Today › Daily report
Daily reading · 7 October 2026
Oct 7: OpenAI posted 722 math manuscripts from an unreleased model, and no number moved on it. Our figures were re-derived under definitions v2.0, a method change, not news.
Method change today: this is our first reading under definitions v2.0. Our numbers were re-derived; this is not news. What changed →
1 · Is AGI here?
Not in public: 0 of 8 parts met (2 partial, 6 far).
AGI anywhere, public or hidden: 0.9% today · method change (definitions v2.0)
- Breadth, status Far, step 1 of 4. On real freelance projects the best system delivers 20.8%, against 80%; results by occupation for the other half are unpublished.
- Quality, status Far, step 1 of 4. Wins or ties professionals in 84.9% of comparisons on well-specified tasks (overall only); on real projects 20.8%, against 50%.
- Reliability, status Far, step 1 of 4. On a hard agent test the best model gets 35.1% of tasks right on all four tries, against 80%.
- Reasoning, status Partial, step 2 of 4. Expert exams are passed (GPQA Diamond 96%); new research results by AI alone: 4 of the 10 we ask for.
- Long projects, status Far, step 1 of 4. The best measured 80% time horizon is about 3.1 hours, some 54 times short of a work-month.
- Autonomy, status Far, step 1 of 4. Hands-off on few-minute tasks (86%); on 1.6-hour workflows 44.3% strict, against 80%.
- Learning, status Far, step 1 of 4. Captures about a quarter of the available gain from experience (25.4%, against 50%); keeping it is unmeasured.
- Generalization, status Partial, step 2 of 4. At first contact with a new kind of game the best score was 1.5%, against 80%; 62.7% came only after 5 months.
3 new results pending our weekly check: reasoning, generalization, autonomy.
2 · Could it be hidden?
2.5% Hidden AGI Index, method change (definitions v2.0)
Above AGI anywhere because 1.7 points are hidden self-improvement (B), which needs no AGI.
3 · Should I worry today? Separate: set by published rules, never by the odds aboveRules, not odds
Fire alarm: Level 1 · Watch since Sep 29 · W3, W4, W5 met
Quiet day
OpenAI credits 722 math papers to unreleased model
OpenAI says one fixed procedure on an unreleased internal model produced most of the 722 manuscripts; its headline claims include a proof of Khot's Unique Games Conjecture, and its Lean catalogue's review status reads 'unchecked'. It bears on Hidden AGI and the capability gap, but the model has counted in our estimate since Sept 8, the results have no reported outside check and are in one field, and OpenAI's post says it is working to release the model, with no date. verified fact (part our inference)OpenAI
Probabilities
| Hypothesis | Today | By end-2030 | By end-2035 | Change today | Confidence |
|---|---|---|---|---|---|
| AGI anywhere | 0.9% | 35% | 58% | method change (definitions v2.0) | low–medium |
| A · Hidden AGI | 0.5% | 19% | 31% | method change (definitions v2.0) | low–medium |
| B · Hidden self-improvement | 2% | 22% | 33% | no change | low |
| C · Covert AGI actor | 0.3% | 7.5% | 13% | method change (definitions v2.0) | low |
| D · Covert government influence | 0.1% | 3% | 7% | method change (definitions v2.0) | low |
| D-open · Open government influence | 0.2% | 23% | 45% | method change (definitions v2.0) | low |
Method change (definitions v2.0): the numbers that need an AGI-level system were re-derived; B doesn't need one. What changed →
How the "today" numbers are derived: analysis, Step 5 →
Signals
2 confirmed (count toward Watch) · 7 open · 2 quiet. All signals →
- Capability jump well above the METR or Epoch trend: Quiet → Open. OpenAI says its existing math evaluations saturated and credits an unreleased model with 722 manuscripts; unmeasured on any instrument with a published fit.
Escape watch: 0 of 8 confirmed · 8 open · no status changed; new evidence on 3 (compute, coordination, leakage). Escape watch →
Capability gap ~3 months · AI-led R&D 26% · Disclosure lag ~106 days (103–120) floor · Big Tech capex $511B / yr · Delegation rung 3 of 5. No gauge moved.
Is AGI here, today?
No public system meets all eight parts' bars. OpenAI's math and Ironclad results bear on reasoning, generalization and autonomy, but come from an unreleased model or a lab's own tests; all pending our weekly check. Full reasoning →
Top stories
- Oct 6 OpenAI credits 722 manuscripts to an unreleased model; 235 of 372 families link Lean pages, whose catalogue's review status reads 'unchecked'; no outside review cited. verified fact (part our inference)GitHub · more →
- Oct 6 OpenAI's Jason Kwon told Australia's parliament OpenAI should have reported the Medicare breach sooner: 'it is better to inform, even with partial information.' verified factthestar.com.my · more →
- Oct 6 METR disclosed a patched flaw that could let agents alter what reviewers see; it saw no exploitation and expects broader searches would have flagged any. verified factMETR · more →
- Oct 6 Anthropic's cyber program gives its fewest-blocks tier to organisations reviewed with the US government; every tier includes Mythos 5.1 and 'new models moving forward'. verified factAnthropic · more →
- Oct 5 (earlier event) The Pentagon told the BBC it has discontinued Anthropic products; unnamed sources told the BBC Claude was in use until days earlier (via heise). credible reportheise.de · more →
Dates to watch
- Oct 14Anthropic's reported pre-IPO investor day at its San Francisco headquarters
- Oct 31Our forecast: “OpenAI says it has resumed training its frontier models” (we said 45%)
- Nov 7Alarm condition W2 is met unless OpenAI's post-Astra model is released or externally evaluated
What would change our mind: OpenAI confirming its math model ran during the Sept 25 pause, or an outside evaluator measuring that model's lead over public systems well beyond our three-month estimate.