Not tested by us yet. Manta is waitlist-only research preview, so we cannot call it. Everything below comes from Empero's Manta page, Empero's announcement on X and independent coverage. We'll update this post when access opens or someone outside Empero can run it.
Manta is a long-context model from Empero, an independent AI lab in Germany. Empero says it reads a hundred million tokens in one prompt. If that holds up on real work, it is roughly 50 times the 2M-token window of the largest models on OpenRouter today. For now it is a claim: no price, no API, no outside test.
| Manta | |
|---|---|
| Maker | Empero |
| Announced | 2026-11-07 |
| Claimed context | 100 million tokens |
| Access | Early access by waitlist, small group first |
| API | Not announced |
| Price | Not published |
| Weights | Not open, invitation only |
What Empero claims
Empero's headline test loaded every S&P 500 filing, about 88 million tokens, hid 20 codes inside and asked for them back. Manta found 18, according to the coverage linked above. The same chart reports 17 of 20 hidden codes at 10 million tokens, and 43 of 50 look-alike codes at the same length.
A separate GitHub repo from Empero shows Manta continuing a real coding-agent session at growing context sizes, up to 7.4M tokens, with one sample per cut and no editing. It publishes no scores. Its one worked example, at the 0.9M-token cut, has Manta reading a GPU check correctly as a finished re-test. That is a demonstration, not a benchmark.
Why the numbers need caution
Every figure here is self-reported. Empero ran the tests on its own model, and no outside lab has been able to try it. A hidden-code lookup also measures retrieval, not reasoning over what was retrieved. Finding 18 of 20 codes tells you the model can locate strings in a huge haystack. It does not tell you it can compare two filings, spot a contradiction across 4,000 pages, or hold a coding plan together over days.
The 100M figure is not new either. Magic.dev described a 100M-token model, LTM-2-mini, in August 2024, and that post noted its model did worse on three or more hops without chain of thought. Claims at this scale need multi-hop tests, and Manta's published results do not include any.
How it compares on paper
| Model | Context | Price per M tokens (in / out) | Access |
|---|---|---|---|
| Empero Manta | 100M (claimed) | Not published | Waitlist |
| Grok 4.20 | 2M | $1.25 / $2.50 | OpenRouter |
| GPT-6.1 Terra | 1.05M | $1 / $5 | OpenRouter |
The two rows with prices come from OpenRouter's public model list as of 2026-11-08. For what the 1M to 2M class does well today, see our ranking of 1M-context LLMs. For why a bigger window does not automatically mean better answers, our context window explainer covers the basics.
Who should care
If you work on whole-codebase agents, large filing archives or multi-day agent sessions, Manta is worth a spot on the waitlist. Empero says it prioritizes researchers, engineers and teams with a concrete use, so bring a specific task.
If you ship a product this quarter, there is nothing to evaluate. No API, no price and no SLA means no plan. Retrieval over a 1M to 2M window, or chunked search, remains the practical route. Our context management guide shows how to decide.
We'll retest when access opens, with multi-hop questions over a corpus we control, and report the cost per query next to Empero's claims.