses for Fast. So the practice of publishing a combined price without narrating the combination is not unique to OpenAI. It just produces a bigger number there, because OpenAI's speed premium is larger.

While we are correcting things: several summaries floating around claim Claude Opus 4.6 charges $10 and $37.50 above 200K. It does not. Opus 4.6 is $5 and $25 flat with no long-context tier, per Anthropic's live pricing page and its own announcement. If you have that number in a spreadsheet, it is wrong.

Set the ceiling at 270,000

Cap the context below the line rather than at the model limit. The window and the price break are different numbers, and no tool we know of warns you when you pass the second one. 270,000 is a reasonable ceiling on all three tiers, and if a task genuinely needs 353,000 tokens then it should cross deliberately and you should see it in a log.

Assert on the pair, not on either half. A request is expensive when it is long and Fast, and a check that only looks at service_tier or only at token count will miss the combination that costs four times sticker. If you are estimating spend anywhere in your stack, the estimator needs both switches or it will under-report by up to 4x on input.

Ask whether the request needs to be long before you ask whether it needs to be fast. This is the one that actually saves money. Because the surcharge hits the entire request, trimming a prompt from just over the line to just under it is worth far more than the tokens you removed, and on a hot path that trim can fund the speed upgrade outright.

The broader point is one we keep running into. Model choice gets all the attention and the rate card underneath it now spans 8x on identical weights. Sol at Batch and Sol at Fast above 272,000 tokens are the same model, and one costs eight times the other on input. Picking the tier is often a bigger decision than picking the model, and it is made by a config flag nobody reviews.

The pages these numbers came off

Prices were read from OpenAI's, Anthropic's and Google's own pages on August 9, 2026. The arithmetic in the worked examples is ours and reproduces the published cells exactly. Nothing here comes from a pricing aggregator.

• OpenAI: API changelog - The August 5, 2026 entry quoted in full, tagged Update against all three GPT-5.6 tiers, and the July 30 rename of Priority processing to Fast mode
• OpenAI: pricing - Every cell in the twelve-row grid, including the Fast long-context rows at $20.00 and $90.00 for Sol, the Batch and Flex rows at half the standard card, and the Short context and Long context column labels that carry no threshold definition
• OpenAI: GPT-5.6 Sol model page - The only place the rule is written out: prompts with over 272K input tokens priced at 2x input and 1.5x output for the full request. The same sentence appears on the Terra and Luna pages
• OpenAI: Fast mode guide - Describes the premium in words only, with no multiplier and no mention of long context combining with it
• openai/codex issue 32486 - Opened July 11, 2026, still open, labelled enhancement. The 372,000 window, 95% effective multiplier, 353,400 usable tokens and 81,400 of exposure are the reporter's figures, as is the model_auto_compact_token_limit workaround
• Anthropic: pricing - The full 1M window at standard pricing for Claude 4.6 and later, the 900k versus 9k sentence, Opus 4.6 at $5 and $25 flat, and the statement that Fast mode pricing stacks with other pricing modifiers
• Anthropic: 1M context for Sonnet 4 (August 2025) - The retired two-tier table, $3 and $15 up to 200K against $6 and $22.50 above it, which is the 2x and 1.5x shape OpenAI uses today. Removed on March 13, 2026 per the GA announcement
• Google: Gemini API pricing - Gemini 3.1 Pro Preview as the only Gemini 3.x model with context tiers, at $2 and $12 up to 200k and $4 and $18 above, plus the Priority rows published with their own long-context columns

Two things we could not settle. We do not know when the Fast long-context cells appeared in OpenAI's table, only that the request they describe became legal on August 5, so we have avoided calling the price itself new. And OpenAI does not say anywhere whether Fast mode can be combined with the Batch API, where Anthropic states plainly that it cannot. If you have tried it, we would like to know what came back.