General Reasoning

Market position

A small, early-stage research lab (headcount reported at ~10, within an 11-50 band) categorised by an RL-vendor directory as an open-source RL environment provider, listed alongside Prime Intellect and Good Start Labs as related vendors. The directory rates overall confidence in its profile as medium, notes no disclosed investors or valuation, and leaves revenue signals and security certifications unknown.

The company positions itself around long-horizon and non-stationary settings rather than single-turn benchmarks, combining benchmark design (KellyBench), an open environment standard and registry (OpenReward / Open Reward Standard) and point-in-time data infrastructure (BackSearch). The site cites prior team contributions at Meta — Galactica, Llama 2 and Llama 3 — plus research on thinking tokens, data-scarce training methods and early RL methods for LLM reasoning.

Technology

Reinforcement learning, multi-agent systems and environment design. Artifacts include the Open Reward Standard, an open specification connecting language models to community-built RL environments with 330+ environments behind one API; the KellyBench long-horizon, non-stationary evaluation environment; BackSearch, a frozen point-in-time news archive with search and fetch; and a GitHub organisation hosting environment repositories and the Apache-2.0 licensed firehorse harness.

Go-to-market

Distribution is primarily through open releases and public artifacts: an open environment registry and API (OpenReward), open specifications and repositories, published benchmarks, and a Hugging Face organisation (no public models or datasets listed as of the snapshot). Recruiting is handled through a public careers page listing three open roles.

Teams that need open, portable RL environments and long-horizon agent benchmarks they can run self-hosted or via managed/API hosting, per a third-party buyer analysis; the wider AI research community is also addressed through open releases.

Geography

An operating research hub in Shoreditch, London, United Kingdom, with the legal entity General Reasoning, Inc. registered in the United States; the SEC Form D lists an address at 2261 Market Street Ste 86270, San Francisco, CA 94114. A third-party profile records the company as not distributed/remote (estimated) and lists no other locations.

History

Founded in 2025, with a US legal entity, General Reasoning, Inc., registered in San Francisco and an operating research hub in Shoreditch, London. An SEC Form D filed 2025-07-11 reported $10,904,992 in equity sold. Public releases followed in 2026: OpenReward (March 2026), KellyBench (April 2026) and BackSearch (July 2026), with press coverage of the KellyBench results in April 2026.

Risks & controversies

No controversies are reported in the sources. Data limitations are notable: investors, valuation, revenue and SOC 2 / certification status are all unknown; the reported total raised derives from a single SEC Form D amount sold rather than a lifetime total, with no stage label. Organisations named in connection with OpenReward — NVIDIA, Nebius, Eigent, OpenAI and Meta — are described by the third-party profile as self-claimed environment providers/contributors on the company's own page, not verified paying customers. Separately, several unrelated entities share similar names (a personal technical blog at generalreasoning.com and a compliance-workflow company at genreason.com), creating identification risk.