MODELSMETR TASK HORIZON (50%)
▲16–20 hrs
doubling every ~4 months (was ~7) · as of 2026-01 (Time Horizon 1.1)
The single best public timeline metric: how long a human-expert task frontier agents can complete half the time. At the current doubling rate, week-long autonomous work arrives within ~12 months — the threshold where agentic labor becomes an economic line item.
METR Time Horizon 1.1 · AI Digest tracker
MODELSSWE-BENCH VERIFIED (TOP)
▲96%
saturated — signal moved to SWE-bench Pro (69.2%) · as of 2026-08-02
Frontier models have effectively solved the standard coding benchmark. Saturation itself is the signal: real-world coding automation is no longer capability-limited, it is deployment-limited.
SWE-bench Verified leaderboard · SWE-bench Pro
CAPITALPREDICTION-MARKET AGI ODDS
▲25% by 2029
50% by 2033; 'weak AGI' by end-2026 · as of 2026-02 (Metaculus)
The crowd's number — and the spread between it and insider claims (2027–28) is the tradeable disagreement this desk exists to referee.
Metaculus AGI question · 80,000 Hours timeline review
MODELSFRONTIER-LAB REVENUE RUN-RATE
▲~$69B (Anthropic)
$9B Dec-25 → ~$69B Jul-26; OpenAI ~$25B, flat since Feb · as of 2026-07
Revenue is deployed capability. Anthropic's ramp — $9B to ~$69B run-rate in seven months, passing OpenAI in April — is one of the fastest in enterprise-software history and means businesses are paying for AI labor at scale now, not in a forecast.
Sacra (Anthropic revenue) · Simon Willison ($47B, May) · SaaStr (passed OpenAI)
COMPUTECOMPUTE BUILDOUT
▲~$725B (2026)
big-4 hyperscaler capex, +77% vs 2025's $410B · as of 2026-07
Capex is a hard-to-fake commitment to the timeline: Amazon ~$200B, Alphabet ~$175–185B, Microsoft ~$120–190B, Meta ~$115–135B for 2026 alone. Power contracted today is capability delivered 2027–2029 — the physical layer confirms or falsifies the software story.
Yahoo Finance ($725B plan) · CNBC (capex scrutiny) · Statista chart
APPSROGUE-AGENT INCIDENT LEDGER
▲82% of firms
reported unexpected agent behavior, trailing 12mo · as of 2026-06 (Gravitee)
Our proprietary dataset-in-progress: every public incident of agents acting off-instruction (OpenClaw mass-deletion, Feb 2026, is entry #1). Incident frequency tracks real-world autonomy deployment — and prices the safety discount.
Gravitee survey · OpenClaw incident
MODELSAI SHARE OF AI RESEARCH
▲70–100% at labs
Anthropic ~70–90% company-wide; Google 75% of new code; Microsoft ~30% · as of 2026
The recursive self-improvement indicator — the mechanism every fast-timeline scenario runs on. Top Anthropic/OpenAI engineers now report AI writing essentially all their code, and ~90% of Claude Code is written by Claude Code. When the research loop closes, every other indicator accelerates together.
Fortune/Yahoo (lab engineers, 100%) · DevOps.com (Google 75%)
COMPUTECOST PER TASK (CONSTANT CAPABILITY)
▼5–40× cheaper/yr
GPT-4-level output: ~$20 → ~$0.40 per M tokens since late 2022 · as of 2026
Capability tells you what is possible; cost tells you when it deploys. Epoch finds prices for a fixed performance level falling 5–40× per year depending on the benchmark — faster than PC compute or dot-com bandwidth ever improved. Falling cost at constant capability is what turns benchmarks into layoffs and margins.
Epoch AI (inference price trends) · MIT FutureTech (algorithmic efficiency)
MODELSINSIDER-STATEMENT LEDGER
◆2027–2030
insider median vs. market's 2033 · as of 2026-08
Every named-insider timeline claim, timestamped and scored on resolution. Current spread: Kokotajlo 50% by ~2029-30; lab leadership privately 2027–28; Metaculus 50% by 2033. The 4–5 year disagreement is the largest mispricing candidate in any asset class.
Kokotajlo interview (primary source) · AI Futures Project update
LABORLABOR DISPLACEMENT SIGNAL
◆US 4.2% (Jun)
payrolls +57k (cooling); 22–25yo in AI-exposed roles −13% since 2022 · as of 2026-06 (BLS)
The lagging indicator everyone watches and misreads. Fast-timeline scenarios predict headline unemployment stays quiet until automation of research completes — so the tell is entry-level hiring in exposed categories (already down double digits for the youngest cohort), not the headline rate.
BLS Employment Situation (Jun 2026) · CNBC jobs report · CBS (entry-level cohort)