The model drawer

We have grown used to thinking that AI models ship when they are ready. It is the image the industry cultivates of itself: labs at the limit, researchers who don’t sleep, a race so tight that every model goes out the instant it stops being embarrassing. Nobody holds anything back. There is no time to hold anything back.

On Tuesday, September 22nd, at 16:31 UTC, Anthropic released Claude Opus 5.5. At 18:00 UTC, OpenAI released GPT-6 Sol and GPT-6 Luna. Eighty-nine minutes apart, from two companies that supposedly don’t share their plans. Both led with the same argument, price. Both talked about cost per task instead of benchmark points. And both did it ten days after their chief executives called, with great solemnity, for pacing the frontier. Fortune headlined its story with a question that answered itself: What slowdown?

It could be a coincidence. Coincidence, in this industry, is remarkably punctual.

The argument of this piece is that it isn’t one, and that the interesting part is not the coincidence but what makes it possible. To counterprogram, you need something to program. To step on a rival’s announcement on the same day, at the same hour, with the same message, you need a model that is finished, evaluated, priced, with the press release written, waiting. You need a drawer.

The coincidence is already three years old

If Tuesday were a one-off, a shrug would do. It isn’t. It is a habit with a track record, and the track record starts almost the same day the race did.

On February 6th, 2023, Google announced Bard in a blog post. The next day, Microsoft had the press summoned to Redmond to unveil the new Bing, built on OpenAI’s technology. Google held its own event on the 8th, in Paris. Three days, three announcements, and a first pattern: when you can’t get there first with the product, you get there first with the statement.

On February 15th, 2024, Google spent the morning presenting Gemini 1.5 Pro and its one-million-token context window. Hours later, OpenAI showed Sora. It didn’t release it. It showed it. A few videos, a web page, no date. That was enough: by the next day nobody was talking about context windows.

On May 13th, 2024, OpenAI called an event on a few days’ notice to present GPT-4o. Google I/O, on the calendar for weeks, opened the following day. Google answered as best it could, posting a teaser of its own assistant an hour before OpenAI took the stage. That December the roles reversed, and it was Google that dropped Gemini 2.0 into the middle of the “12 Days of OpenAI”, a campaign designed precisely to occupy the entire calendar.

Since then the list has only grown denser. Claude 3.7 Sonnet and GPT-4.5 in the same week of February 2025. Gemini 2.5 Pro and GPT-4o image generation on the same March 25th. Claude Opus 4.1, the open-weight gpt-oss models and DeepMind’s Genie 3 on the same August 5th. GPT-5.1, Gemini 3 Pro and Opus 4.5 inside eleven days in November, followed by Sam Altman declaring an internal “code red”. And on February 5th of this year, Opus 4.6 and GPT-5.3-Codex, both scheduled for ten in the morning Pacific, in a week when the two companies were also fighting over Super Bowl ads.

It had already happened once this month: Fable 5.1 on the 1st, Gemini 3.8 Flash and Muse Spark 1.3 on the 2nd, GPT-6 Astra on the 3rd. Four companies, seventy-two hours. Tuesday the 22nd was not an anomaly. It was the second time in three weeks.

With a launch every four days, collisions come cheap

The serious objection to all this is statistical, and it deserves to be taken seriously before going any further.

The Neuron ran the numbers in July: 20 major model launches in 2023, 42 in 2024, 62 in 2025, and 36 in the first seven months of 2026. One every ten days, then every seven, every five, every four. At that pace, two labs landing in the same week proves nothing. It is what a saturated calendar looks like. Add that everyone aims for the same slots — Tuesday to Thursday, morning in California, clear of holidays and earnings calls — and the collisions happen on their own, with nobody spying on anybody.

It is a good argument. It explains the crowded weeks. It explains November 2025 and probably March 2025.

What it doesn’t explain is the resolution. Chance produces same-week coincidences. It does not produce two launches scheduled for the same hour, or an eighty-nine-minute gap, or an event improvised for the exact eve of a conference that had been on the calendar for months. Nor does it explain the symmetry of the message. Tuesday was not two arbitrary models colliding; it was two price cuts. Opus 5.5 arrived 40% cheaper to run than Opus 5. Sol and Luna came in 50% below GPT-5.6. Two companies do not separately discover, on the same morning, that the argument of the quarter is cost per task.

Chance fills the week. Somebody picks the hour.

To counterprogram, you need something in the drawer

This is where the metaphor starts doing work, because counterprogramming has a material requirement that almost never gets mentioned: inventory.

