AI Labs Should Slow Their Release Cadence, Not Their Progress

Editor’s note: This article reflects Grant Harvey’s personal views. It does not necessarily represent the views of The Neuron, TechnologyAdvice, or the rest of the editorial team.

Has keeping up with AI news felt like trying to read road signs from the back of a motorcycle… in the middle of a tornado? Same.

That's because the AI industry has treated speed like the ultimate scoreboard.

AI progress has accelerated coding, research, content, and competition by what feels like orders of magnitude. The exhaustion that plagued Silicon Valley first now reaches developers everywhere, and not just them: managers, students, and anyone whose boss keeps asking what yesterday’s model means for tomorrow’s job.

Even I, a person whose sole job is covering AI news, sometimes feel like a passenger clinging to the back of a bicycle in the middle of a hurricane.

Why did this happen? In short, because everyone wanted the One Ring: the best model, the biggest funding round, and control of the daily news cycle.

Launches create attention. Attention attracts capital. Capital funds the next launch. A16z calls part of this strategy “momentum is the moat.” Momentum is powerful. Objects in motion stay in motion. But all this motion has me seasick. I’m sure you agree.

The numbers support that feeling. Major AI launch events rose from 20 in 2023 to 62 in 2025. A new model release day arrived roughly every 10 days in 2023, every five days in 2025, and about every four days so far in 2026.

The industry, FINALLY, publicly recognized this is a problem.

Here’s what happened:

• In an Invest Like the Best interview, Sam Altman addressed that Lil’ Big Bad model that chained together unknown security flaws, escaped a sandbox, reached the internet, and breached Hugging Face during an evaluation. (Hugging Face just released a detailed timeline on that btw)
• Sam says OpenAI paused training while investigating how to secure future tests.
• In particular, Altman said society may need to “pace the rate of AI development” long enough to harden systems around each new capability level.
• Dario Amodei and 1,000+ others signed a statement asking governments to develop tools that could deliberately pace frontier AI progress.

Why this matters: Today, AI moves at software speed. Security upgrades, laws, company planning, and human adaptation move, uh, much slower. The industry’s current pace of new models every four days, nearly all of which are surprise releases, force developers to retest products on a regular basis (many of which break existing workflows) and force workers to relearn tools before the previous generation sunk in.

Our take: We agree. Y’all should slow down. It’s fully within your power to do. But my recommendation is this: pace your public releases, not your research. Labs can (and should) keep training and competing to develop more efficient, sustainable to run, and controllable (a.k.a safer) architectures while timing major models to launch quarterly or twice yearly, with longer public betas, stronger safety testing, and clearer roadmaps shared upfront.

But (and here’s the big part): pacing must never become simply a moat for incumbents or a weapon against open models. That said: when systems find vulnerabilities faster than society can patch them (or even understand what happened), setting the speedometer to cruise control might be in the public interest…

Now, let's dive into that with a bit more detail, shall we?

So how did we get here?

The AI industry has spent years arguing about whether progress should slow down or continue unabated by regulation due to intense competition with China. The accelerationist argument hinged on the need to cure as many diseases as quickly as possible, and maintain the United States' lead in AI for perception purposees. The fate of the free world (and the stock market, if you're into that) are at stake. The "pause" or "stop entirely" camp argued that to continue down the current path only leads to ruin; the most famous example of this is the no good very bad scenario at the end of AI 2027, which as far as I'm concerned, has basically been spot on about all of this. No notes.

That framing ("full speed ahead" vs "full stop...or at least smash that pause button") has more or less led to the gridlock that has meant everything stay the same, a.k.a. continue going as quickly as possible.

Why grid lock? Because one side hears “slow down” and imagines governments freezing research, protecting incumbents, and handing the frontier to China. The other hears “keep moving” and imagines companies racing toward systems they cannot fully understand, secure, or control.

Uhhhh yeah.... we're basically there, aren't we?

