AGI Made Clear › Pacing the frontier
AI labs say we need to pace AI. So why is the race still so fast?
Anthropic and OpenAI have both started arguing for pacing frontier AI development. Then, on 22 September, they released new models on the same day. So what exactly does pacing mean, and who decides when the gap between capability and control is safe?
On 22 September 2026, Anthropic released Claude Opus 5.5.
Later the same day, OpenAI released GPT-6 Sol and GPT-6 Luna.
At first glance this looks like business as usual in the AI race.
New models. Higher capabilities. Lower prices. Rival laboratories releasing on each other's heels.
Except something has changed.
Both Anthropic and OpenAI have recently been arguing publicly for pacing the development of increasingly powerful AI. Anthropic describes Opus 5.5 as its first release since it called for pacing the frontier.
So how can AI companies call for pacing while continuing to ship more capable models?
The answer tells us what the AI slowdown debate is actually about. Pacing does not mean stopping. It means trying to prevent capabilities from advancing faster than our ability to understand, secure and control them.
That is a more complicated idea, and a more interesting one.
What actually happened this week
Anthropic released Claude Opus 5.5 on 22 September 2026.
Anthropic says the model is substantially better than Opus 5 and cheaper to run: input tokens drop from $5 to $4 per million, output from $25 to $20, and cache reads from $0.50 to $0.20. The company puts the overall saving at around 40% on typical workloads, which comes mostly from the cache change rather than a 40% cut to the headline prices.
Anthropic says Opus 5.5 achieved its strongest results yet on its automated behavioural audit, with improvements in preventing sandbox escapes and harmful actions. Two external organisations, METR and Frontier Design, tested the model before release.
The company also says something that is easy to skip past, and should not be: that building evaluations which reliably catch every failure before deployment "remains an unsolved problem".
OpenAI then released GPT-6 Sol and GPT-6 Luna.
Both are roughly half the price of the models they replace. Sol costs $2 per million input tokens and $10 per million output, down from $4 and $20. Luna costs $0.10 and $0.50, down from $0.20 and $1.20. Cached input reads are discounted by 90%.
So on the same day that two laboratories are arguing for pacing, capable AI got meaningfully cheaper.
That apparent contradiction is worth taking seriously rather than treating as hypocrisy.
What "pacing" actually means
The word is easy to misread.
Pacing does not mean stop building AI. It does not mean stop releasing models. It does not mean freeze capabilities at today's level.
Anthropic has described two different forms of it.
The first happens inside a company. A laboratory decides to prioritise safety over speed when the two conflict. That might mean delaying a training run, running additional evaluations, improving security, strengthening monitoring, restricting what a model can access, or holding back deployment until a particular risk is better understood.
The second happens across an industry. Here the problem is competition. If one laboratory slows down for safety reasons while its competitors do not, the cautious one simply loses ground. Anthropic calls this a race-to-the-bottom dynamic, and argues that avoiding it requires coordination that is lawful, verifiable and effective.
That second form is a much bigger ask than the first. The first is a company decision. The second needs somebody outside the company to make it stick.
OpenAI has already done it, and said so
This is not purely theoretical, which is what makes the week interesting.
On 18 August 2026, OpenAI disclosed that it had temporarily slowed part of its frontier development. It introduced "a two-week pause in reinforcement learning (RL) training on our latest models intended for deployment", and said its largest planned frontier RL run "remains on hold" while smaller evaluations were carried out.
The reasons were specific rather than general. OpenAI cited a security incident and preliminary evidence that its upcoming Astra model may meet what it calls the Critical cybersecurity capability threshold.
The principle it stated was simple: "our standards for monitoring, alignment, and security must stay ahead of those risks". And on the decision itself: "we temporarily slowed the pace of scaling".
That is pacing in practice. The laboratory did not stop developing AI. It slowed one specific part of the process while safeguards were strengthened, and it said which part and why.
So why are new models still being released?
Because the argument is not about whether development continues.
It is about which parts should slow, under what circumstances, and what evidence should be required before they speed up again.
Anthropic makes the distinction explicitly. It says it largely understands the risks presented by today's models and believes existing practices can manage them, though not perfectly. Those practices include alignment testing, external evaluation, cybersecurity safeguards, biological-risk evaluations, monitoring and capability-specific protections.
But it also warns that more serious risks could emerge quickly as models become substantially more capable.
In other words: today's model may clear today's bar. The concern is whether tomorrow's model needs a much higher one, and whether anyone will have built it in time.
Opus 5.5 is a test of the idea
Anthropic is clearly aware of how the timing looks. Its announcement does not hide the fact that this release follows directly from its own call for pacing. It argues instead that the release shows what its version of pacing looks like in practice.
External evaluation before release. The strongest alignment results it has recorded. Safeguards matched to the model's capabilities.
Whether those safeguards prove sufficient is a separate question, and not one anyone can answer today. But the position is coherent: pacing does not mean refusing to release a model that meets the current standard. It means raising the standard as capabilities rise.
The honest test of that position is not this release. It is the first release where the standard genuinely gets in the way.
The complication: AI is helping build AI
The pacing debate gets considerably more urgent when you look inside the laboratories themselves.
Anthropic has published measurements of how much AI is already contributing to its own AI research and development. As of August 2026, it says Claude "leads" 26% of measured AI R&D work, and that the share where AI at least collaborates substantially exceeds 90%.
It is also specific about the limit: "Claude is not operating fully autonomously for any measured subset of AI R&D work."
