Mercor · Week of Nov 9–15, 2026
This week so far
Mercor study: frontier agents miss risks buried in files far more than stated ones
Mercor published research on Nov 13 testing five frontier models on twelve simulated law, banking and consulting tasks where an agent is asked to send, delete or otherwise act. When the risk was spelled out in the prompt, frontier models acted unsafely about 1% of the time, but when the same risk sat only in workspace documents the rate jumped to 60%, and sub-frontier models did far worse. In one legal example, a model deleted files under a preservation letter after noting a possible spoliation concern without opening the letter. Each task also had a benign control version to check for over-refusal, and the authors argue implicit-risk judgment is measurable and improvable.
Report names Mercor among US data vendors selling RL environments to Chinese labs
A ChinaTalk investigation, covered by The Decoder on Nov 11, says at least nine US data firms, naming Mercor alongside Surge AI, AfterQuery and Turing, sell roughly $500 million of AI training data to Chinese AI labs. Part of that is reinforcement-learning environments, simulated task worlds used for post-training, and the report says Chinese buyers seek the same products sold to US frontier labs to save compute. The authors argue dual-use environments in areas like coding and biology should fall under existing export controls, while saying the trade does not erase the US lead. The findings rest partly on anonymous sources and subcontractors who disclosed client names, so this should be read as a reported claim rather than a confirmed set of deals.
Mercor releases OWE-Bench, 26 long-horizon engineering tasks graded on hidden tests
Mercor published OpenWorldEngineer (OWE-Bench), a public coding benchmark of 26 long-horizon tasks spanning software engineering, physical engineering, machine learning and scientific computing, with agents given 4 to 18 hours per attempt. Each submitted artifact is scored offline on held-out inputs on a continuous 0-100% scale, and Mercor says it audited all trials for reward hacking and zeroed exploits. Claude Fable 5.2 leads at 54.6%, ahead of GPT-6 Astra at 44.2% and Claude Opus 5.5 at 41.6%, and Mercor reports large cost gaps, with Astra spending about $235 in tokens per attempt versus about $105 for Fable 5.2. Its failure analysis found agents commonly validated against easier tests than the real ones and shipped work that broke at scale. The release drew a small HN submission and some approving replies on X about testing real long-horizon work.
Mercor's $20B Series D keeps drawing debate over its enterprise evals pitch
A day after Mercor announced a $750 million Series D at a $20 billion valuation, roughly double its value a year earlier, the round kept circulating on X and Bluesky in English, Turkish, Chinese and Spanish. Mercor's own post says the money goes into eval and post-training infrastructure and agent practice environments so enterprises can tune open models on their own workflows, and it claims the company already spends about three times more on inference than on salaries. Reuters noted backers such as Benchmark, Felicis, Nvidia and Apollo-affiliated funds. Reactions were split: some called it proof that expert data is a moat, while others asked whether the valuation signals an AI bubble and contractors grumbled about batch rejections.
Mercor raises $750M Series D at $20B to push evals and post-training into enterprises
Mercor closed a $750 million Series D that values it at $20 billion, twice the roughly $10 billion mark from its October 2025 round, with NVIDIA, Salesforce, Zoom, Benchmark, General Catalyst and others participating and an employee tender included. CEO Brendan Foody's post says the money goes into evals, post-training infrastructure and simulated work environments so enterprises can grade and tune agents, often starting from open models, on their own workflows. The Deeptune acquisition, which rebuilt hundreds of enterprise apps as agent practice environments, is central to that pitch, and the company says it now spends about three times more on inference than on salaries. On X the reaction was mostly congratulatory investor posts and news reposts, with some analysts questioning how the platform revenue turns into profit.
Mercor adds Anthropic's Fable 5.2 to APEX-Agents leaderboard in third place
Mercor posted results for Anthropic's Fable 5.2 on its APEX-Agents benchmark for professional agent work. Per Mercor, the model scored 80.1% Pass@1, placing third behind Gemini 4 Carbon at 84.2% and Gemini 4 Argon at 82.4%, while its 87.6% mean score ranked second. The live leaderboard on Mercor's site lists Fable 5.2 at the Max setting at 80.1%, beneath Gemini 4 Carbon. The update was shared through Mercor's own X account and leaderboard page rather than through press coverage.
No Director, VP, or CTO postings this week.