Lisan al Gaib's 2026 AI forecast, checked: 8 of 14 test targets met, 3 within a point
Mostly accurate and often too cautious: four targets fell by more than 20 points, two were missed. Seven weeks remain.
On 1 January 2026 the X account Lisan al Gaib (@scaling01), who runs his own test, LisanBench, posted the scores he expected AI to pass during the year on 14 benchmark lines. By 11 November eight count as met, three sit within a point of the line, two are missed and one cannot be scored. The table runs from the biggest overshoot to the biggest miss.
| Test | His target | Best so far |
|---|---|---|
| FrontierMath Tier 4 | 75% | 100% |
| GSO | 65% | 88.2% |
| Cybench | 70% | 93% |
| ARC-AGI-3 | 60% | 81.2% |
| WeirdML v2 | 85% | 93.6% |
| SimpleBench | 90% | 94.9% |
| FrontierMath Tiers 1–3 | 95% | 95.1% |
| ARC-AGI-1 and -2 | 95% | 98.5%, 95.0% |
| Epoch Capabilities Index | 175 | 174.2 |
| SimpleQA Verified | 95% | 94.0% |
| Terminal-Bench 2 | 90% | 87.4%, retired |
| SWE-bench Verified | 90% | 83.5% |
| Humanity’s Last Exam | 80% | 54.8% |
| GDPval, wins only | 70% | no figure |
His 80% stretch target for ARC-AGI-3 fell too. He also expected METR’s time horizon (the length of task an AI agent finishes half the time) to pass 24 hours. The last published figure is 17.4 hours, and METR stopped updating the measure in September.
He gave Gemini 4 an 80% chance of release in 2026, Claude 5.5 70% and GPT-6 60%. All three shipped by mid-October.
What it means
The forecast was fairly accurate, and where it was off it was mostly too cautious: 11 of 14 targets are met or within a point, and half were passed outright. We count Terminal-Bench 2 as met although its best score is 87.4%: in July its makers moved to a harder version 3, writing that many tasks had become saturated, and the post had said these tests would get saturated. Of the two misses, OpenAI stopped reporting SWE-bench Verified in February after finding flawed tests. Humanity’s Last Exam is the one large miss on a test still in use, 25 points short.
Caveats
- Terminal-Bench 2 is counted as met by judgement, not by score: its best result is 87.4% ± 1.8% against a 90% target, on version 2.1 (84.7% on the original 2.0 task list). Counting scores alone, 7 of 14 targets are beaten.
- This is a count on 11 November. Seven weeks remain, and one new model could flip the three near-misses, including ARC-AGI-2, which stands at exactly 95.0% against a target of more than 95%.
- Two figures come from the labs themselves. Cybench’s 93% is from Anthropic’s report on Claude Opus 4.6. OpenAI’s latest GDPval figure, 84.9%, counts ties as wins; the last wins-only figure in Epoch AI’s data is 49.7%, for GPT-5.2 in December 2025.
- METR’s 17.4 hours has a range of 8.5 to 55 hours, and METR says its results above about 16 hours are unreliable. No model released after April has a score.
- The post also forecasts markets, company valuations and elections. This site tracks AI capabilities, costs and footprint, so those are not scored.