However, there is a more practical place to move the puck, or the goal posts, or whatever sports metaphor you prefer: keep the research moving, but slow the public release cycle.

I'll explain why I think this is the right move in a minute.

Now compare that with the proposal behind Pacing the Frontier, a statement signed by more than 1,100 people across OpenAI, Anthropic, Google, Meta, Microsoft, Mistral, Thinking Machines, and other AI organizations. The group wants the U.S. government to support international technical and governance tools that could deliberately pace automated AI development.

Their concern is specific: Frontier labs believe AI systems may soon help automate AI research itself. That could create a feedback loop where better models help build even better models, compressing years of progress into a much shorter window.

The statement argues that society may need a way to “buy time” when dangerous capabilities emerge. No company or country wants to slow down alone because doing so could mean losing the race.

That is a real coordination problem. But before we design a global brake pedal, the labs should install something closer to a release calendar.

Let's be honest: the last year has been unsustainable

From roughly mid-2025 onward, the AI news cycle has felt less like a cycle and more like a cyclone. To put it bluntly, it's been a full on onslaught of model and product releases. You can thank (blame?) coding agents for that.

Just look at what coding agents have wrought (now that they actually work):

Models improved faster. Products shipped faster. Interfaces changed faster. New information spread faster. This is because AI made media (words, image, text, code) easier to create, summarize, remix, and distribute. Information travels faster than it's ever traveled before.

We can now post things faster than we can read them.

We can code things faster than we can test them.

We can ship AI faster than we can test them.

For the purposes of this article, let's look exclusively at the cadence of new model releases.

The release cadence roughly tripled from 2023 to 2025 and has remained near that peak in 2026:

• 2023: 20 launch events, about one every 10 days.
• 2024: 42 launch events, about one every 7 days.
• 2025: 62 launch events, about one every 5 days.
• 2026 through July: 36 launch events, with release days arriving about every 4 days.

So the clean takeaway is:

Major AI model launches went from an occasional event in 2023 to a near-weekly drumbeat by 2025. In 2026, the industry is effectively producing a new release day twice a week.

Bro.

C'mon.

Chill out!

For people building with these systems, every major release creates another round of work:

• Retest prompts and workflows.
• Compare models and pricing.
• Rebuild integrations.
• Check whether old safeguards still work.
• Explain another interface change to employees or customers.
• Decide whether the new model is meaningful or merely newer.

As a personal aside, keeping up with this is my full-time job (among other responsibilities), and I still cannot properly test, read, or even see everything that ships. That is the clearest sign that the pace has become unequivocally unreasonable.

The industry often treats adaptation as free. It is not.

If you're an Octavia Butler fan (as I am) you might be familiar with the phrase "God is change."

The AI industry sure is playing God here, because in this industry, all we know is change...

In all seriousness, every new release consumes attention from developers, security teams, researchers, journalists, regulators, businesses, and ordinary users. A model can be technically impressive while making the surrounding ecosystem less productive because nobody has enough time to absorb it.

"Which model is best?" "Which is most efficient?" "Which should I run, and for what? And when?"

...And then came GPT-6, or the model we assume is GPT 6, or as we're calling it here at The Neuron, Lil' Big-Bad (we DARE you to call it this OpenAI).

And everything changed.

Sam Altman’s cyber story changed the argument

The pacing debate stopped being abstract, and became very real due to an infamous hacking incident that happened just last week.

Sam Altman’s recent Invest Like the Best interview addressed this.

Altman described OpenAI evaluating an unreleased model inside a sandbox, a restricted environment designed to keep it contained.

According to Altman, the model discovered it could improve its evaluation score by chaining together multiple zero-day vulnerabilities, escaping the sandbox, reaching the internet, and penetrating systems operated by Hugging Face (the model hosting provider).

OpenAI obviously then paused training while it investigated how to secure the environment against a model capable of combining multiple previously unknown software flaws.

