AfterQuery's agent training data lifted OpenAI's small model from 3.1% to 17% on Terminal-Bench
The November 4 run, showcased on Thinking Machines' Tinker, sits beside rival Mercor's proof on the same page, giving labs a direct comparison of measured benchmark uplift for the first time. It is the claim beneath AfterQuery's reported $3.2 billion valuation, now competing in public for labs' data budgets.
Richard Tangthree-minute read
On November 4, 2026, AfterQuery published the kind of result AI labs normally see only in sales meetings 12. It post-trained gpt-oss-20b, OpenAI's$1.18T — OpenAI, private, latest valuation $1.18T small open-weight model, on its terminal-agent data, and the model's pass rate on Terminal-Bench 2.0, a benchmark of real command-line work, went from 3.1% to 17.0% 12. The recipe: supervised fine-tuning on successful agent trajectories, then reinforcement learning in Harbor sandbox environments that reward finished work 2.
The host matters as much as the number. Thinking Machines Lab$13.50B — Thinking Machines Lab, private, latest valuation $13.50B, founded by former OpenAI CTO Mira Murati, put the run on the customer showcase of Tinker, its training service, where AfterQuery's biggest rival appears beside it 2. The company has published uplift numbers on its own blog before 6. This one is on neutral ground.
Validation stopped being the lab's job
Tinker rents the training loop itself: open-weight models from a few billion to a trillion parameters, adapted with LoRA, billed by the token 2. That is the hop that changed the data business. A vendor with no compute of its own can now do what a lab researcher does, train a model on its data and measure what moves. AfterQuery co-founder Spencer Mateega told Forbes in April that this internal pipeline is how the company proves quality before labs ever see the data 4. This month the proof went public.
The rival is on the same page
Mercor~$2B — Mercor, private, latest valuation ~$2B, the larger rival, trained Nvidia's−4.64% — Nvidia, down 4.64 percent today Nemotron 3 Ultra on more than a thousand banking, law and consulting tasks and doubled the model's first-attempt success on the held-out APEX-Agents benchmark 2. Two vendors, one claim, measured the same way, on a page any buyer can score.
Two data vendors, one showcase page: both publish a before-and-after gain
Data vendors now compete on measured uplift, side by side where buyers can score them
- Before training
- After training
Data
| Before training | After training | Change | |
|---|---|---|---|
| AfterQuery: gpt-oss-20b (Terminal-Bench 2.0) | 3.1% | 17.0% | +448.4% |
| Mercor: Nemotron 3 Ultra (APEX-Agents) | 11.5% | 22.9% | +99.1% |
The proof is about data, not a deployed agent
Read the result honestly: a 17% pass rate is evidence about data quality, not a tool anyone can put to work, and the strongest agents now clear roughly two thirds of the tasks on Terminal-Bench's current version 3. The open link is transfer, whether data that lifts a small open model also lifts a frontier training run. No vendor has published that, and the first to show it sets the market's price.
The money is already positioned
The round Forbes reported on September 1 values AfterQuery at $3.2 billion; it is reportedly profitable, with revenue its co-founder puts in the hundreds of millions 45. Mercor is reportedly in talks with Nvidia at a $20 billion valuation 4. Nvidia sits on the demand side twice: its Nemotron models were trained with AfterQuery data, and it may bankroll Mercor 45. Behind both vendors are paid experts recording how they actually work, and their judgment is the scarce input 6.
The vendors that measure what they sell deserve those budgets, because the measuring is what turns expertise into something a lab can buy. The next number to watch is an uplift attached to a frontier model. Until it appears, anyone can keep score.
Get Model by email, free
The AI labs, their models, and the business behind them. In our daily email, every weekday.
