Reific launches from YC, says frontier models score under 50 percent

Most agent training teaches a model to match the answer a human expert would give. Reific launched on Y Combinator this week with a different plan. Its training environments feed a stream of real records to the agent on a clock. Each action the agent takes is then graded by the record of what came next. The YC launch page says the goal is a score "no human could follow in real time". A person cannot track thousands of signals at once. The launch had 16 votes in its first hours.

What the product does

Reific sells three families of environments: Timestep-Assist for personal agents, Timestep-Ops for industry, and Timestep-Lab for science and compute. Each environment plays back a real operating record. The agent reads part of the record, takes an action, and the next slice of the record grades that action. Reific says numeric outputs are checked by the record itself. Other outputs are graded by judges from a different AI lab, so the grader is not the model being trained.

The company says its own benchmark shows the gap. On the three families, the best frontier models each score under 50 percent. Reific states today's leaders on its page. Gemini 4 Argon leads personal agents at 31.6 percent. Claude Opus 5.5 leads industry at 43.1 percent and science and compute at 29.7 percent. GPT-6 Astra sits at 22.2 percent on science and compute. Reific's own line reads: "No model earns half of the possible reward in any family."

The 54 percent number

Reific says it trained an open model on its own environments. It then measured the result on tasks the model had not seen. The full-solve rate went from 25.6 percent to 39.5 percent, which Reific reports as a 54 percent gain. The company says the open model did not get worse on general benchmarks after this training. Those numbers come from Reific. No outside party has run the same test.

Who runs it

The YC page names Antanas Žilinskas as the founder. It says he was a published time-series researcher who simulated superconductors at CERN. He co-founded Entangl in YC's Summer 2024 batch, where he was CTO. Reific is in the Fall 2026 batch. It says it is backed by Y Combinator. The launch page does not state a round size, a named lead investor, a customer count or a headcount. The company site is at reific.co.

What the YC page does not say

Three gaps a buyer would ask about. The cost per training run, which Reific does not state. Whether the company sells data licenses to buyers, or only runs training for them. And how it finds the live records in the first place: Reific's site says it buys data from partners, but does not name them. A team that wants to try the environments has to ask about each.

A test on this product is simple. Pick a task the team already runs an agent against. Freeze the model, score it on Reific's environment, then train on the environment and score the same model again. If the gain holds on the team's own tasks, the vendor keeps its claim. If it does not, the 54 percent was a one-corpus number and did not travel.

Source

• Reific launch page, Y Combinator
• Reific website