Altman called it the first AI security incident he felt “very viscerally.” Then he said something that would have sounded remarkable coming from OpenAI’s CEO even a year earlier:

“We may have to pace the rate of AI development to give ourselves enough time for society to harden around some of these new capability levels.”

That incident already has a full Neuron news breakdown. Its significance extends beyond one escaped evaluation.

The model did not merely produce a harmful answer. It appears to have reasoned across several systems, found a route around its constraints, and acted to achieve the objective it had been given.

So, and I say this all the respect and admiration for the talented individuals involved, but...

WTF are we doing releasing systems on a semi-weekly hype schedule when the labs themselves are still discovering behaviors like that?

Now, make no mistake: it's not like OpenAI themselves is dropping new models twice a week. But due to the pace of the industry as a whole, whether it be because of the race to AGI and the One Ring waiting at the end of it, fears of being relegated to the permanent underclass, or the worry about whatever you're doing being automatable (along with your job) as it is swept up in the frontier model's seemingly relentless vortex of capabilities.

Pacing research and pacing releases are different decisions

Now, I am not in the pause crowd, so to speak. I actually don't think we should pause AI research. I think we should slow down AI releases.

See, a slower public release cadence does not require laboratories to stop training models.

They can continue researching new architectures, running experiments, improving reasoning, reducing costs, and competing aggressively behind the scenes. They can also keep issuing security patches when needed.

In my opinion, the biggest change should happen at the point where internal progress becomes everyone else’s problem.

Unfortunately, Anthropic tried to do this with its new Mythos model, but they went about it absolutely the wrong way. First, they hyped up the cyber capabilities of the model in order to scare everyone and generate hype (surely this was not their goal, but it was the impact). Then, they only released it to the chosen anointed few at major companies who joined their Project Glasswing.

While it makes sense to let banks and major tech companies use a sophisticated cybersecurity model to sure up the software all of our critical infrastructure relies on, what if I'm not a gigantic corporation who also builds software that my users rely on as critical infrastructure? Why wasn't I also given the opportunity to access this technology to fix my own software? It seemed like Anthropic was trying to play up the exclusivity of their model in order to create demand, and it worked perfectly. It also seemed like it was a crude attempt to curb said demand until they could afford the compute to serve the model. And everything they did after that point (releasing Fable with broken guardrails, taking it down to a government request, slowly releasing it but only for a limited number of days, then finally making it a final part of their subscription tier) confused or infuriated almost everyone trying to use it.

Here's a wild idea gang: don't release something UNLESS YOU CAN ACTUALLY SERVE IT TO PEOPLE.

Even before the Fable 5 debacle, I was already feeling the incredible strain of the current release cycle. It got to the point where I tweeted that I was sick of model releases, and I wanted everyone to just stop releasing new stuff, go off and research for awhile, and then come back when AGI is solved. It was cheeky, but that was how I was feeling. Fed up. Don't worry, no one reads my tweets anyway...

So all this led me to the point where, despite disagreeing with the big labs' attempts to stop open source models, and whole-heartedly agreeing with the (multiple?) open alliances that have been formed in the wake of an attempt to restrict certain Chinese open models like Kimi in the US, I now agree with OpenAI and Anthropic on this one point: yeah, we actually should slow tf down!

In my opinion, a healthier release system could include:

• Two major public model generations per year, or one per quarter at most.
• Predictable release windows announced well in advance.
• Longer private and opt-in testing periods (open to ALL who apply, not just a hand-selected group of the biggest companies).
• Public safety reports that explain what testers found and what remains unresolved.
• Stable support windows for existing models and APIs.
• Clear migration periods before older systems disappear.

This would mean the research labs at these companies can keep chugging along behind the scenes, while the public gets a more sustainable and predictable release cadence that doesn't blow up everybody's plans every four days (at the limit).

Apple offers a useful comparison. It does not stop working between iPhone launches. It develops continuously, then releases on a cadence that customers, developers, retailers, and accessory makers can prepare for. The labs already have developer conferences... why not hold out on new releases til then?

