Nov 2025 · 7 months agoModel
Claude Opus 4.5 answers within a week
Six days after Gemini 3, Anthropic retakes the coding and agentic frontier. The lead at the top now changes hands in days, not years.
THE MODERN AI ERA · 2012 → TODAY
The Race to Superintelligence
Eight frontier labs. One question: who reaches the top first? Every model below is plotted by when it shipped and how capable it was — watch the race climb toward the frontier.
Watch the race ↓Latest milestones
A quick read of the newest events before diving into the full chart.
Nov 2025 · 7 months agoModel
Gemini 3 takes the lead
Google's strongest jump since Gemini 1.0 tops reasoning and multimodal boards and ships into Search's AI Mode on day one.
Nov 2025 · 8 months agoOpen
Kimi K2 Thinking stuns the leaderboards
Moonshot AI's open-weight trillion-parameter reasoning MoE matches or beats closed frontier models on agentic benchmarks — the DeepSeek shock, repeated.
The capability race
Each line is a lab; each dot is a model release. Height is its Capability Index? at launch. Hover or tap any dot for detail.
Highlight
Axis
2012
2026
drag to play through time · ←/→ step releases
Standings
| # | Lab | Top model | Index | Gap | Released |
|---|
Methodology
How the Capability Index is curated and how the timeline is maintained.
Scoring criteria
The index is a 0–100 editorial composite of frontier standing at launch, weighing reasoning, coding, knowledge, multimodality, tool use and visible adoption signals.
Update process
Release dates come from primary announcements, papers or project repositories. New models are added when they materially shift the competitive frontier or define a new open baseline.
Caveats
This is not an official benchmark, leaderboard or safety evaluation. It compresses many dimensions into a single line so the broad race is easier to inspect.
Example read
GPT-4 sits high because it created a clear 2023 frontier step-change; DeepSeek-R1 sits high because it brought reasoning performance into open-weight competition.
Use the data
The full curated dataset behind this page — labs, models, landmarks and sourced milestones. Free to cite and reuse with a link back.
Milestones that bent the curve
The releases and moments that defined each leg of the race — every entry sourced.
Deep Learning Era2012 – 2016
Deep neural networks crack computer vision on GPUs — the breakthrough that kicked off the modern AI boom, years before the Transformer.
- Sep 2012ResearchAlexNet ignites the deep-learning revolution#Krizhevsky, Sutskever and Hinton train a deep convolutional network on two GPUs and win ImageNet by a landslide — a 15.3% top-5 error, ~10.8 points clear of the field. Learned features beat hand-engineering, and the modern AI era begins — co-author Ilya Sutskever later co-founds OpenAI.
Foundations2017 – 2021
The architecture and the scaling laws that set the stage.
The ChatGPT Era2022 – 2023
AI goes mainstream and the frontier becomes a multi-lab race.
Reasoning & Agents2024 – 2025
Models learn to think before they answer — and to act.
- Aug 2024PolicyThe EU AI Act becomes law#The first comprehensive, risk-based AI law puts legal guardrails on frontier models — governance becomes a real constraint on how fast and how openly labs can ship.
- Oct 2024ScienceAI takes two Nobel Prizes in one week#Hopfield and Hinton win Physics for neural-network foundations; a day later Hassabis, Jumper and Baker share Chemistry for AlphaFold and computational protein design. The field's founders become laureates.
- Jan 2025OpenDeepSeekThe DeepSeek moment#An open reasoning model rivalling o1, trained for a reported fraction of the cost, erases ~$1T of market value in a day and resets assumptions about who can compete.
ASI Watch2025 →
Agents that work for hours; the lead measured in months.
- Jul 2025ScienceGoogleGold at the Math Olympiad#Gemini Deep Think officially reaches gold-medal standard at the IMO — with OpenAI reporting the same — solving competition proofs in natural language at elite human level.
- Nov 2025OpenKimi K2 Thinking stuns the leaderboards#Moonshot AI's open-weight trillion-parameter reasoning MoE matches or beats closed frontier models on agentic benchmarks — the DeepSeek shock, repeated.
No milestones match — try a broader search or clear the filter.
Standings, story mode, compare — and fresh data
Charted Grok 4.1, GPT-5.1, Gemini 3 Pro and Claude Opus 4.5, and added 11 sourced milestones through late 2025 (Sora, the Nobel week, computer use, MCP, o3 on ARC-AGI, IMO gold, gpt-oss, Kimi K2 Thinking). New features: a live standings table, story captions on replay, two-lab compare, an MMLU axis toggle, keyboard stepping, share links, JSON/CSV downloads and an Atom feed. Fixed chart end-labels, deep links vs autoplay, sticky-header overlap and the Capability Index tooltip.
Extended the timeline back to AlexNet (2012)
Added a new "Deep Learning Era" opening with AlexNet, and plotted AlexNet and the 2017 Transformer as foundational landmarks on the capability chart — whose time axis now compresses the early years and stretches the fiercely competitive recent ones.
Added methodology, model sources, SEO files and accessibility fixes
Every model dot now has a source link, a permalink and keyboard-accessible tooltip details.
Launched the capability-race timeline
Initial static version with frontier-lab lanes, model dots, milestone filters and source-backed events.
About this timeline
What you're looking at
A focused history of the modern AI era — from the 2012 deep-learning breakthrough and the 2017 Transformer to the frontier models of today — framed as a race between the labs pushing toward artificial general, then super, intelligence. Every milestone links to a primary source.
The Capability Index
The vertical axis is an illustrative 0–100 composite of each model's frontier standing when it launched, blending reasoning, knowledge and coding ability. It's hand-curated for narrative clarity — a way to see the race — not an official benchmark. Ship dates are real.
Why "Race to ASI"?
Because that's the stated goal. Several of these labs were founded expressly to build superintelligence — and the gap between them at the top of the chart is now measured in months, not years.