When the company selling you a technology and the company’s own research lab disagree about how fast that technology is improving, the disagreement is worth more than either side’s headline claim. That is the situation Anthropic finds itself in this month. Its chief executive, Dario Amodei, has published an essay arguing that artificial intelligence is on an exponential curve toward something he calls “Powerful AI”. The system cards for Anthropic’s two newest frontier models, written by Anthropic’s own researchers, say they cannot find evidence of that acceleration. For UK businesses deciding how much of their strategy to bet on imminent superintelligence, that contradiction is the most useful piece of information to come out of the lab in months.
The pitch and the paperwork
Amodei’s essay, titled “Policy on the AI exponential”, reaches for a Lord of the Rings image: political institutions move like Treebeard the Ent, too slow to keep up with the Hobbits. The Hobbits, in this telling, are AI capabilities — and they are not just fast, they are getting faster all the time. He uses the word “exponential” six times, including in the title. The central claim is direct: “AI’s scaling laws, which predict an exponential increase in general cognitive capabilities with increasing computing power, now have over a decade of empirical evidence behind them. If these scaling laws continue for only a year or two longer, we are likely to get what I’ve called Powerful AI.”
It is a confident forecast. It is also, as Mashable’s Chris Taylor pointed out in a sharp piece on 15 June 2026, hard to square with what Anthropic’s researchers have written in the documentation for the models Amodei is promoting.
Critical Context: Amodei is about to benefit from an Anthropic IPO. When a forecast of near-term, exponential capability gains arrives alongside a public offering, the forecast is also a sales document. That does not make it wrong — but it changes how much weight a buyer should give it.
The system card for the preview of Claude Mythos — the model Amodei singles out in the essay for its cybersecurity implications — contains this line: the intelligence “gains we can identify are confidently attributable to human research, not AI assistance … early claims of large AI-attributable wins have not held up.” In other words, the improvements came from people, not from AI accelerating its own development.
The test built to find an exponential, and what it found
The more pointed contradiction sits in the system card for Anthropic’s other new frontier model, Fable 5. Here the researchers did not stumble on the question by accident. They set out to look for it. Using a benchmark called the Epoch Capabilities Index, they specifically tested whether there was a feedback loop — the kind of self-reinforcing acceleration that would lead toward what the field variously calls AGI or digital superintelligence.
The finding, in the researchers’ own words: “We do not observe a sustained, AI-attributable 2× acceleration in the pace of our AI progress.”
Strategic Reality: A “2× acceleration” is the modest end of what an exponential requires. Anthropic’s researchers went looking for it in their own flagship model and reported that it is not there. That is not a sceptic’s hot take — it is the builder’s own measurement.
There is a real distinction worth holding onto here. Models are still getting better. Each release does more than the last. What the Fable 5 card disputes is not improvement but acceleration — the claim that the rate of improvement is itself increasing, which is what the word exponential actually means. A technology can advance steadily, even impressively, whilst its growth curve flattens rather than steepens. Plenty of genuinely transformative technologies have done exactly that.
What the two claims actually say
| Source | Claim | What it implies for capability growth |
|---|---|---|
| Amodei essay | Scaling laws predict exponential gains; “Powerful AI” within 1–2 years | Acceleration is real and near |
| Claude Mythos system card | Identified gains attributable to human research, not AI; early AI-attributable wins “have not held up” | No self-improvement loop yet |
| Fable 5 system card | No sustained, AI-attributable 2× acceleration observed | No exponential in the flagship model |
Where the exponential claim comes from
Pressed on his evidence, the trail leads to a single document: a 2020 paper, “Scaling Laws for Neural Language Models”, co-authored by Jared Kaplan — then at OpenAI, now a co-founder of Anthropic. That paper has carried an enormous amount of strategic weight for six years. A great deal of frontier-lab fundraising, and a great deal of public expectation, rests on the assumption that its curves extend cleanly into the future.
They may not. The AI researcher Gary Marcus has argued since 2022 that “there are serious holes in the scaling argument”, warning that the field “may already be running into scaling limits in deep learning, perhaps already approaching a point of diminishing returns.” He pointed to evidence that scaling “starts to falter on some measures, such as toxicity, truthfulness, reasoning, and common sense.” Marcus was widely dismissed at the time. The arrival of GPT-5 — capable, but not the superintelligence its loudest advocates had promised — has made his caution look better than his critics’.
Reality Check: The strongest empirical evidence for an exponential is a 2020 paper. The strongest empirical evidence against a near-term exponential is the 2026 system cards for the models built on that paper’s assumptions. Recent measurement beats old extrapolation.
