Surge AI is a bootstrapped human-data and RLHF leader for frontier labs (reported $1.2B 2024 revenue). In 2026 it built a large in-house benchmark and RL-environment program (DAYJOB, GDP.pdf/.xlsx, HANDBOOK.md, Chartography, CoreCraft, and the Tuesday Work Index composite), published post-training transfer studies, launched Post-Training Runs and an enterprise offering, and released sample harnesses under Apache-2.0. Anthropic reports results on Surge's GDP.pdf in its Claude Sonnet 5 system card, and Microsoft publicly credits Surge's expert raters. OpenAI's adoption of GDP.pdf is so far claimed only by Surge.
Forbes: 'Surge claims it has not yet raised any outside funding'. Chen owns an estimated 75% and is described as the youngest member of The Forbes 400. Net worth $18B as of 2026-10-05.
Why it matters: No disclosed outcome from the July 2025 raise talks. The company still describes itself as bootstrapped.
70 real-world spreadsheet tasks across 12 knowledge-work domains. The top score is Gemini 4 Argon at 38.3%. Released on GitHub (Apache-2.0) and on Hugging Face (surgeai/GDP.xlsx).
Why it matters: Extends the 'GDP' professional-document suite to Excel workflows.
130 tasks (50 healthcare, 80 finance). Tasks average 13.6 hours (healthcare) and 16.6 hours (finance) of expert time and are graded against binary expert rubrics. Claude Opus 5.5 passes 24.7% in healthcare and 23.9% in finance, while the median configuration passes 0.6% and 2.5%. Tasks run in containerised Harbor environments.
Why it matters: Surge's flagship long-horizon professional environment family. The blog positions it explicitly against OpenAI's GDPval (shorter prompts, larger working environments, longer horizons).
One epoch of RL on 1,700 Surge coding tasks (1,000 repository and 700 terminal tasks) with Kimi K2.7 Code improved all six external benchmarks. Examples: SWE-Bench Pro from 60.1 to 64.8 and Terminal-Bench 2.1 from 67.4 to 82.0.
Why it matters: Surge now publishes its own RL post-training runs and sells 'Post-Training Runs' as an offering.
Combines eight Surge benchmarks: Chartography, HANDBOOK.md, Antidote, Hemingway-bench, ComplexConstraints, GDP.pdf, CoreCraft and Riemann-bench. In the launch post the leader was Claude Fable 5 (Adaptive/Max) at 66.8, with GPT 5.6 Sol at 66.7. The live leaderboard now shows Claude Opus 5.5 at 83.9, ahead of GPT 6 Astra at 78.7.
Why it matters: A single headline number Surge can promote across its whole suite.
The dayjob (updated 2026-09-23), handbook (2026-09-25) and gdp-xlsx (2026-09-29) repos are Apache-2.0 and ship harnesses and sample tasks. Datasets are under the 'surgeai' org on Hugging Face. Full task sets for DAYJOB and HANDBOOK.md require contacting benchmarks@surgehq.ai. The gdp-pdf, chartography and complex-constraints repos show no license.
Why it matters: Data-ops teams can run Surge's environments using standard container tooling, without custom integration work.
Surge says it 'ran that pipeline end to end, including researcher recruitment, structured submissions, quality control, and expert review', and recruited 28 experienced technical AI safety researchers. Anthropic's announcement (2026-08-28) confirms a baseline of '28 human safety researchers' but does not name Surge or say who recruited them.
Why it matters: Shows Surge recruiting expert researchers for frontier-lab safety research. The figure matches Anthropic's, but Surge's role is self-claimed.
Microsoft wrote: 'We built a blind side by side human evaluation with one of our partners, Surge, using their pool of professional raters to measure various models on these traits.' The evaluation covered 1,276 tasks and compared MAI-Thinking-1 with Claude Sonnet 4.6.
Why it matters: A frontier lab names Surge as a partner in its own launch post, so Microsoft is now a primary-source-confirmed customer.
100 tasks pairing professional, domain-specific charts with questions, tested on 30 model configurations. The best configuration reaches 45.0% mean pass@1. Surge's live leaderboard now shows Gemini 4 Argon at 71.6%.
Why it matters: Adds a professional chart-perception eval to the Surge suite. Surge says frontier models score about 90% on CharXiv, so this one is far from saturated.
Training on 363 long-horizon office-workflow tasks, with no software-engineering tasks, raised SWE-Bench Pro pass@1 by 5.8 points according to the abstract. Surge's research page rounds this to 5.7.
Why it matters: Argues that non-coding enterprise environments also help train coding agents.
Post-training on Surge's long-horizon multi-tool tasks improved all five external evals in the paper: Toolathlon +9.6 pp, tau2-Bench +5.3, BFCL-V4 +3.5, SWE-Bench Pro +5.8, Terminal-Bench 2 +2.8.
Why it matters: Surge is publishing its own post-training results as evidence that its environments transfer to other benchmarks.
65 agentic tasks in five enterprise domains (Finance, Medical Billing, Insurance, Logistics, HR), each governed by a 20 to 124 page policy document. Under strict grading the strongest model passes 36.2% of trials. The GitHub repo (Apache-2.0) contains a 3-task sample, and the full set is available by contacting benchmarks@surgehq.ai.
Why it matters: An enterprise RL environment that tests whether agents keep to standing company policy over long tasks.
Expert-authored questions over professional PDFs. The paper abstract says the best model passes 30.7% of items and the worst passes 2%. Authors: Suhaas Garre, Emily Ritchie, Sushant Mehta, Edwin Chen.
Why it matters: Formal write-up of the benchmark Anthropic (confirmed) and OpenAI (Surge-claimed) have reported results on.
Surge says GDP.pdf appeared in the multimodal section of OpenAI's GPT-5.6 benchmark table, alongside MMMU Pro. OpenAI's page (openai.com/index/gpt-5-6/) returned 403 and could not be checked. The post's metadata shows 2026-07-09 and a 2026-09-16 re-render date. Surge's live GDP.pdf leaderboard does list GPT 5.6 Sol at 30.7%.
Why it matters: If accurate, OpenAI uses a Surge benchmark despite reportedly ending its data relationship in Sept 2025. This is self-reported by Surge and unverified.