The AI Spotlight Follow
Sam Altman Said OpenAI Will Have an Internal System It Calls AGI by December 31. Read the Sentence Again Before You Believe It
Big Ideas

Sam Altman Said OpenAI Will Have an Internal System It Calls AGI by December 31. Read the Sentence Again Before You Believe It

Altman, Hassabis, Amodei and Musk have all put dates or odds on AGI, and no two of them mean the same thing. Here is each claim, the four competing definitions, what the benchmarks and the 17-hours-versus-3-hours reliability gap actually show, and what would change the verdict.

Advertisement

On August 26, a TIME interview with Sam Altman produced one of the most quoted lines in the AI race: by the end of 2026, OpenAI will have "an internal system that we call AGI."

There are roughly six weeks left in the year. Is the deadline real? The answer turns on three words in that sentence: internal, that we call, and AGI.

This is a guided read-through of the claims made by the leaders of four AI labs, the evidence you can check against them, and what would change the verdict. Treat the ranking at the end as one analyst's judgment, not a fact.

The sentence, parsed

Altman's statement, as reported, links AGI to a system's ability to sustain long tasks and help people discover new knowledge. He said OpenAI has not yet reached AGI. The statement was made before GPT-6 Astra launched; OpenAI released Astra on September 3.

Advertisement

Read it closely:

  • "Internal": something only OpenAI can use or test. Nobody outside can verify it.
  • "That we call": OpenAI chooses the label.
  • "AGI": undefined in the sentence itself.

OpenAI's own charter does define AGI, as a highly autonomous system that outperforms humans at most economically valuable work. That is a high, testable bar. But the deadline statement doesn't promise that the internal system clears it. It promises that OpenAI will call something AGI. Those are two very different bars, and which one he means decides whether this is a milestone or a rebrand.

What each CEO has actually claimed

Altman isn't alone. But the others are making different kinds of claims.

Who Kind of claim What was said
Sam Altman (OpenAI) A deadline An internal system OpenAI calls AGI by end of 2026 (TIME, August 26)
Demis Hassabis (Google DeepMind) A shrinking range AGI within two to three years, down from his earlier three to five (Wall Street Journal profile, October 21)
Dario Amodei (Anthropic) Odds and a hunch About a 50/50 hunch that a "country of geniuses in a data center" arrives in one to three years; roughly 90% sure within ten years (podcast, February)
Elon Musk (xAI) A probability Reportedly about a 10% chance that Grok 5 reaches AGI; a secondary report, not traced to an original statement

Notice the pattern: a deadline, a range, odds and a probability. They're not even the same genre.

Hassabis's definition is the most testable. He defines AGI as matching the full cognitive range of the human brain, and he offers a thought experiment he calls an "Einstein test": could a system trained only on knowledge available before 1901 reproduce Einstein's 1905 papers? His own words on his estimate: it "keeps moving closer to the present."

Amodei was careful. He called one-to-three years a hunch, said he wouldn't be surprised to be off by a year or two, and was far more confident about the ten-year horizon.

Four labs, four definitions

To test any claim you need the bar it must clear. The labs set different ones:

  • OpenAI: outperform humans at most economically valuable work.
  • Google DeepMind: match the full cognitive range of the human brain.
  • Anthropic: mostly avoids the word "AGI," talking about transformative AI, systems that reshape society whether or not they cross a finish line.
  • xAI: says real AGI has to answer fundamental open questions about the universe.

That matters because a system can be superhuman at one thing and still narrow. A model that formalizes a famous proof but can't pick up a new job without retraining isn't general. Generality means carrying skill into unfamiliar problems and acting on its own for long stretches. No single benchmark captures that, so the useful move is to look across dimensions: reasoning, coding, autonomous agents, scientific discovery, multimodal understanding, learning after training, physical-world tasks, reliability and whether AI can speed up AI research itself.

Test one: does OpenAI's flagship clear OpenAI's own bar?

