Manta
Empero, a German AI lab, announced Manta, a long-context model it says reads 100 million tokens in one prompt. Access is invite-only and nothing is independently verified.
Empero claims Manta holds about 100 million tokens of context in a single prompt. In its own recall test on 88 million tokens of SEC filings, it found 18 of 20 hidden codes. It also showed a 7.4 million token coding session as anecdotal examples, with no quantitative scores.
- Date
- Friday 6 November 2026
- Lab
- Empero
- Kind
- model
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Claimed context window | 100 million tokens company claim, not independently verified | company |
| Hidden code recall at 88M tokens (20 codes) | 18 of 20 (90%) test data was 346 MB, 1,269 S&P 500 10-K and 10-Q filings | company |
| Hidden code recall (50 codes) | 43 of 50 (86%) secondary test as reported by coverage | company |
Early access is by waitlist under a confidentiality agreement. Empero announced Manta on X as @EmperoAI on 2026-11-06, and the lab's page is empero.org/manta.
Sources
This record was partly confirmed: some claims could not be checked on 8 November 2026. How we check