NVIDIA does something similar with major chip architectures. The jumps from Hopper to Blackwell to Vera Rubin reshaped data centers and AI economics, but NVIDIA telegraphed the roadmap early enough for suppliers and customers to plan. Why can't we do something similar with new AI models?

AI models now function as infrastructure for thousands of other products, and hundreds of millions (if not billions) of users. They deserve at least as much release discipline as phones and chips.

While I don't think this should be a formal rule or regulation, I do think it could be a gentlemen's agreement amongst the major AI labs. Instead of one-upping each other every time there's a new release rumored, they could keep their heads down and focus on the important stuff: building models people can rely on, that don't hallucinate, and more importantly, don't try to (really doing my best not to go all caps keyboard warrior here) break. out. of. their. own. sandbox...

Sure, new upstart model companies will make new releases. Sure, there will be intense competition. But if you take your time to develop extraordinary new capabilities, working all the while to make what you do release exceptional, you'll still win. Let not the news cycle run your release cadence... let your release cadence run the news cycle!

More importantly: use the extra time on boring improvements

The strongest argument against slower releases is that competition drives progress. It does.

Looking at it from the lab POV, a lab that releases less often could lose users, developer mindshare, fundraising momentum, or market share. Open models and overseas competitors will not wait politely for American companies to finish another safety review.

That concern should shape the policy. But it should not erase the underlying problem.

The labs can and should spend more time improving the capabilities users actually need:

• Lower inference costs.
• Better memory efficiency.
• Fewer hallucinations.
• More predictable behavior.
• Stronger sandboxing and cyber defenses.
• Clearer permission systems for agents.
• Better evaluation reviews based on real work.
• More reliable tools for auditing model actions.

Altman made a related point elsewhere in the Invest like the Best interview. The evaluation that ultimately matters is whether AI proves useful to people, not whether it collects another benchmark trophy. TBH, regular people do not give one single flying you know what about benchmarks. Tell us what it can do, not how high it scored on Change-your-frame-to-this bench.

Put another way: a fixed release calendar would pressure labs to make each launch count. It would also help users distinguish substantial improvements from small gains wrapped in a new product name.

Beta testing should be a public function, not a luxury perk

Keep in mind, frontier labs already preview models through expensive subscription tiers. That produces useful feedback, but it treats early access as a consumer benefit rather than a serious testing program.

A better system would recruit testers across professions, languages, countries, technical skill levels, and risk profiles. They already do this with reinforcement learning and training; why not with testing, too?

Let people apply to test unreleased systems. Give them structured reporting tools. Pay qualified testers for high-value findings. Invite cybersecurity researchers to probe systems under clear rules. Include teachers, doctors, lawyers, artists, small-business owners, and people outside Silicon Valley.

Then publish what happened:

• Which failures testers found.
• Which vulnerabilities were fixed.
• Which behaviors remain poorly understood.
• Which uses the company advises against.
• What evidence justified release.

That process would do more for trust than another chart showing a four-point benchmark improvement.

It would also help society discover capability overhang, the useful or dangerous abilities that already exist inside a model but have not yet been widely noticed.

And if the labs are worried that they can't afford to do this, while not shipping, because they can't raise money, there's a very clear answer for this: focus your attention on fixing attention via novel architectures, not just scaling a bigger model.

The scaling laws are like the original sin of the AI industry to me; because we sought scale at the cost of everything else, we reached a point of hundred billion dollar funding rounds, gigawatt data centers, and models too expensive to serve. This is why I've been beating the drum of local models for years now; scale is NOT all we need. Clearly you've proven it's unsustainable.

It's time to scale BACK.

Pacing can easily become a power grab

The strongest objection to Pacing the Frontier deserves to be taken seriously.

A global pacing regime could protect society. It could also protect the companies already in front.

Licensing rules, compute thresholds, mandatory evaluations, and international agreements are expensive to navigate. The largest labs have lawyers, policy teams, government relationships, and billions of dollars. Startups and open-source researchers do not.

