Anthropic's AGI Timeline 2027: When Does AGI Actually Arrive? OpenAI vs Anthropic Breakdown

Anthropic expects AGI by early 2027. OpenAI just shipped what it calls a step toward it. And the Financial Times reports OpenAI has overtaken Anthropic.

This week, three signals converged around the same question — when does AGI actually arrive — and they came from the two labs racing to build it.

The most striking detail isn't the date. It's that Anthropic is the only frontier lab with a public AGI timeline, and the timeline — drawn directly from CEO Dario Amodei's own essays and official government submissions — points to early 2027. Not "maybe." Not "in principle." Early 2027.

OpenAI, meanwhile, launched GPT-6 Astra on September 3 and had its president, Greg Brockman, call the model AGI in a press briefing. The Financial Times reported this week that OpenAI believes it has overtaken Anthropic with Astra. And Anthropic's CEO spent the weekend calling for everyone to slow down.

Here's what Anthropic's AGI timeline actually says — and why, as of mid-September 2026, we're now inside the narrowest part of it.

Anthropic's AGI Timeline: What They've Actually Said

Anthropic's AGI forecast isn't a slide deck or a leak. It's in writing, in multiple public places:

• Machines of Loving Grace (Dario Amodei's essay): powerful AI "could arrive as early as 2026, though there are also ways it could take much longer."
• March 2025 OSTP recommendations: "we expect powerful AI systems will emerge in late 2026 or early 2027" — systems "smarter than a Nobel Prize winner across most relevant fields."
• Jack Clark (Anthropic co-founder) confirmed in September 2025 that something like this "remains Anthropic's view."

Anthropic describes this capability level as a "country of geniuses in a datacenter" — AI systems with interfaces for text, audio, video, mouse and keyboard, internet access, able to autonomously carry out tasks lasting hours, days, or weeks, running at 10x–100x human speed, with millions of instances repurposed from training compute.

The key operational detail, per the OSTP submission: "Powerful AI technology will be built during this Administration." That implies a high probability — possibly 80% — on powerful AI before January 2029, and a decent probability by mid-2027.

So when people ask "when does Anthropic think AGI arrives?" the honest answer from the company's own public statements is: late 2026 to early 2027, with around a 6% chance by early 2027 according to one detailed internal-modeling exercise (by researcher Ryan Greenblatt on LessWrong, who finds the timeline unlikely but worth taking seriously).

The Timeline in Detail: What Early 2027 Looks Like From the Inside

Anthropic's forecast implies a compressed sequence:

• September 2026: Research engineers accelerated ~5x. AIs successfully complete most engineering tasks that would take decent human engineers months, and 90% of tasks taking a few weeks. Full automation of junior research engineering is functionally achievable with management.
• December 2026: Full automation of research engineering. AIs complete massive greenfield projects — like reimplementing a Rust compiler from scratch in C with similar performance.
• February 2027: Full automation of AI R&D itself. Engineering multiplier ~200x, AI R&D multiplier ~35x.
• March 2027: Powerful AI. Engineering multiplier ~600x, AI R&D multiplier ~100x.

That's the forecast. Note the tightness: December 2026 to March 2027 is a 3.5-month window from full research-engineering automation to what Anthropic defines as powerful AI. That only works if AI R&D acceleration compounds fast.

Anthropic's own engineer-acceleration numbers give a sense of where they think they are today: research engineers accelerated ~5x as of September 2026. That's consistent with the timeline — they believe they're already inside the ramp.

OpenAI's Position: Astra, Brockman on AGI, and the FT Report

While Anthropic is publishing its timeline, OpenAI is shipping what it believes is a landmark.

GPT-6 Astra, launched September 3, saturates ARC-AGI-3 at 99.9%, ExploitBench at 100%, and FrontierMath Tier 4 at 98%. Greg Brockman called it AGI in a press briefing. The FT reported this week that OpenAI says it has overtaken Anthropic.

OpenAI's own timeline disclosure came from Chief Scientist Jakub Pachocki, who told staff: "we want to take this maybe somewhat unusual step of sharing our internal goals and goal timelines towards these very powerful systems... automated AI research by March 2028." He added: "these particular dates, we absolutely may be quite wrong about them."

So the two labs' public timelines are:

Anthropic is the more aggressive on the calendar — and it's the only one framing it as a forecast rather than an ambition.

What This Means for the AGI Race Right Now

Three things worth noting as of mid-September 2026:

1. We are inside Anthropic's forecast window. The September 2026 milestone — 5x engineer acceleration, mostly automated — is now. The clock on the 3.5-month ramp to March 2027 started, by their own model, this month.

2. The two labs are telling different stories about pace. OpenAI's Pachocki is more cautious (March 2028 automation goal, "we may be quite wrong"). Anthropic's Amodei is more specific and more confident on the early-2027 date. And yet Amodei spent Saturday calling for an immediate global slowdown. The tension between "AGI by early 2027" and "we must slow down" is not a contradiction — but it is worth sitting with.

3. The benchmarking conversation is shifting. With GPT-6 Astra saturating ARC-AGI-3 (a benchmark specifically designed to measure progress toward AGI), the question is no longer whether benchmarks can track progress. It's whether they can track it fast enough to matter. Astra's 99.9% on ARC-AGI-3 is, by the benchmark's own standard, effectively human parity.

Why Anthropic's Timeline Matters Even If It's Wrong

Ryan Greenblatt — who did the most careful public modeling of Anthropic's forecast — puts the probability of powerful AI by early 2027 at around 6%. He thinks the timeline is very unlikely.

But he also gives Anthropic credit for doing something no other frontier lab does: making a falsifiable prediction at all.

And there's a reason the 6% matters even if the 94% wins. If Anthropic is right, the next 18 months contain some of the most consequential capability jumps in the history of the field — and the safety infrastructure, policy, and public understanding are not currently scaled to match. That's the argument Amodei made Saturday in his slowdown essay, and it's the same argument he's been making since Machines of Loving Grace.

The timeline isn't just a prediction. It's a risk assessment.

The Bottom Line

As of September 14, 2026:

• Anthropic expects AGI by early 2027 and has said so in official submissions.
• OpenAI launched GPT-6 Astra (which Brockman called AGI) and targets automated AI research by March 2028 — a later date, expressed with more uncertainty.
• The Financial Times reports OpenAI believes it has overtaken Anthropic.
• Anthropic's CEO is simultaneously calling for the industry to slow down.

If you're trying to track where AGI stands this week, the most informative single data point may be this: the two labs most responsible for building it are now publishing concrete timelines — and they disagree on the date by roughly six months.

In a field where the median expert prediction a decade ago was "decades away," a six-month gap between the two leading labs is, in itself, a signal.

Published Monday, September 14, 2026. Sources: Anthropic's Machines of Loving Grace essay and March 2025 OSTP recommendations; Jack Clark's confirmation on X (September 2025); Ryan Greenblatt's LessWrong analysis of Anthropic's AGI timeline; OpenAI's GPT-6 Astra launch announcement, Greg Brockman's press briefing, and Jakub Pachocki's internal timeline disclosure; Financial Times report on OpenAI overtaking Anthropic; IPS/CBS News/Guardian reporting on Dario Amodei's slowdown essay. Follow AIPress for ongoing coverage of AGI, ChatGPT, Claude, OpenAI, Anthropic, and the AI industry.