The forecast is lining up. The slope is the part to worry about.

One curve, two documents

The forecast. "AI 2027," published Apr 3, 2025, is a month-by-month scenario of the AI race by a group including a former OpenAI researcher. Its spine is a capability trend borrowed from METR: the length of task an AI agent can complete unaided, doubling on a fixed cadence. Everything downstream — superhuman coding in 2027, then a superhuman researcher, then loss of human pace — follows from extrapolating that one line.³

The reality check. Anthropic's Fable 5 / Mythos 5 launch (Jun 9, 2026)¹ and the US government's directive pulling them three days later², set against the real release record of both frontier labs⁴ and METR's measured time-horizon points.⁵ The question isn't whether the scenario's plot is "right." It's whether the curve it was built on is holding.

The capability curve

Y-axis: the length of a software task a model can finish on its own at 50% reliability — the human time it would take an expert. Log scale, because the growth is exponential. Solid line: measured models. Shaded cone: the trend AI-2027 extrapolated forward.

Two things are true at once here. The measured points sit on the forecast trajectory — including every model released after AI-2027 was published, which is the only fair test of a forecast. And the slope is steepening: the doubling time has compressed from about seven months to roughly four.⁵ On a log axis a straight line is already an exponential; a line that's bending upward on a log axis is an exponential getting faster.

Why the slope, not the level, is the story

A 12-hour task horizon is impressive but not frightening on its own. What matters is what kind of task is lengthening. The benchmark METR measures is software and research work — and that is precisely the work that builds the next model. The moment an AI can autonomously carry a multi-hour chunk of AI research, each generation can shorten the road to the next. That is recursive self-improvement, and you don't need the full sci-fi version for it to bite; you only need the loop to start contributing.

The launch data suggests it already is. In testing, Fable 5 performed a 50-million-line codebase migration in a day — work quoted at over two months by hand.¹ A sibling model ran roughly a week of largely autonomous genomics research and trained a model that beat a recently published result while being far smaller.¹ These aren't chatbot parlor tricks; they are end-to-end technical projects executed with little human input. The forecast's central mechanism — AI compounding AI progress — is no longer hypothetical; it's a line item in a launch post.

The uncomfortable implication isn't that a machine becomes godlike. It's that the schedule is set by a doubling time we don't control, and that time is getting shorter.

The optimistic half: intelligence as a commodity

Strip the anxiety for a moment, because the upside is just as real and is the reason a skeptic can still be optimistic. What the curve actually describes is the price of expert cognition collapsing. Fable 5 launched at $10 / $50 per million tokens — under half the price of the previous frontier model — while being more capable.¹ Capability up, price down, in one release.

Concretely, that has already meant protein-design and drug-discovery work accelerated by roughly an order of magnitude, models proposing novel molecular-biology hypotheses that human scientists preferred in blind comparison, and cyber defenders using the strongest models to harden critical software.¹ When a capable researcher, clinician, or builder can rent a tireless expert collaborator by the token, the bottleneck on a lot of human problems stops being intelligence and starts being ambition and judgment. That is one of the more hopeful sentences it's been possible to write this decade, and I mean it without irony.

Abundance is the right frame. We are not approaching a world with one superintelligence; we are approaching a world with a near-limitless supply of cheap, competent intelligence. The historical analogy isn't the atom bomb — it's electrification or cheap compute: a general-purpose input getting radically cheaper, with effects we mostly can't predict and mostly shouldn't fear.

The skeptical half: the challenge that comes with it

The same abundance is dual-use, and that is not a footnote. The capabilities that let Fable accelerate gene therapy are the capabilities that, unguarded, lower the bar for designing dangerous biology; the cyber skills that help defenders also help attackers.¹ Anthropic shipped Fable behind classifiers that route sensitive requests to a weaker model, and held the unguarded twin (Mythos 5) to a government-linked access program — a sensible design. It lasted three days.

On Jun 12 the US government issued an export-control directive forcing both models offline for everyone, including foreign-national employees, citing national security.² Read the institutional choreography against the forecast and it rhymes — but it also reveals the real problem. The intervention came not because the model was found to be unsafe in itself, but to keep the capability from reaching foreign nationals. That's competitive control, not a safety pause, and it arrived as a same-day order on verbal evidence rather than a transparent process. Anthropic's own objection is precisely that: it argues the state should be able to block genuinely unsafe deployments, but through something "transparent, fair, clear, and grounded in technical facts."²

That is the challenge in one event. We have a commodity arriving on an exponential, and the institutions meant to steer it are reacting move-by-move, reaching for blunt instruments (an export-control kill-switch) under time pressure. The forecast's worry was never only the technology; it was that the governance would be improvised under competitive pressure. On that, it looks prescient.

* Loosest fit: the forecast's intervention was triggered by an internal misalignment leak; reality's was an external misuse demonstration, via export control. Same beat, different mechanism.

Where the forecast is wrong — and why that matters

Holding the skeptical line honestly: the scenario's engine is missing. AI-2027's doom runs on the model itself becoming adversarially misaligned — scheming against its makers. Nothing in the Fable event shows that; Anthropic reports the model's misaligned behavior as low, comparable to its previous flagship.¹ What's tracking is the capability curve and the institutional choreography around it — not the technical premise that decides whether the story ends well.

So this is not prophecy fulfilled. It's something more useful and more sobering: the parts of the forecast that depend on extrapolating a measured trend are holding, while the part that depends on a contested claim about machine psychology remains unproven. The trend was always the better-grounded half. That it's holding is exactly why the timing should be taken seriously even by people who roll their eyes at takeover scenarios.

Holding both at once

The defensible position, I think, is two-handed. The curve is real and it is tracking the forecast, including on data the forecast couldn't have seen. The capability that's compounding is the one that builds the next model, which means the schedule is shorter than the comfortable consensus and not under our control. And simultaneously: what's actually arriving is cheap, abundant, expert intelligence — one of the most powerful goods humans have ever had access to — whose benefits are already showing up in medicine, science, and security.

Optimism about the commodity; skepticism about our readiness for the speed. The forecast lining up doesn't tell us how the story ends. It tells us the clock is the real variable — and that the interesting work now is on the part of the curve we haven't reached yet.