04 / You are here
Where is AI right now?
Every few weeks a lab ships a new model and claims the top spot. This page zooms out: who is leading today, how far along the road to human-level AI we are, and what the evidence actually says.
SNAPSHOT · UPDATED OCTOBER 2, 2026The big picture
The race to
superintelligence
Think of it as six stages, from chatbots to machines smarter than any person. Tap a stage to see what it means and the evidence for where we stand.
The six stages are ARQIV's student-friendly version of the five-level ladder OpenAI reportedly uses internally (chatbots, reasoners, agents, innovators, organizations), plus a final superintelligence stage. Where the needle sits is our editorial judgment from the signals below, not an official measurement.
The scoreboard
Who's leading
right now?
There is no single scoreboard. Switch between three views and watch the leader change. That's the first lesson: "best" depends on how you measure.
The evidence
Four signals
we watch
Hype is cheap. These are measurements that are hard to fake, and the reasons the needle sits where it does.
Signal 1 · How long can AI work on its own?
From minutes to two-workday tasks in three years
Researchers at METR time how long a task takes a skilled human, then check whether an AI can finish it on its own at least half the time. The tasks are mostly software, machine-learning and cybersecurity work, and the measure is how long a task takes a person, not how long the AI runs. In METR’s corrected Time Horizon 1.1 data (as of September 2026, when METR stopped regular updates), that task length doubled about every 4.2 months from 2023 onward (likely range about 3.4 to 5.2 months), or about every 6.2 months over the longer series back to 2019. The longest estimates are the least certain.
Each dot is a model. The vertical scale multiplies: every gridline is a much bigger jump than the one below it. Values are approximate, from METR's time-horizon measurements (its current method, Time Horizon 1.1, which starts with GPT-4 in 2023). The arrow on the last dot means "at least": METR notes results above 16 hours are not yet reliable. This chart has reached the end of METR’s current test. Its tasks top out at about 16 hours, and newer models are too capable to measure with them, so there are no new points after April 2026. We’ll add new points if METR releases a new set of longer tasks.
Signal 2 · Brand-new puzzles
ARC-AGI-3
Video-game-like puzzles the AI has never seen, where it must figure out the rules by playing. Built to be easy for people and hard for AI.
Sept 2026. The test's creators say even acing it would not be "proof of achieving AGI." ARC Prize
Signal 3 · Making new discoveries
Finding what humans missed
Anthropic says its Claude Mythos model found security holes in every major operating system and web browser. Mozilla reported fixing 271 Firefox bugs it found. Some outside reviewers argue the claims are overstated.
Anthropic held Mythos back from the public in April 2026 because of these risks. Anthropic · Mozilla · Background
Signal 4 · What forecasters expect
General AI by…
The median guess of about 1,900 forecasters on Metaculus for the first "general AI" system, with the middle half of the guesses between 2028 and 2037.
Forecasts move fast and experts disagree widely. Metaculus
The latest flagships
The models
on the field
Most of the big labs shipped a new top model within a few weeks of each other in September 2026. Tap a card to open its profile: what it’s good at, independent test scores, what it works with and what it costs.
Gemini 4 Carbon
Oct 28, 2026 · Google’s new top model, built on Gemini 4 Argon; for now only some cyber-defense experts can use it. Leads the smarts index and the people’s vote.
Meta
Muse Wonder
Oct 12, 2026 · Meta’s first frontier model; its bigger “Full” version ranks fourth on the independent smarts index.
OpenAI
GPT-6.1 Sol
Sep 29, 2026 · Upgraded GPT-6 Sol: close to GPT-6 Astra for a fifth of the price.
Anthropic
Claude Opus 5.5
Sep 22, 2026 · Second on the independent smarts index, just behind Gemini 4 Carbon.
xAI
Grok 4.7
Sep 21, 2026 · Latest in the Grok 4 line.
DeepSeek
DeepSeek V4.1 Flash
Sep 10, 2026 · Low-cost model from China.
Alibaba
Qwen3.8 Max
Sep 2, 2026 · Alibaba’s flagship (China), latest update.
Moonshot
Kimi K3
Jul 16, 2026 · Moonshot AI’s open-weight flagship (China).
Think like a scientist
Why do the
scoreboards disagree?
Different tests measure different things. The "smarts test" (Artificial Analysis Intelligence Index) averages ten tests run by an independent group: real work tasks done as an agent, coding, knowledge and long-document reading, and hard science questions. The "people's vote" (Arena, formerly LMArena) asks thousands of people to pick which of two anonymous answers they like better. A model can be great at exams but less pleasant to talk to, or the other way round.
Settings matter. The same model can score very differently depending on the setup around it: how long it may think, which tools it gets and what it can remember. On ARC-AGI-3, GPT-6 Astra scored 62.7% in the organizers' standard setup, where it carries forward only the notes it chooses to keep. Using memory features OpenAI designed for it, which let it reuse its earlier hidden reasoning, it scored 99.9%.
Leads are short. In 2025–26 the top spot changed hands every few weeks. Being #1 today says little about next month.
Companies grade their own homework. Launch announcements pick the tests that make the model look best. Independent tests are worth more.
Put curiosity to work
Read it. Try it. Question it.
Explore new AI research, or take a paper-first investigation into your classroom.