The pitch is who holds the pen
Most deployed AI models do not automatically update their underlying weights when they make mistakes in production. A law firm that wants a model to review contracts according to its own partners' standards typically needs a separate post-training process, often involving specialist researchers or outside consultants. Trajectory wants to make that process accessible to the people who understand the work itself.
The customer supplies what Trajectory calls a value function: a way to score model outputs against the standards that matter for its particular work. The platform then handles the post-training process, evaluation and deployment. The proposition is that domain experts can define what a good result looks like without having to build the underlying training infrastructure themselves. The company calls its product the first self-serve platform for post-training on production data.
The approach also moves responsibility toward the customer: if the scoring criteria reward the wrong behavior, post-training can make a model better at producing exactly the wrong result.
That is a real gap in the market, and the founding team is not a random one. The company has 18 people. Its three founders are Stanford classmates: Ronak Malde and Michael Elabd worked at Google DeepMind, with Malde among the researchers DeepMind acqui-hired from Windsurf, and Arjun Karanam worked on generative foundation models at Apple. Sequoia led a $45 million Series A at a $300 million valuation, with NVIDIA, Bessemer Venture Partners and Bain Capital Ventures taking part. Builders at Harvey, Rogo, Mercor and NVIDIA are already training on the platform, the company says.
Read the launch post in order
The order of the launch post matters. The two capabilities that carry the argument come with a label attached. Continual Grading, which learns a customer's standards from a few graded examples and applies them across production data, is a research preview. Group Size 1, which trains on individual production interactions so a model keeps improving after it ships, is also a research preview. What is generally available today is the base loop: connect a model, supply tasks with a way to score them, train in one click, deploy in one click.
The ease claims do not quite agree with each other, either. The company's homepage says the platform takes just two code changes to get started. Its launch post says about five lines of code. Neither figure is the point on its own, and both are the company's own. But a reader comparing the two pages learns something useful about how much of the launch is finished product and how much is a direction.
Pricing would settle more of that question, and there is none to check. No price was listed on Trajectory's public site as read on November 2, 2026. A self-serve product with no published price is, for now, a product whose cost is learned in a sales conversation.
One quantified benchmark, three company claims
The launch offers one quantified result. Working with NVIDIA and Harvey, Trajectory says it took a Nemotron Lightning model from 0% to an 8.3% all-pass rate on Harvey's LAB benchmark, matching Opus 5.5's score at 100x lower cost and 3.8x faster speed. It also says it post-trained Nemotron 3 Ultra on frontier legal tasks within 24 hours of that model's release, and that one unnamed customer trained more than eight models for long-horizon agentic tasks in two weeks.
Take the 8.3% apart before repeating it. All-pass is a strict measure: the reported result means 8.3% of the evaluated cases met the benchmark's full passing criteria. That is an improvement from the reported zero baseline, but it also means the model did not achieve an all-pass result on most evaluated cases. The benchmark belongs to Harvey, a customer. The matching score, the 100x cost figure and the 3.8x speed figure are Trajectory's claims, with no independent evaluation linked in the launch post.
The launch post makes a broader comparison than the earlier results Trajectory published in August. It now says the 8.3% result matches Opus 5.5, at 100 times lower cost and 3.8 times faster speed. These are new company-reported comparisons, not independently verified results. The launch post does not link to a separate evaluation establishing the cost and speed methodology, or provide enough detail to reproduce the comparison. They should be read as Trajectory's claims rather than an independently established performance advantage.
The competitive framing underlines the bet. Trajectory identifies the field it is entering as Thinking Machines Lab and its Tinker platform, Applied Compute, valued at $3.25 billion last month, and Prime Intellect. An 18-person company at $300 million is not being valued against those names on current evidence. It is being valued on the possibility that the person who knows the work, rather than the lab that built the model, ends up controlling how models improve. That is a coherent thesis. On November 2, it is still a thesis with one benchmark behind it.
None of that makes the claims false. It does mean the launch offers limited independent evidence: one quantified benchmark comparison, alongside two additional customer or deployment claims without independently published evaluation results. The $300 million valuation is therefore being announced alongside a product thesis and early customer adoption, not a publicly reproducible performance record.
Sources
- [1] Trajectory — "Own Your Value Function: Introducing the Trajectory Platform," trajectory.ai, November 2, 2026 (read first-hand — $45M Series A led by Sequoia at a $300M valuation, with NVIDIA, Bessemer and Bain Capital Ventures participating; "the first self-serve product for post-training on production data"; Continual Grading and Group Size 1 both labeled research previews; builders at Harvey, Rogo, Mercor and NVIDIA already training; one unnamed customer trained more than eight models in two weeks; a Nemotron Lightning model taken from 0% to an 8.3% all-pass rate on Harvey's LAB benchmark, matching Opus 5.5's score at 100x lower cost and 3.8x faster speed; Nemotron 3 Ultra post-trained on frontier legal tasks within 24 hours of its release, working with NVIDIA and Harvey; "about five lines of code")Read source
- [2] Trajectory — "Continual Learning Cheap Enough to Meter," field report, trajectory.ai, August 11, 2026 (read first-hand — the company's earlier published result on the same benchmark: post-trained Nemotron 3.5 Lightning reached 8.3% on Harvey LAB, up from 0% at baseline, above Opus 4.6 at 6.6% and post-trained Nemotron 3 Ultra at 5.8%; the November launch post's Opus 5.5, 100x cost and 3.8x speed comparison is broader than this earlier report and links no methodology for it)Read source
- [3] Trajectory homepage, trajectory.ai, read November 2, 2026 (read first-hand — "takes just two code changes to get started"; Continual Grading and Group Size 1 shown as research previews; testimonials from NVIDIA, Clay, Decagon and Harvey; SOC 2 certified; the signup entry point is presented as "Join the waitlist"; no pricing listed)Read source
- [4] Marina Temkin, "Trajectory raises $45M to make AI model training accessible to nontechnical experts," TechCrunch, November 2, 2026, 23:13 UTC (read first-hand — 18-person company founded about a year ago by three Stanford classmates: Ronak Malde, Michael Elabd and Arjun Karanam; Malde among the researchers Google DeepMind acqui-hired from Windsurf, Elabd also worked at DeepMind, Karanam worked on generative foundation models at Apple; customers include Decagon, Mercor and NVIDIA alongside Harvey and Rogo; the competitors Trajectory identifies: Thinking Machines Lab's Tinker, Applied Compute, valued at $3.25 billion last month, and Prime Intellect)Read source