Cosine

Cosine describes itself as The Sovereign AI Lab: it trains AI models and specialized coding agents for organizations that need frontier capability inside secure environments. Founded in 2022 and backed by Y Combinator, it is now based in London and positions around regulated and sovereign deployment rather than the open benchmark race it first became known for. Its Genie agent drew wide attention in 2024 for topping SWE-bench at roughly 30 percent, the highest by any company at the time; the Genie branding has since been retired and the company now leads with the Lumen model family.

Lumen is a family of coding models rather than a single one. Scout is post-trained from Devstral 123B and runs cheaply and on-device for the work around the coding loop. Outpost is post-trained from Kimi K2.6 and handles everyday production implementation, trained with behavioral reinforcement learning specifically against the failure modes developers complain about: messy patches, duplicated logic, unnecessary abstraction, and agents that claim a task is finished before it is.

Sovereign, aimed at frontier-scale reasoning, is announced but not yet shipped, targeted for late 2026 and to be trained on UK soil using the Isambard-AI supercomputer under the UK Government's sovereign AI program. The models are proprietary and post-trained for enterprise and niche languages including COBOL, Fortran, Verilog, Rust and complex SQL. Cosine evaluates them on three internal benchmarks of its own design: Niche-Bench for legacy-language robustness, Slop-Bench for maintainability and Vibe-Bench for collaboration.

The agent reaches developers through two surfaces. The CLI installs via Homebrew and runs terminal-native with a local-to-remote execution model, moving between planning, implementation and review modes. Every agent turn is a lightweight git commit, so any change can be undone instantly.

A live todo list tracks complex jobs, Memory persists project conventions and architecture decisions across sessions, language servers supply symbol-level understanding for references, definitions, renames and diagnostics, and Swarm mode spawns specialized child agents that break a task down and execute in parallel. Cosine Cloud runs multiple tasks at once in a shared project across engineers, product managers and stakeholders, and remote agents keep the terminal local while long-running work executes in the cloud. MCP connections are auto-installed with zero configuration, covering GitHub, Jira, Postgres, Figma, Stripe, Slack and Linear.

The company is explicit about not locking customers into its own models: the CLI ships a menu of twenty-one models across eight providers, with per-model credit multipliers ranging from Lumen Scout at 0.1x to GPT 5.5 and Claude Opus at 2.75x.

Deployment is the commercial center of gravity. Cosine offers fully managed public cloud, a managed single-tenant environment with enterprise controls, and fully air-gapped installation inside the customer's own perimeter, with Scout available on-device. That posture underpins an industry coalition co-designing Lumen Sovereign whose members include BAE Systems, Babcock, BT, HSBC, Lloyds, LSEG, NatWest, PwC, Thales UK, Leonardo UK, QinetiQ, Fujitsu, Deloitte, Vodafone, EY and The Alan Turing Institute.

Pricing is credit-based across three published tiers: Starter at $19 a month with 4M credits, Team at $199 with 47M, and Enterprise at $999 with 240M, with add-on credits priced per million and enterprise or private deployment scoped through sales. Credits are consumed by agent work, model calls and cloud execution, so cost varies with task size and model choice. Notable gaps for a vendor selling into regulated industries: no security attestation, trust center, SSO or RBAC documentation appears anywhere on the site, and no public API or SDK for embedding Cosine is published.