This is the heart of it. An exponential is a claim about the future drawn from a line on a past chart. The further you extend the line, the more load-bearing the assumption that nothing bends it. Anthropic’s own 2026 measurements suggest the line is already bending.
Why this matters more for buyers than for builders
For a frontier lab, the exponential narrative is close to existential. It justifies the valuations, the capital expenditure on compute, and the urgency of the safety case. For a UK business buying and deploying these tools, the incentives run the other way. You gain nothing from believing the curve is steeper than it is, and you lose a great deal — in wasted capital, abandoned roadmaps, and credibility with your own board — if you plan for a capability jump that does not arrive on schedule.
Competitive Reality: Your suppliers are commercially motivated to tell you transformation is imminent. Your job is to plan for the capability that exists, plus a conservative estimate of the next 12 months — not the capability promised in a launch essay.
The practical risk is not that AI fails to improve. It is that organisations make decisions sized for an exponential and then have to unwind them. Headcount cut in anticipation of an autonomous agent that ships a year late. A product roadmap that assumes reasoning reliability the current models do not have. A capital commitment to infrastructure justified by a usage curve that flattens. The cost of believing the hype is paid by the buyer, not the seller.
How this reshapes the AI decision
| Stakeholder | If you believe “exponential” | If you plan for “steady improvement” |
|---|---|---|
| Board / finance | Front-load spend; size for disruption | Stage investment against delivered capability |
| Operations | Redesign around agents now | Pilot, measure, expand what works |
| Strategy | Bet on first-mover superintelligence | Compound advantage from reliable, deployed tools |
| Risk | Plan for runaway capability | Plan for capable-but-fallible systems |
The hidden challenges of a flattening curve
A decelerating, or merely linear, improvement path creates problems that the exponential story conveniently hides. Four are worth naming.
The first is sunk-cost lock-in. Infrastructure and tooling bought on the promise of next year’s capabilities can become expensive ballast when those capabilities slip. Mitigation: prefer reversible, consumption-based commitments over large fixed bets until a capability is in your hands, not in a roadmap.
The second is the reliability gap. Steady improvement on benchmarks does not automatically mean improvement on the specific, messy reliability your workflow needs. A model that is two per cent better on an index can still be unusable for your edge cases. Mitigation: test against your own data and tasks, not vendor benchmarks.
The third is narrative fatigue inside your own organisation. If leadership repeatedly promises transformation that does not land, teams stop believing the next, genuinely useful, rollout. Mitigation: under-promise on timing, and let delivered wins build the case.
Warning ⚠️: The most damaging outcome is not a project that fails. It is a portfolio of real, modest AI wins that never gets backed because the organisation was holding out for the revolution that the marketing promised.
The fourth is misreading the safety conversation. If the headline risk is framed as imminent superintelligence, attention drifts from the risks that are actually live today — fabrication, data leakage, automation of flawed decisions at scale. Mitigation: govern for the system you have deployed, not the one in the press release.
The strategic takeaway
The useful signal in this episode is not that Amodei is wrong — nobody can be certain about a multi-year forecast, and he may yet be vindicated. The useful signal is that Anthropic’s own researchers, measuring Anthropic’s own models, could not find the acceleration their chief executive describes. When the people closest to the technology report something more modest than the people selling it, the modest reading deserves the weight.
For UK decision-makers, that points to a posture rather than a prediction. Treat current AI as a genuinely powerful tool that is improving steadily — not as a countdown to a discontinuity. Three principles follow:
- Plan for the capability you can test today, plus a cautious read of the next 12 months. Treat anything beyond that as optionality, not as a commitment.
- Stage spend against delivered results. Reversible, consumption-based investments protect you whether the curve steepens or flattens.
- Govern the systems you actually run. The live risks are reliability and misuse, not science-fiction autonomy.
Take Action: Before your next AI investment decision, ask one question of every supplier projection: is this capability something I can verify in a pilot this quarter, or is it a line extended from a chart? Fund the first. Discount the second.
None of this argues against adopting AI. The opposite, in fact: a business that builds compounding advantage from reliable, deployed tools will be in a stronger position than one that froze, waiting for a revolution — or one that overbuilt for a revolution that arrived late. The companies that do best with AI over the next two years will probably be the ones that took the exponential narrative least literally.
Source and attribution
This analysis draws on reporting by Chris Taylor for Mashable, “Anthropic CEO says AI growth is exponential. Anthropic research says otherwise.” (15 June 2026), which surfaced the contradiction between Dario Amodei’s essay and the system cards for Anthropic’s Claude Mythos and Fable 5 models, with additional reporting by Timothy Werth. Gary Marcus’s critique of scaling laws is drawn from his published writing from 2022 onward. This is original Resultsense analysis; the editorial interpretation and the implications for UK businesses are our own.