But the direction is the point. AI is increasingly helping researchers build better AI, which raises the possibility of a feedback loop: better AI helps researchers develop AI faster, which produces better AI, which accelerates the research further.
That is why pacing is becoming harder to think about. The speed of development may increasingly depend not on how fast human researchers work, but on how much of the research itself is automated.
This is where recursive self-improvement enters
Recursive self-improvement is usually presented as distant or science-fictional. The strongest version involves a system autonomously designing and building a more capable successor, which then improves itself again - our sister site sets out what that loop would actually require.
We are not there, and Anthropic says so plainly.
But parts of the process are beginning to automate, and Anthropic says one reason it publishes these measurements is to help governments and the public understand how close frontier laboratories may be moving towards more automated AI development.
That is a notable thing for a company to volunteer, and it changes what pacing has to measure. Eventually the question may not only be how powerful the latest model is, but how quickly that model is helping create the next one.
Pacing may be about the gap, not the speed
Probably the most useful way to read this debate is to stop asking how fast AI is progressing, and ask instead how large the gap is between capability and our ability to manage it.
Imagine capability improves by ten steps. If alignment, monitoring, cybersecurity and oversight improve by twelve, rapid progress may be manageable. If capability improves by ten while control and understanding improve by two, the same speed looks very different.
That appears to be the actual principle underneath most of the current argument.
The objective is not to make AI progress slowly. It is to stop safety progressing more slowly than capability.
But who decides whether the gap is safe?
This is where it gets difficult.
At the moment, most of that judgement sits with the laboratories building the models. They develop the systems. They run many of the evaluations. They decide when a model meets their deployment standard. They decide when development pauses and when it resumes.
External evaluators are increasingly involved, which adds real scrutiny. Anthropic used two of them on this release.
But the larger question is unresolved. If a frontier laboratory believes its safeguards are sufficient to keep scaling, who verifies that judgement?
The company? Independent evaluators? Governments? International bodies? Other laboratories? Some combination?
Pacing sounds straightforward until somebody has to decide when to press the accelerator again.
Competition makes it harder still
These laboratories do not develop in isolation. They compete for customers, for researchers and for investment. And AI development is increasingly discussed as international competition, particularly between the United States and China.
That creates a genuine tension rather than a rhetorical one.
Slow too little, and safety may fail to keep pace with capability. Slow too much, and another laboratory or another country continues regardless.
Which is why coordinated pacing keeps coming up. A company can slow itself. Slowing a frontier is a different problem entirely.
Meanwhile the economics keep accelerating
Here is the part that gets least attention, and may matter most.
Even if frontier training becomes more cautious, the cost of using advanced AI keeps falling. Both of this week's releases cut prices substantially.
That means the AI transition can keep accelerating without any jump in frontier intelligence at all. Existing capabilities get cheaper. More companies can afford them. More developers build with them. More agents get deployed. More tasks become economic to automate.
Pacing the frontier is not the same as pacing adoption. They are two different curves, and only one of them is being discussed.
Right now, adoption may continue speeding up even as laboratories become more careful about the largest training runs.
This changes what the slowdown debate is
It is tempting to split the discussion into two camps: one that wants development to continue, one that wants it stopped.
The real questions are closer to these.
Which capabilities require stronger safeguards? What evidence should a laboratory provide before continuing to scale? Who verifies that evidence? When should development pause? What would allow it to restart? Can competitors coordinate without simply handing an advantage to whoever defects? And what happens when AI itself starts accelerating AI research?
Those are harder than "should we slow AI down". They are also the ones that will actually be decided.
What to watch next
More transparency from laboratories. Anthropic has started publishing measurements of AI involvement in its own R&D. If others adopt comparable reporting, we get a much clearer picture of how quickly development itself is automating. If they do not, we have one company's self-report.
Independent evaluation. External organisations are increasingly testing frontier models before deployment. What matters is the extent of their access, their independence, and whether their findings can actually stop a release.
Training pauses. Temporary internal pauses may become more common at particular capability thresholds. The question is whether they stay voluntary or become part of a standard.
Government involvement. Anthropic has said public policy should play a larger role as AI becomes more capable. What that role looks like is heavily contested.
AI building AI. Watch the share of frontier research performed by AI systems. If that number rises quickly, the pace of development becomes much harder to predict, and much harder to pace.
The race has not stopped
This week makes one thing clear. The frontier race is still moving quickly.
Anthropic released a more capable Claude. OpenAI released cheaper GPT-6 models the same day. Capabilities improved. Prices fell. AI became more accessible.
None of that looks like a slowdown.
But that may be because slowdown is the wrong word. Pacing is the more precise one. Keep moving, but try to ensure that the ability to understand, monitor and control increasingly capable systems moves at least as fast.
Whether that balance survives intensifying competition is the much bigger question. We may find out sooner than expected.
What this article does not establish
It does not establish that either company's safeguards are adequate. Both describe their own testing; Anthropic states plainly that evaluations reliably catching every failure before deployment is an unsolved problem.
It does not establish that pacing is working. One disclosed pause and one release with external evaluation are evidence that the idea is being applied, not evidence that it is sufficient.
It does not establish that the two companies mean the same thing by the word. They have described related but not identical practices, and neither has proposed a shared definition.
It does not establish that AI is close to conducting its own research. Anthropic's own measurement says the opposite: not fully autonomous in any measured subset.
What it does establish is that two frontier laboratories now describe their release decisions in terms of a gap between capability and control, and have each published something concrete about how they are managing it. That is new, and it is checkable.