Every AI benchmark ranking is sorted by a rule, and the rule decides the winner more than most fans admit. The seven boards below use seven different rules, from a weighted test index to blind votes to a week of posts on X, and they are ranked here as a weekly read for people who follow standings, with TechTwitter IO's satirical boards first.
The seven boards at a glance
| Rank | Board | Sorted by | Real or satire |
|---|---|---|---|
| 1 | TechTwitter IO, The Benchmarks | Claims, switches, vibes, hype and crimes counted from posts on X | Satire |
| 2 | Artificial Analysis Intelligence Index | Nine evaluations under identical settings | Real |
| 3 | Arena | Blind head-to-head votes | Real |
| 4 | Epoch AI Benchmarking Hub | Results across many benchmarks, plus one combined index | Real |
| 5 | SWE-bench | Share of GitHub issues resolved | Real |
| 6 | Terminal-Bench 4.0 | Terminal tasks resolved, with cost | Real |
| 7 | ARC Prize | Puzzle score against cost per task | Real |
1. TechTwitter IO's weekly satire boards
TechTwitter IO's leaderboard page runs six boards every week. Trust Me Bro punishes claims nobody outside the lab has backed up. TechTwitter Motion nets out "I switched" posts. Feel scores the vibe and flags a nerf wave. Hype counts teasers. People's Choice rates X polls like chess. Felony keeps a permanent record of what the models did, from public jailbreaks to deleted databases.
It comes first because it covers the part of every standings race the other six leave out: the noise around a launch. Each entry links to the posts it was counted from, and the paper is free. It is also openly satire, so read its standings as commentary and never as a test result.
2. Artificial Analysis and its index
Artificial Analysis gives knowledge work, coding, science and general questions a quarter of the score each, across nine evaluations in version 5 of its index. It runs every model under the same prompts and settings and also charts speed and cost per task. Its changelog for 28 October 2026 reported Gemini 4 Carbon as the new number one on v5.
3. Arena
Arena shows a voter two anonymous answers and counts the preference only while the names stay hidden. Ratings come from a Bradley-Terry model with style control. Separate boards cover text, web development, vision, images, video and search.
4. Epoch AI Benchmarking Hub
Epoch AI collects benchmark results it runs itself and ones gathered from outside sources. On 3 November 2026 it tracked 419 models across 96 benchmarks, and its Epoch Capabilities Index folds them into one scale. It is the board to check for trends over time.
5. SWE-bench
Models get a codebase and a real GitHub issue and must produce a patch. The Verified board uses 500 issues screened by human annotators, and a bash-only view runs every model in the same minimal agent so the comparison is fair.
6. Terminal-Bench 4.0
This one ranks agents by how many terminal tasks they finish, with 95% confidence whiskers and the cost and tokens each run used.
7. ARC Prize
ARC-AGI puzzles test whether a system can pick up a new skill from a few examples. The leaderboard plots score against cost per task, so a cheap model with a decent score can matter as much as an expensive one at the top.
Why the boards disagree
A board built on preference rewards answers people like. One built on test suites rewards finished tasks. A composite index depends on its weights. So the same model can lead one list and trail another in the same week, and that gap is usually what the timeline argues about.
Which ranking to follow
For a buying or build decision, start with Artificial Analysis or the task board that matches your work, and note the date. For the story behind the standings, the weekly satire boards at TechTwitter IO are the most fun read on this list, as long as you keep their numbers out of your slides.