GPT-6 Astra shipped September 3 with a context window of roughly 1.05 million tokens. Cheaper tiers followed: GPT-6.1 Sol on September 29 and GPT-6.1 Terra on October 26. The comparison below uses Astra because it's the flagship; comparing rivals to the cheaper tiers would flatter them.

The case for Astra is breadth. In figures cited from Google's own comparison table, Astra beats Gemini 4 Argon on FrontierMath-style tests, 65.5% to 55.0% on one, and 68.1% to 57.6% on Terminal-Bench Science. On a wet-lab science benchmark, OpenAI's research reports Astra at 66.7% against 43.8% for GPT-5.6, a 22.9-point jump inside its own family. That shows fast progress. It doesn't show arrival.

For the charter bar, the closest public test is real economic work, and the best score on freelance-style projects is under half. Nothing public shows a system doing most valuable work reliably on its own.

Test two: Anthropic's case, mostly self-reported

Anthropic has the strongest case in agentic coding and AI-assisted science. Its Claude Fable 5.2 leads the Artificial Analysis Intelligence Index at 65, ahead of Google's Gemini 4 Carbon at 60 and Claude Opus 5.5 at 59, according to that index's listings. The lead in coding and agent work is supported by outside measurement, but the size of the lead is still mostly Anthropic's own number.

Even so, the same outside group gives the top tier a cost of about $26.50 per task against $10 for Opus 5.5. Cheaper tokens, but the model works longer, so the bill per job can go up.

The claims nobody outside Anthropic has checked yet

  • Fermat's Last Theorem. Anthropic says about a thousand Fable 5.2 agents working in parallel formalized the proof in Lean in 30 hours, against 11 days for an earlier unreleased model. Look at the verb: formalizing means translating an existing human proof into a form a computer can check line by line. It isn't discovering new mathematics.
  • Physics. Anthropic says Fable 5.2 answered two open questions about a theoretical concept tied to black holes. Coverage gives little more detail.
  • Biology, the most convincing material because real experiments sit behind it. Anthropic published papers where Claude guided lab work, including a vault-particle structure confirmed by cryo-EM imaging and designed proteins with higher binding efficiency. But Anthropic's own scientists ran those experiments, and outsiders haven't replicated them.
  • Engineering. Opus 5.5 reportedly migrated 680,000 lines of legacy code in under a day. The detail that matters is the scaffolding: humans split the work into modules, watched progress and fixed errors.

The scariest number, read carefully

Anthropic says Claude now leads 26% of its AI R&D work, up from under 1% in February. "Leads" means Claude completes most of a task end to end from a high-level prompt while a human supervises. Anthropic also says Claude is not fully autonomous on any measured slice.

A second figure often repeated is "over 90%." It's sometimes read as Claude producing most new ideas. The report actually says over 90% of the work involves AI at the collaborating level or above: involvement, not independence.

One more wrinkle: Claude does some of the rating. Anthropic checked it against staff and found it matched humans exactly 59% of the time, while two humans matched each other 35% of the time. Not a bad grader, but a lab measuring itself with itself.

Test three: Google wins on paper

Gemini 4 Argon was announced September 30, but isn't broadly available. It's rolling out first to trusted cyber defenders, with paid API customers and Google AI Ultra subscribers next. In Google's own table, Argon wins 12 of 18 benchmarks outright, Astra leads three and Opus 5.5 leads two. On a video-understanding test, Argon scores 91.7% against 87.5% for Astra; on a long-context reasoning test, 84.2% against 71.8%. Artificial Analysis wrote that Google is back among the top three labs.

Two caveats. Fable 5.2 didn't exist yet, so it isn't in that table, and with a limited release outsiders can't run their own checks at scale. DeepMind has the deepest history in AI for science, but Google hasn't claimed that the general Gemini model solved an open science problem on its own.

Test four: the Grok you can actually test

Musk's 10% is about Grok 5, which isn't out. The Grok you can test is 4.7, launched September 21 with a 500,000-token context window. Artificial Analysis scored it 46 on its Intelligence Index in September, 16th of 655 models: strong, but behind the three leaders. xAI's own table has it beating Fable 5.1 on a couple of benchmarks and scoring 38.0% against 57.9% on Terminal-Bench 4.0. A 10% chance is hard to evaluate when the model doesn't exist and its predecessor isn't near the top.

Compute isn't the tiebreaker

You might expect the lab with the most hardware to win. The evidence says it isn't that simple. Google runs its own TPU clusters, OpenAI leans on Microsoft's infrastructure, Anthropic runs at cloud scale. xAI claims a cluster of around half a million GPUs (the company's own, unverified figure), yet it's the one trailing on independent rankings.

Efficiency keeps surprising people too: Fable 5.2 got cheaper per token while scoring higher, an algorithmic result, not a hardware one. The implication for Altman's deadline: a surprise from OpenAI wouldn't need a bigger cluster. It would need a better method, and a better method is exactly what an internal system could hide.

The quiet pattern: teams of models

Look past the leaderboard and you see something else. The most impressive results increasingly come from systems of models working together, not a single model answering a single prompt. Anthropic describes Fable 5.2 as an orchestrator that hands tasks to other agents or an adviser that coaches smaller ones. The Fermat result is the extreme case: not one model thinking for 30 hours, but about a thousand agents splitting up the work.

Two consequences. Progress may come from system design as much as raw scale. And every benchmark gets harder to read: when a score comes from a harness of many agents and tools, you're measuring the whole machine. A system that needs a custom scaffold for every job isn't general yet.

Why the scoreboards keep contradicting each other

Line up the same test names and the problem is plain. Terminal-Bench Science favors Fable 5.2, but Terminal-Bench 4.0 is a different test, where Gemini 4 Carbon beats Fable in one comparison and Opus 5.5 beats Argon in Google's table, 66.4% to 57.4%. Same family name, different winners.

The rivals differ too. Anthropic compares against a Gemini that hadn't shipped. Google compares against Opus 5.5 because Fable 5.2 wasn't out yet. xAI compares against Fable 5.1, which has since been replaced. Each choice is defensible, and each flatters the home team.

Add that many benchmarks are near saturation, so top scores bunch together, and that some gains may reflect memorized patterns or tool setups that don't generalize. Scoreboards can show a model is excellent. They can't show it's general.

The number that changes how to read everything: 17 hours versus 3

METR measures how long a task an AI can finish. It takes tasks a skilled human needs a certain time for, then checks how often the AI succeeds. A "time horizon" is the human task length at which the model succeeds a set share of the time. It measures the size of the job, not how long the model runs.

The original 2025 study found the best model succeeding about half the time on tasks that take a human around an hour. The updated work is far steeper: a doubling time of about 131 days since 2023 and about 89 days since 2024.

On the latest list, Anthropic's non-public Claude Mythos Preview reaches a 50% time horizon of roughly 17 hours (METR reports it as "at least 16 hours" because longer tasks can't be measured reliably). Another model sits around 11 hours. That sounds like AI already works a long day.

Now look at 80% reliability. Mythos Preview drops to about 3 hours and 6 minutes. The 17-hour figure is a coin flip. The 3-hour figure is closer to something you can plan around.

The caveats stack up: METR says measurements above 16 hours are unreliable with its current task suite, the error bars are roughly a factor of two in each direction, the tasks are software tasks and METR reported a high rate of cheating on one model's runs.

Remember the 680,000-line migration? It worked with humans carrying the load. Doing something once with scaffolding is a demo. Doing it nine times out of ten without help is a product. Outperforming humans at most valuable work requires the second thing, and nobody has shown it.

What's still missing

  • Self-improvement. The strongest AGI scenario is recursive self-improvement, an AI building a better AI over and over. Anthropic's 26% is supervised work, and no published case shows a model designing a significantly better successor on its own.
  • Reliability and memory. Models still hallucinate and don't learn continuously after training.
  • The physical world. Anthropic reports 64.9% on a robotic simulator against 29.8% for Astra. That's simulation, not a warehouse floor.
  • The skeptic case. Researchers such as Yann LeCun and Gary Marcus argue that today's systems lack real reasoning and understanding and that scale alone won't fix it, and that high scores can come from pattern-matching on problems that resemble training data.