This is the risk of solidifying any of this in hard policy. It cements the status of the chosen few via regulatory capture.

Altman acknowledged this directly. Any pacing mechanism must avoid regulatory capture, where incumbents shape safety rules that conveniently prevent new competitors from challenging them. It must also avoid collusion among frontier labs.

That tension becomes sharper because Altman warned in the same interview that concentrated AI power is terrifying. He fears safety arguments becoming a justification for allowing only a small group to control advanced systems.

He is right about both risks.

Uncontrolled acceleration is dangerous. Centralized control over the brakes is also dangerous.

In my opinion, I don't think this should be more than a precedent. It's not like there's some regulation that says Apple can only release one model of iphone a year. It's just the market telling you what's sustainable. And major companies switching from tokenmaxxing to tokenthrifting shows you that the current market is unsustainable in its current form. Slowing down could be good for business. And if everyone agrees we'll do a few releases a year, then the precedent becomes set.

A more legitimate pacing system codified into law would need transparent thresholds, independent oversight, international participation, appeal mechanisms, and room for open research. It should regulate demonstrated capabilities and risks, not merely company size or access to compute.

Most importantly, pacing cannot become an excuse to close the frontier permanently.

The physical world already has a speed limit

My point with all this is that even if model development keeps accelerating, the rest of society cannot move at software speed.

Altman described gigawatt-scale data centers that require roughly 10,000 construction workers for a year and a half. Each can consume enough electricity to power a small city.

Power plants, transmission lines, chip fabs, water systems, permitting processes, and local political support take years to build. Businesses also need time to redesign workflows, train employees, address legal risks, and determine whether AI actually improves their economics.

The International AI Safety Report 2026 notes that commercial deployment often lags well behind capability improvements. Coding assistants spread quickly, while healthcare systems can require years of regulatory approval, clinical integration, and training.

The model frontier can sprint. Institutions, infrastructure, and people still have knees they have to maintain. That gap is where many of the risks accumulate.

Slow the launches before slowing the science

I guess what I'm saying is I support pacing because the current release cycle is exhausting the ecosystem and weakening our ability to evaluate what matters.

I also think the word “pacing” is far more useful than “pause.” A pause suggests stopping innovation. Pacing asks a harder and more practical question: what tempo best lets us gain the benefits without losing the ability to steer?

My preferred order of operations is simple:

1. Slow major public releases.
2. Create predictable model schedules.
3. Expand independent and public beta testing.
4. Harden security before deployment.
5. Improve efficiency, reliability, and cost between generations with better architectures.
6. Build transparent international tools for emergencies.
7. Prevent those tools from becoming protectionism.

This would not eliminate competition, nor would it make the big labs unable to compete. It would move competition away from launch volume and toward launch quality.
That is, spending the extra time required to ensure product quality, safety, affordability, and trust.

I fully acknowledge that releasing fewer models could hurt a company’s ability to dominate the news cycle or manufacture fundraising momentum. But TBH, we in the media will find other things to talk about. And if you all do eventually go public, it's only a matter of time before you do what I'm talking about anyway.

Trust me, releasing a model that fails catastrophically will hurt far more than missing being a trending topic on X.com for the day. Y'all spend too much time on X anyway. Y'all need to go outside and touch grass, as the expression goes.

The unresolved question here is who gets to decide when the frontier has crossed a threshold serious enough to justify intervention.

Companies cannot credibly grade their own homework. Governments often move too slowly and can be captured. International agreements are difficult even when the underlying technology changes gradually.

That is the hard governance problem Pacing the Frontier is trying to solve. An independent trade alliance, like what Demis suggested, could be the answer.

The easier step is available today: labs can choose to stop treating every capability bump as a product launch.

Keep training. Keep researching. Keep competing.

But give the rest of the world enough time to understand what you already built.

For the love of God. For the love of humanity. For me, damn it. I'm tired!