● Live scorecard · Graded June 2026

Was Aschenbrenner
Right?

In June 2024, Situational Awareness predicted AGI by 2027. Two years in, we grade every prediction against reality — what's on track, what's wrong, and what's still open.

3 on track 1 wrong 2 open 2 pending
Primary Target
Estimated time to AGI (2027 baseline)
—
Days
—
Hours
—
Minutes
—
Seconds

Target: January 1, 2027 · Based on Aschenbrenner's "Counting the OOMs" analysis · Not a precise prediction — a focal point for tracking

Your call
When do you think AGI arrives?
2019
GPT-2 — Preschooler level
Could barely string together coherent paragraphs. The baseline that showed language models had potential.
✓ Reached
2023
GPT-4 — Smart high schooler
Aces AP exams, writes sophisticated code, reasons through competition math. Benchmarks rapidly saturated.
✓ Reached
2024
PhD-level benchmarks cracked
GPQA (expert PhD questions) being approached. Claude 3 Opus at ~60% vs 80% for in-domain PhDs.
✓ In progress
2025–26
Outpace college graduates
Models able to perform at the level of a skilled recent grad across most professional tasks. Agent capabilities mature.
⟳ Approaching
2027
AGI — AI researcher level
Models capable of doing the work of an AI researcher/engineer. Could automate AI research itself, triggering intelligence explosion.
⟳ Forecast
2027–29
The Project begins
USG/national security involvement. Government AGI project launches. No startup handles superintelligence alone.
◌ Future
2029–30
Intelligence explosion
Hundreds of millions of AGIs automating AI research. A decade of progress compressed into 1 year.
◌ Forecast
2030s
Superintelligence
Vastly superhuman AI systems. Decisive economic and military advantage. The free world's survival at stake.
◌ Forecast
2030s
Trillion-dollar clusters
US electricity production grows tens of percent. Hundreds of millions of GPUs. Industrial mobilization unlike anything since WWII.
◌ Forecast
OOM Tracker — Effective Compute Scaleup
Orders of magnitude vs GPT-2 baseline
"With each OOM of effective compute, models predictably, reliably get better. If we can count the OOMs, we can extrapolate capability improvements."
— Leopold Aschenbrenner, Situational Awareness (2024)
Raw compute
~2 OOMs/yr
Algorithmic efficiency
~0.5 OOMs/yr
Unhobbling gains
~1 OOM total
GPT-2 → GPT-4 total
~5 OOMs
GPT-4 → 2027 (proj.)
+5 OOMs

Source: Epoch AI public estimates + Aschenbrenner analysis. Dashed bar = projection.

Frontier Labs — Status Overview
Best available public data · Updated June 2026
OpenAI
🇺🇸 USA
Flagship modelGPT-5.5 / 5.5 Pro
Annualized revenue$25B+ (est.)
NotableLeads FrontierMath T4
Release cadence~6 weeks/flagship
Anthropic
🇺🇸 USA
Flagship modelClaude Fable 5 (Mythos)
Annualized revenue~$19B (est.)
NotableNew tier above Opus
StanceSafeguarded frontier
Google DeepMind
🇺🇸 USA
Flagship modelGemini 3.1 Pro / 3.5
NotableGemini Spark agent
AdvantageTPU + multimodal lead
Context window2M tokens native
Chinese Labs (est.)
🇨🇳 China
Key playersDeepSeek, Qwen, MiniMax
Flagship modelsDeepSeek V4, Qwen 3.7 Max
StrategyOpen-weight + low cost
Gap vs frontier~3–6 months (est.)

The Scorecard — Every Prediction vs Reality

Source: Situational Awareness (June 2024) · Graded June 2026
Target Prediction Reality check (2026) Verdict
2025/26 Models outpace college graduates across knowledge work Frontier models at ~83% on knowledge-work benchmarks; agent products in production ✓ On track
~0.5 OOM/yr Compute + algorithmic scaling continues at trend One-year retrospectives find the pace roughly supported by evidence ✓ Holding
$500B/yr Massive AI capex acceleration Investment exceeding projections; accelerating faster than forecast ✓ Exceeded
— Open-source fades; proprietary algorithms create a durable US moat DeepSeek V4 / Qwen open-weight models within 3–6 months of frontier ✗ Wrong
2027 AGI: models do the work of an AI researcher/engineer Agentic coding strong (~80% SWE-Bench Pro) but autonomous research unconfirmed ⟳ Open
2027/28 US government launches formal AGI project National-security involvement growing; no formal Project announced ⟳ Open
2027–29 Intelligence explosion: decade of progress in 1 year Too early to grade ◌ Pending
2030s Superintelligence; decisive geopolitical advantage Too early to grade ◌ Pending

Source: situational-awareness.ai · Verdicts based on public retrospectives & benchmarks

Frequently asked questions

Was Aschenbrenner right about AGI by 2027?

Partially on track as of mid-2026. Compute scaling has roughly held, AI investment has exceeded his projections, and frontier models perform at or above the level he predicted for 2025/26. But full AGI — models autonomously doing AI research — remains unconfirmed, and his call that open-source would fade has been clearly wrong.

What is Situational Awareness?

A 165-page essay published in June 2024 by former OpenAI researcher Leopold Aschenbrenner, predicting AGI by 2027, an intelligence explosion to superintelligence by decade's end, trillion-dollar compute clusters, and a US-China race over AI. Aschenbrenner went on to found Situational Awareness LP, a hedge fund now managing over $5 billion.

How far behind are Chinese AI labs?

As of 2026, open-weight models like DeepSeek V4 and Qwen 3.7 Max trail the proprietary frontier by an estimated 3–6 months — much closer than Aschenbrenner's framework assumed, driven by aggressive open-weight releases and dramatically lower pricing.

Stay ahead of the curve

Weekly briefings on AGI progress, lab news, and milestone updates. No hype — just signal.