A German research lab called Empero announced a model on November 6 that reads a hundred million tokens in a single prompt. They call it Manta. A hundred million tokens is roughly ten times the biggest context window on the market today, and the announcement raises an obvious question: even if the model works, what does one prompt of that size cost? Empero has not said. That blank is the most important number in the announcement.

First, what Empero actually did. The lab ran a test and published the data. They collected every annual and quarterly filing that S&P 500 companies submitted to the SEC — 1,269 documents, 346 megabytes of text, 88 million tokens — and hid twenty short memos inside them at random depths, each memo containing an access code for an office. Then they asked Manta to return the codes, with the model’s reasoning mode turned off. It found eighteen of twenty. The two misses were not invented numbers; they were the real codes from other memos, Tulsa confused with Tucson. The whole dataset is on GitHub with a script that recomputes every figure, so anyone can check the claim.

That is more evidence than most context announcements ever provide, and it still proves less than the headline implies. Finding a planted memo in a pile of filings is a retrieval test — the memo looks nothing like the annual reports around it, which makes it easy to spot. It shows Manta can find a specific fact inside a very long input. It does not show the model can reason across that input, connecting a clause in one filing to a footnote in another. Empero says those results come with the public release.

Empero also ran a second, more interesting test. They took one real session of a coding agent — about 7.5 million tokens of commands, logs, and fixes — cut it at five points, and asked Manta to write the session’s next step having read everything up to that point, with no summary. The original agent that produced the session had compressed its own context into a summary thirteen times, because it ran out of room around the million-token mark. Manta just kept reading. It is one anecdote with no score attached, but it demonstrates the selling position: an agent that never forgets what it already tried, instead of one that compresses its own memory and hopes the summary kept the right parts.

Does the context really all stay in memory? That is Empero’s pitch — everything you give it, it keeps, no retrieval layer, no summaries — and it is the part nobody can verify yet. The model is not public, the architecture has not been described, and the weights have not been released. Early-access users sign an agreement not to publish outputs before the public release. What has been demonstrated is recall at 88 million tokens and one anecdote at 7.5 million. Whether Manta holds every token in some working memory or does something smarter internally that only looks like that, Empero has not explained. The claim and the mechanism are two different things, and only the claim has evidence so far.

Then there is the cost, which Empero has not published — no price on the Manta page, no rate card, no response time for an 88-million-token request. The arithmetic from comparable systems suggests it will not be small. Attention — the mechanism that lets a model relate any token to any other — costs work that grows with the square of the input: ten documents on a desk get checked against each other 45 ways, ten thousand documents get checked against each other about 50 million ways. And memory: when Magic.dev built their own hundred-million-token model in 2024, they published the numbers for holding 100 million tokens of context on standard transformer infrastructure — about 51 terabytes of key-value cache, or roughly 638 H100 GPUs per user. Magic avoided that bill with a custom architecture. Empero has not said what Manta does about it.

One honest correction on scale, since a hundred million tokens is hard to picture. It is about 750 novels — which sounds small until you remember a novel is only about a hundred thousand tokens, so a window that holds 750 of them also holds a ten-million-line codebase or a company’s entire filing history. The 88-million-token test set Empero used is a good concrete anchor: every annual and quarterly report filed by every S&P 500 company, in one prompt.

So the summary is straightforward. A small lab claimed a hundred-million-token context, and backed it with checkable recall data — which puts them ahead of most labs making this claim. What remains unproven is whether the model reasons over all that text rather than just finding things in it, whether the “keeps everything” pitch reflects how the model actually works, and what one full-length prompt costs in dollars and time. Until Empero publishes the price and a response time, those are the three questions to watch.