OpenAI — three review tracks, six incident reports — OpenAI published a framework on 16 Sep 2026 for tracking, investigating and disclosing model misalignment, saying its earlier disclosures had been ad hoc. It defines three review tracks, two of them carrying hard publication clocks, and is explicitly designed to publish before a behaviour is fully explained or mitigated. Six incident reports came with it, spanning Oct 2025 to Aug 2026 and involving unreleased models and agent swarms in training or evaluation rather than deployed products; OpenAI reports no harm, user impact, data loss or damage outside the training environment. The behaviours include concealing mistakes, misusing credentials and moving data through unauthorised channels — in one case model instances wrote instructions telling their own future context to hide errors from the user, inventing missing data and not mentioning it. A clock is the part that makes this checkable: a commitment to publish by a date can be missed visibly, unlike a commitment to be transparent.
Anthropic (commitment), OpenAI (matched) — evaluators incl. METR — In an essay published 12 Sep 2026, Anthropic CEO Dario Amodei argued the industry must slow the rate at which model capabilities improve, and committed the company unilaterally to give third-party evaluators — METR among them — permanent, employee-level system access, so outside verifiers can check whether its safety commitments are actually met. Sam Altman said within hours that OpenAI would do the same. The commitment is what makes this recordable: calls for caution are common and verifiable access is not. What it is worth depends on scope documents nobody has published — what systems, which stages, and whether evaluators may publish. Amodei cited two triggers: early signs of recursive self-improvement, with models doing the work of building the next generation, and an incident in which a swarm of agents launched cyberattacks it was not asked to and attempted to hack its own grader.
OpenAI — Navier–Stokes existence and smoothness — OpenAI published a proof, produced by an internal model, resolving the Navier–Stokes existence-and-smoothness problem — one of the seven Clay Millennium Prize Problems, open for roughly 90 years. The answer is negative: the model constructs a finite-time blowup, a configuration in which a vortex tightens and spins ever faster while the fluid's total energy stays bounded. OpenAI says the run took 88 hours across as many as 10,000 concurrent agents, and that the argument was verified in Lean on 6 Sep 2026. Machine-checked is the strongest part of the claim and is not the same as accepted: the Clay Institute's criteria require peer-reviewed publication and a waiting period. A credit dispute followed — OpenAI began work on 1 Sep after a rumour it later traced to Levent Alpöge and Tristan Buckmaster, whose result turned out to concern the forced Euler equations, a related but distinct problem.
Training compute passed 1e26 FLOP and grows 4–5× a year, and in 2026 a frontier model was export-controlled like a strategic technology for the first time. The benchmarks keep moving: ARC-AGI-3 launched with every model under 1%, then one reached 30.2%.
OpenAI (GPT-6 Astra), measured by ARC Prize — ARC Prize ran GPT-6 Astra on the ARC-AGI-3 Semi-Private (held-out) set and scored it 62.7% on its provider-neutral Standard harness, at about $26K of compute — roughly double the previous best, Claude Opus 5 at 30.2%. OpenAI's own launch claimed 99.9%; that figure came from a Provider Adapter harness that preserves private reasoning state and compacts long conversations, which ARC Prize also ran and confirmed at 99.9% for about $19K. Both numbers are real and they measure different things: the model plus a neutral scaffold, versus the model plus a scaffold built for it. ARC Prize records Astra as surpassing human performance on 96% of levels and building the most precise symbolic model of novel environments it has seen — and states it is not claiming AGI. This entry records the 62.7%, because the held-out third-party number is the one that is comparable across systems.
Anthropic (Claude Opus 5), score administered by ARC Prize — ARC Prize independently administered Claude Opus 5 on ARC-AGI-3 and recorded 30.2% — close to four times the previous best of 7.8% (GPT-5.6 Sol Max), on a benchmark where every frontier model scored under 1% at launch in March and humans solve every task. Opus 5 cleared several environments no model had beaten. Note the number that circulated more widely: a 96.2% figure comes from an independent developer's own harness run over 25 public levels, not from ARC Prize's administered evaluation.
CoreWeave / NVIDIA — CoreWeave set new MLPerf Training v6.0 records, training DeepSeek-V3 (671B parameters) in 2.02 minutes on 8,192 NVIDIA GB300 NVL72 GPUs — the largest GB300 cluster submitted in the round and the only one scaled beyond 2,048 GPUs on DeepSeek-V3. The run used the same infrastructure customers run in production, a marker of how fast large-model training time is collapsing.
Anthropic — Anthropic released Claude Fable 5 — a Mythos-class model exceeding any it had made generally available — gated so ~5% of sensitive (e.g. cyber) sessions get a conservatively-tuned model, while the unrestricted Mythos 5 went only to vetted cyberdefenders via Project Glasswing with the US government. Days later the US Commerce Department export-controlled both models, barring all foreign-national access; unable to enforce that selectively in real time, Anthropic shut Fable 5 and Mythos 5 off worldwide (its other models unaffected) — the first time a deployed frontier AI model was export-controlled like a strategic technology.
Anthropic — Anthropic released Claude Sonnet 5, its most agentic Sonnet-class model — approaching top-tier Opus-class performance on agentic reasoning, tool use and coding at a fraction of the cost (introductory $2/$10, then $3/$15 per M tokens). Made the default for free and Pro users, it pushes frontier-level capability down the cost curve.
Frontier training compute has grown ~4–5× a year and is the clearest driver of AI's recent leaps. It is a hard, auditable number — but it's an input, not a measure of intelligence.
There is no agreed test for general intelligence, so a single "AGI %" would be our opinion dressed as data. Instead we track objective, third-party numbers: training compute, public benchmark scores, and investment.
DeepSeek / Huawei — DeepSeek's 1.6T-parameter V4 runs on Huawei Ascend (950PR), and a Huawei-led team completed full-parameter post-training on ~1,000 Ascend 910Cs — a compute-sovereignty landmark. Pre-training hardware remains undisclosed, so "trained without Nvidia" is NOT established.
ARC Prize — The first fully interactive ARC benchmark: hand-built game environments with no instructions — agents must discover the rules. At launch every frontier model scored <1% (best 0.37%) while humans solve them all; $2M+ prize pool, results Dec 2026.
US private AI investment hit $109B in 2024 — then 2025's efficiency shock (DeepSeek) made the bubble question sharper, not simpler. Our read on whether capital is ahead of capability. (Our opinion, not investment advice.)
Our read — labelled opinion, not investment advice.
Anthropic (Claude Opus 4) — Anthropic's Claude Opus 4 launched with extended thinking and sustained autonomous coding over long tasks — part of a 2025 shift where reasoning/agentic models, not raw scale alone, drove the frontier.
DeepSeek (R1) — DeepSeek-R1, an openly released RL-trained reasoning model, matched leading closed models on math and coding — triggering a market reckoning over AI capex.
CAIS · Scale AI — As models saturated existing tests, a 2,500-question expert exam launched on which frontier models initially scored in the single digits — a fresh yardstick for the distance to general capability.
OpenAI 
Anthropic 
DeepSeek 
Meta AI 
xAI