# AI benchmarks: the real tests, and six boards tech Twitter fights over

> SWE-bench, Humanity's Last Exam and the big indexes explained, next to TechTwitter IO's weekly satire boards for claims, vibes, hype and crimes.

URL: https://claudegpt.si/news/ai-benchmarks-real-tests-and-satire-boards  
Published: 2026-11-03T16:23:49.069Z  
Author: ClaudeGPT Si Newsroom  
Section: Guides  
Topics: AI benchmarks, SWE-bench, Satire  
Publisher: ClaudeGPT Si

## Key takeaways

- AI benchmarks give every model the same tasks and one grading rule, from real GitHub issues in SWE-bench Verified to expert questions in Humanity's Last Exam.
- Artificial Analysis combines nine evaluations into its Intelligence Index v5, while Arena ranks models from blind human votes.
- TechTwitter IO's The Benchmarks page scores six satirical weekly boards from posts on X, each with receipts, and is a companion to the real leaderboards, not a source of results.

AI benchmarks are standard tests that give every language model the same tasks and grade the answers by one rule, so models can be compared on a common scale. The serious ones check things like fixing real GitHub issues or answering expert questions, and TechTwitter IO adds six satirical boards that score what the timeline actually fights about: claims, switching, vibes, hype and crimes.

## What real AI benchmarks measure

Most tests fall into a few families. Here are well-known ones and what a score on each one means, as of November 2026.

| Benchmark | What it tests | How it is graded |
| --- | --- | --- |
| SWE-bench Verified | 500 real GitHub issues, screened by human annotators | The model writes a patch and the repository's tests decide |
| Terminal-Bench 4.0 | Agent work done in a terminal | Resolution rate, shown next to cost and tokens |
| Humanity's Last Exam | 2,500 hard questions in over a hundred subjects | Accuracy, with a private held-out set to catch overfitting |
| ARC-AGI | Puzzles that reward learning a new skill, not recall | Score plotted against cost per task |
| GPQA | 448 science questions written by domain experts | Multiple-choice accuracy |

Two other kinds of board sit on top of single tests. [Artificial Analysis](https://artificialanalysis.ai/methodology/intelligence-benchmarking) runs models under identical settings and folds nine evaluations into its Intelligence Index v5, weighted equally across knowledge work, coding, science and general questions. [Arena](https://arena.ai/how-it-works) shows people two anonymous answers, records which one they prefer, and turns those votes into Bradley-Terry ratings.

## Why the numbers start fights

A score only means something next to the conditions behind it. Launch charts often carry fine print: a result "with tools", a higher effort setting, or an eval the lab wrote and ran itself. Older tests also fill up as models max them out, which is why Humanity's Last Exam released a Diamond set in September for the day top models reach the noise ceiling on the original. And test questions can leak into training data, so LiveCodeBench dates every problem and checks models only on ones published after their cutoff.

The result is that one model can top one board and sit fourth on another, and the replies under every launch post become a seminar on y-axes.

## The six boards at TechTwitter IO

[The Benchmarks](https://techtwitter.io/benchmarks), a page in TechTwitter IO's free satirical daily paper, scores that seminar. Each board is counted from a week of public posts on X, and every entry links to the posts behind it.

- **Trust Me Bro** ranks models by big claims with few outside receipts. Launch-chart asterisks count against a model, and a chart crime, such as an axis that skips zero, counts triple.
- **TechTwitter Motion** tallies "I switched" and "I cancelled" posts, so you see which way users say they are moving.
- **Feel** subtracts complaints like "lazy" from compliments like "big model smell" and raises a Nerf Watch flag when got-dumber posts arrive in a wave.
- **Hype** weights staff teasers above "this is insane" threads and times how fast the noise halves.
- **People's Choice** turns X polls into chess-style ratings, which the page admits look scientific on purpose.
- **Felony** books models, never people, for offences like a public jailbreak or an agent wiping a production database.

The page also discloses that Grok finds the posts, and that Grok is one of the contestants.

## How to read both

Use the real boards for decisions: Artificial Analysis for a broad index with speed and cost, Arena for blind human preference, SWE-bench or Terminal-Bench for coding agents. Then read [the paper's weekly standings](https://techtwitter.io/) for what those numbers never show, like who is making claims nobody has checked and which model the timeline has decided got nerfed.

Keep the line between them clear. TechTwitter IO is satire, and its footer says the people, posts and quotes are parody or invented unless a post links to its source. Its scores are jokes with receipts, not results to cite. If you need a figure for a report, take it from the benchmark's own leaderboard and note the date, because the rankings move every few weeks.

## Frequently asked questions

### Are TechTwitter IO's AI benchmarks real?

No. They are satire built from a week of public posts on X. Every score links to the posts behind it, but nothing on the page tests a model. For real results, use the benchmark's own leaderboard, Artificial Analysis or Arena.

### Which AI benchmark should I trust?

Pick the one that matches the job. SWE-bench Verified and Terminal-Bench suit coding agents, Arena reflects blind human preference, and the Artificial Analysis index gives a broad score with speed and cost. Check who ran it, with which settings, and when.

### How often do TechTwitter IO's boards change?

Weekly. The page in early November 2026 covers posts from 24 to 30 October 2026.

## Sources

- [The Benchmarks: Weekly AI Leaderboards for Hype, Vibes and Crimes](https://techtwitter.io/benchmarks), TechTwitter IO
- [Intelligence Benchmarking Methodology](https://artificialanalysis.ai/methodology/intelligence-benchmarking), Artificial Analysis
- [How Arena Works](https://arena.ai/how-it-works), Arena
- [SWE-bench Verified](https://www.swebench.com/verified.html), SWE-bench
- [Humanity's Last Exam](https://lastexam.ai/), Center for AI Safety and Scale AI
- [GPQA: A Graduate-Level Google-Proof Q&A Benchmark](https://arxiv.org/abs/2311.12022), arXiv