The days before Tuesday left it fairly exposed. On September 6th, word was already circulating that GPT-6 Sol was in internal testing. On the 13th, its identifier turned up in OpenAI’s API. On the 15th, Altman announced a big ship “this week”, boat emoji included. Nothing sailed that week. The boat left the following Tuesday, eighty-nine minutes after Anthropic’s, which is either an extraordinary coincidence or the first documented case of a ship waiting for another one to leave so it wouldn’t have to go alone.

A model whose name already answers in the API is not a model in development. It is a model in storage.

And this is not a rare case. It is how things normally work. OpenAI said it spent more than six months on GPT-4’s safety and alignment after training it. Mythos Preview was in use inside Anthropic from late February and did not reach its partners until April. The academic literature on internal deployment takes it for granted: each company’s most capable systems spend weeks or months in internal use before they ship. Koray Kavukcuoglu, of Google DeepMind, said it last Thursday without meaning to. Gemini 4 is already in post-training, they can already see the results, they are already excited, and it will come out “as soon as possible”. It is hard to describe a drawer any better.

It pays to be precise about what this shows and what it doesn’t. No lab has ever admitted to holding a finished model in order to ruin a competitor’s day. The official reason for every wait is safety, which does in this industry the job that “scheduling conflicts” does everywhere else. And often it will be true: a tally of release delays published in August found only two cases with a stated reason, and both were safety evaluations.

But there is no need to picture a warehouse of finished models gathering dust for a year. Something far more modest will do: slack. Between the day a model could ship and the day it has to ship there is a margin of days or weeks, and that margin does not belong to the safety team. It belongs to comms. The drawer isn’t deep. It only has to be deep enough to choose the day, and the day gets chosen by looking across the street.

A launch is no longer a technical milestone, it’s a slot in the schedule

Television has been doing this for decades. When the rival network premieres its big series, you air the final, the premiere or the special you had been saving for another occasion. Not to win the night. To make sure the other side doesn’t win it clean.

The mechanism in AI is the same, and it has to do with how this stuff gets bought. A launch that has a news cycle to itself produces a full week of comparisons, threads, tests and headlines with a single protagonist. Above all, it produces the moment when an engineering team says “let’s try it.” A shared launch produces something else: a table. Two columns, prices side by side, and a conclusion along the lines of “it depends on your use case”, which is the polite way of saying nobody is switching providers this week.

That is the real product of counterprogramming. Not persuading anyone that your model is better, but making sure the rival’s isn’t the only topic of conversation for seven days. Neutralizing is a lot cheaper than winning.

It has costs, and we are the ones paying them. The first is evaluation. Whoever has to decide which model to use gets two launches, two price sheets and two sets of benchmarks picked by interested parties on the same morning, and independent evaluators take days to say anything useful about either. The second is trust. If the release date is set by the competitor’s calendar, the date says nothing about the model. Not that it just became ready, not that it has been ready for a month, not that it could have used two more weeks.

The third is the most uncomfortable. The same industry that tells us it cannot slow down, that competitive pressure leaves no margin, that every week counts, shows every few months that it does have margin: it has it every time it waits for a rival. The proposal to pace the frontier is debated as though it were an unprecedented sacrifice. Pacing already happens. It’s just that the schedule is set by the competitor’s marketing department rather than by an outside evaluator.

What’s still inside

If the argument is worth anything, it should be able to predict, so here goes.

On Tuesday, September 29th, OpenAI holds its DevDay, and Altman has already said he is saving things for it. Anthropic promised on the 22nd itself that Sonnet 5.5 and Haiku 5.5 would arrive “over the coming weeks”, a window elastic enough to land exactly where it will be most annoying. My bet is that at least one of the two shows up within forty-eight hours of DevDay, before or after, and that OpenAI will not walk on stage empty-handed even though it shipped two models a week ago. If I’m wrong, it will be on the record here. If I’m right, that will be too, and it won’t have taken a frontier model to see it coming.

Then there is the most visible drawer in the industry. Google has not shipped a flagship since November 2025, skipped the Gemini 3.5 Pro it had hinted at, and has a Gemini 4 in post-training with “as soon as possible” as its only date. When it comes out, I would bet it won’t be in a quiet week. And that somebody, at some other lab, already has the blog post written for that day.

None of this turns the race into theater. The models get better, the prices really do fall, and the pressure between labs is real. But a race in which the runners jointly choose the moment to cross the line looks less like a race than like a television show. We had got used to reading every launch as news about the state of the technology. It is time to read it as what it is: a scheduling decision about something that already existed. The drawer doesn’t call into question what is inside it. It is only a reminder that the date on the label is not the date of manufacture.