Here’s a position that won’t win me friends in the group chats: the most interesting thing about Reflection AI right now isn’t the model. It’s that the model doesn’t exist yet, publicly, and the valuation has tripled anyway.
Axios has a scoop that a powerful new model from the Nvidia-backed startup is coming, and that it could reshuffle the AI race. I believe the first part. I review tools for a living, which means I spend most of my week in the gap between “coming” and “usable,” and that gap is where a lot of very expensive promises go to die.
What we actually know
The funding story is remarkable and easy to verify. Reflection raised $2 billion in October 2025 at an $8 billion valuation. It raised another $2 billion in March 2026, pushing the number to $20 billion. In April 2026, CEO Misha Laskin confirmed the latest financing closed at a $25 billion pre-money valuation. The New York Times framed the earlier round as a bid to compete with DeepSeek, and as another sign of investor fervor that hasn’t cooled despite the usual concerns.
Then there’s compute. In June 2026, CNBC reported a multiyear agreement with SpaceX potentially worth up to $6.3 billion. In July 2026, Bloomberg reported Nebius will sell $1 billion in AI capacity to the company. Add those up and you get a startup that has secured more guaranteed access to GPUs than most national research programs.
And the awkward part, per reporting from early March 2026: the frontier open-weight model at the center of the pitch still hadn’t shipped publicly, and Asimov, the company’s code research agent, was behind a waitlist. The release date and details of the new model are still undisclosed.
Why a reviewer reads this differently
Investors are buying a thesis. I’m waiting on an artifact. Those are not the same purchase, and they fail in different ways.
When I test a coding agent, the questions are mundane and unglamorous. Does it hold context across a real repo or just a toy one? What happens on the fifth tool call when the schema drifts? How much does a typical task cost, and does that cost stay stable under load? Can I run it where my data already lives? None of those questions are answerable from a valuation, a compute contract, or a scoop. They’re answerable from weights and an API, and neither has arrived in public hands yet.
That’s not a knock on Reflection’s ambitions. It’s an observation about sequencing. Capital and silicon come first now; the thing you can download comes later. That ordering has become normal, and I don’t think the industry has fully reckoned with how strange it is.
The open-weight angle is the part worth watching
If the company delivers a genuinely frontier open-weight model, it changes the math for people like me and probably for you. Open weights mean you can test the thing yourself instead of taking a benchmark chart on faith. You can run evaluations on your own data, fine-tune for your own domain, deploy without sending your codebase through someone else’s endpoint, and compare honestly against DeepSeek and the rest.
That’s the real stakes here, more than the race framing. The competitive pressure on closed labs from a credible American open-weight release would be significant. But the operative word is released. An open model behind a waitlist is a closed model with better marketing.
What I’d want on day one
When this does ship, the review checklist is pretty simple, and I’d encourage you to use the same one on any model launch, not just this one.
- Are the weights actually downloadable, and under what license? “Open” covers a wide range of terms.
- Can you reproduce the headline benchmark numbers on your own hardware? If not, why not?
- What’s the real cost per task on workloads you care about, including retries and failed tool calls?
- Does Asimov come off the waitlist, or does the agent story stay a demo while the base model carries the news cycle?
- How does it hold up on long-horizon work, where most agents quietly fall apart?
My honest read
I’d put Reflection in the category of “earned attention, unproven product.” The funding is real, the compute is contracted, the team has investors who’ve seen a lot of pitches and still wrote nine-figure checks. That’s meaningful signal. It is not a substitute for a model I can break.
So I’m curious rather than converted. If the release lands and it’s as strong as the scoop suggests, the open-weight tier gets a serious new entrant and a lot of tooling decisions get reconsidered, mine included. If it slips again, we’ll have watched a company reach $25 billion on a roadmap, which tells you something about this moment that no benchmark chart will.
Either way, I’ll be downloading it the day it’s available and reporting back with numbers instead of adjectives.
🕒 Published: