As part of my set of forecasts on all the major competitors to the frontier AI labs, Reflection AI is maybe the least-well covered that has a real chance of achieving a breakthrough. It was founded in early 2024 by two DeepMind researchers, Misha Laskin and AlphaGo author Ioannis Antonoglou, and spent its first year in stealth building coding agents. It has since raised about $4.6 billion and repositioned as America's open frontier lab, the answer to DeepSeek.
Its March Series C at a $25 billion pre-money valuation got so little coverage that Wikipedia still lists the October 2025 round as its latest, which is why I called it a sleeper among the "neolabs," the billion-dollar startups now challenging the big AI labs. It is the fastest-hiring lab on that bench, around 230 people, up from roughly 60 last fall. (Remember when OpenAI had a hard cap of 150 employees?)
My forecasting did not reveal a very promising picture, though. The flagship open-weight model, Beam 1.0, shipped on October 16 and scores 30 on the Artificial Analysis index, tenth among open-weights models. Meanwhile the burn is spectacular, roughly $150 million a month renting compute on SpaceX's Colossus 2 campus under a lease either side can end on 90 days' notice, plus a $1 billion Nebius deal.
I asked a series of date questions about when Reflection would achieve certain milestones, from which I derived my overall view of the company.
| Question | Forecast |
|---|---|
| Flagship open-weight model release date | Shipped Oct 16, 2026 (forecast Jan 2027, Oct 2026 to Oct 2027) |
| A Reflection model is the #1 open-weights model before 2028 | 11% |
| Annualized revenue run rate, end of 2027 | $60M ($5M to $380M) |
My core timeline behind Reflection AI:
- Early 2024. Laskin and Antonoglou leave DeepMind and found Reflection, spending the year in stealth on autonomous coding agents.
- March 2025. Launch with $130 million across a seed and Series A (Sequoia, CRV, Lightspeed), at about a $545 million valuation.
- October 2025. A $2 billion round at $8 billion, with Nvidia among the backers, and the repositioning as America's open frontier lab.
- March 2026. The quiet Series C, roughly $2.5 billion at $25 billion pre-money, with JPMorgan's Security and Resiliency Initiative among the reported participants.
- May 2026. The DOE Genesis Mission agreement across 17 national labs, and Pentagon network agreements alongside other leading labs.
- June 2026. Training compute comes online on SpaceX's Colossus 2 campus, rented at roughly $150 million a month on a lease either side can end with 90 days' notice.
- July 2026. The $1 billion Nebius deal, headcount reaching about 230, and the CTO repeating later-this-year guidance for the flagship.
- August 2026. The White House finalizes the open-weight security-review exemption. Still no model.
- October 2026. Beam is unveiled in early access on October 5, and the Beam 1.0 weights land on Hugging Face on October 16 under Apache 2.0, a 501-billion-parameter mixture-of-experts. The FT reports Nvidia in talks to deepen its investment or buy the company (Reuters).
The release-date median landed on January 25, 2027, just past "later this year." The 80% interval ran from mid-October 2026 to October 2027, and Beam 1.0 shipped on October 16, at its early edge. Two things pushed my August forecast late. Reflection has been on rolling later-this-year guidance since February, and The Information reports the plan is a first model this year with the larger version in early 2027. That matters because this question resolves on the flagship, not a warm-up release, so a staggered launch pushes the date right. The Colossus 2 compute also only came online in June, and nobody compresses a frontier pretraining run plus post-training into a quarter. One tailwind the research surfaced: the White House finalized an exemption for US open-weight models from pre-release security reviews, removing a policy blocker I would have priced in.
Can it take the open-weights crown? 11%, that a Reflection model holds the #1 open-weights slot on Artificial Analysis at any point before 2028. The crown currently belongs to MiMo-V2.6-Pro at a score of 46, and Beam 1.0 debuted sixteen points behind it. First models from new labs essentially never top the global board. What holds 11% up is that the crown has changed hands every month or two, resolution requires only a single instant at #1, and Reflection is explicitly building for this exact leaderboard with more compute than any first-time open-weights lab has ever had.
The revenue median is $60M annualized by the end of 2027, with a wide interval, $5M to $380M. What surprised me is the pipeline behind the fat right tail: a DOE Genesis Mission agreement spanning 17 national labs, classified Pentagon network agreements, a Dell AI Factory channel, and a 250MW Korean sovereign AI factory with Shinsegae. All memoranda, none priced. Sovereign and federal deals are lumpy, they can add $100M overnight or convert to nothing. The median case has Reflection tracking Mistral's curve, roughly $312M ARR two and a half years in, but starting later. Against a $150M monthly burn, even the good revenue scenarios are not the point. The model is.
The bench gave Reflection a frontier median of October 2031, within striking distance of SSI despite far less capital, precisely because open weights make its progress visible. Visibility is also the risk.