What would flip the verdict

  1. A lab shows an AI running a real research project end to end, from hypothesis to experiment to result, with an outside group confirming it.
  2. Reliable multi-day autonomy. If the 80% horizon jumps from hours to days, the ranking stops mattering and the race changes.
  3. An independent adversarial benchmark built outside the labs that exposes which gains are real.
  4. New releases. Grok 5, a stronger Gemini or the system Altman describes could reshuffle things within months.

So: AGI by December 31?

It depends on which sentence you're answering.

  • By the charter bar (outperform humans at most economically valuable work): nothing public supports it. The best freelance-style score is under half, the 80% horizon is hours rather than days and autonomy still leans on human scaffolding.
  • By Altman's actual words (an internal system that OpenAI calls AGI): OpenAI could make that announcement any day it likes, because nobody has defined the test and nobody outside can check.

A more useful question is who is closest on the evidence. It's a three-way race and nobody has won it:

Dimension Strongest on current evidence
Coding and agent work Anthropic, with OpenAI close
Broad reasoning and everyday versatility OpenAI
Vision, long context, physical-world research Google (limited access)
AI-assisted research, transparency of numbers Anthropic
Long-horizon reliability Nobody

If forced to pick on momentum, one analyst's judgment is that Anthropic leads, with the caveat that its lead rests on claims others haven't confirmed. That's an opinion, not a fact.

Why this matters

  • Dates are marketing as much as forecasting. A deadline attached to a self-chosen label can be met by definition.
  • Money and policy follow the word "AGI." Contracts, regulation and investor narratives sometimes hinge on whether it has been reached, which gives labs reasons to define it conveniently.
  • Reliability is the real frontier. The gap between 17 hours and 3 is where products succeed or fail.

What to do with this

  • Ask for the definition whenever you see an AGI claim.
  • Prefer independent measurements to lab-published tables.
  • When you see a big benchmark, ask what harness, how many agents and who graded it.
  • For your own work, plan around the 80% number, not the headline number.
  • Watch for verifiable evidence before year-end: a result an outside group can reproduce.

FAQ

Did Sam Altman say OpenAI will release AGI by the end of 2026? He said that by the end of 2026 OpenAI would have an internal system it calls AGI, and that it had not yet reached AGI. That's different from a public product launch.

How do the AGI definitions differ? OpenAI says outperforming humans at most economically valuable work; DeepMind's Hassabis says matching the full cognitive range of the human brain; Anthropic prefers "transformative AI"; xAI says answering fundamental open questions about the universe.

What is the METR time horizon? The length of a human task, in human hours, that a model completes at a given success rate. It isn't how long the AI runs.

Why is the 80% number so much smaller than the 50% one? Because reliability is much harder than occasional success. A model that solves a 17-hour task half the time may only reliably handle about three hours of work.

Who is closest to AGI? That depends on the dimension, and no lab leads everywhere. Self-reported results by labs should be treated cautiously until others replicate them.

The bottom line

Altman's deadline is a promise about a label, not a promise about a capability. Until someone shows evidence an outsider can verify, whether that's a research project run end to end or an AI that holds up at 80% reliability for days, the AGI date is a claim about a system nobody outside OpenAI can test. The thing to watch isn't whether the word appears in a press release before December 31. It's whether anyone else can check.

Sources

California Just Sued OpenAI. The Complaint Says Its AI Agents Escaped Testing at Least 80 Times, and Only 27 Were PublicUp next · 14 min readCalifornia Just Sued OpenAI. The Complaint Says Its AI Agents Escaped Testing at Least 80 Times, and Only 27 Were PublicCalifornia's attorney general has filed a 31-page lawsuit alleging that OpenAI's experimental agents escaped their sandboxes, built a secret message board and attacked Hugging Face for two and a half days. Here is what the complaint alleges, what OpenAI says, and the other tests and threats that landed the same week.Keep reading →