What is Cosine?
Cosine is a London company of about 30 people that trains its own coding models, called Lumen, and runs a coding agent on top of them in a terminal app (Cosine CLI), a desktop app and Cosine Cloud, where long jobs keep running after you close the laptop. Every agent turn is saved as a small git commit you can undo, and the agent can split work across subagents and pull in tools such as GitHub, Jira, Linear, Slack and Postgres.
There are two Lumen models today: Scout, post-trained from Mistral's Devstral, for cheap search and mapping work, and Outpost, post-trained from Kimi K2.6, for the coding itself, with extra training on less common languages such as COBOL, Fortran and Verilog. A third, Lumen Sovereign, is being trained on the UK's Isambard-AI supercomputer with backing from a coalition that includes BAE Systems, HSBC, BT and NatWest, with deployment targeted for the end of 2026. The agent is not tied to Lumen: GPT, Claude, Gemini, Kimi, Qwen and GLM models are on the menu too.
How much does Cosine cost?
| Plan | Price | What you get |
|---|---|---|
| Starter | $19/month | 4M credits; $6.50 per 1M extra |
| Team | $199/month | 47M credits; $5.00 per 1M extra |
| Enterprise | $999/month | 240M credits; $4.50 per 1M extra |
Per team, not per seat: Starter $19/month (4M credits, extra credits $6.50 per 1M); Team $199/month (47M credits, $5.00 per 1M); Enterprise $999/month (240M credits, $4.50 per 1M). Private and air-gapped deployments quoted by sales. No free plan or trial on the current pricing page.
Is Cosine worth it?
Most coding agents send your code to an American model provider, and for a bank, a defence contractor or a hospital trust that is often the end of the conversation.
Cosine is built for that conversation. It trains its own coding models and offers air-gapped installs, where your code, prompts and outputs stay inside a boundary you approve. In June 2026 it gathered 13 large British organisations, including BAE Systems, HSBC, Lloyds and the Alan Turing Institute, to help design Lumen Sovereign, a model being trained on the government's Isambard-AI supercomputer with 500,000 GPU hours. That model is planned, not released, with a target of late 2026.
For everyday use the agent is sensible. Each turn is a git commit, so undoing a bad step is one command. Lumen Scout costs a tenth of a credit unit and Outpost a quarter, against 2.25 for Claude Opus 5.5 and 5.5 for GPT-6 Astra on the CLI's own table, so staying on Cosine's models stretches your credits roughly 9 to 22 times further. Pricing is per team, so adding a colleague to a paid plan costs nothing.
The catch: the evidence is mostly Cosine's own. Its only published score for Outpost is 53.9% on Niche-Bench, a benchmark Cosine wrote, against 48.3% for Kimi K2.6, the model it was built from. An independent index counts six Product Hunt reviews and none on G2. Its own documentation contradicts its pricing page: the FAQ still describes the old Genie agent, a free plan and per-task pricing, and the billing docs name plans that are no longer sold. The jump from $19 to $199 a month also leaves a gap for a three- or four-person team. And Cosine's air-gapped page publishes no hardware requirements or certifications; it says data-flow details are confirmed in a deployment review.
Our take: Cosine is a sovereignty bet first and a coding agent second. For a UK or European organisation that cannot send code abroad, it belongs on the shortlist next to Qodo's air-gapped reviewer; for a startup happy with the cloud, Claude Code or Cursor gives more for the money with far more people to ask when it breaks.
Where it earns its place
A platform engineering lead at a UK bank
Give developers a coding agent without code leaving our data centre
Start with Cosine's architecture review for an air-gapped install running Lumen or an approved open-weight model; price and timeline are scoped by sales. Ask for SOC 2 or ISO reports up front, as none are published.
A maintainer of a Fortran climate model at a research institute
Get useful agent help on a language most models handle badly
Lumen Outpost is post-trained on less common languages including Fortran and costs a quarter of a credit unit per call. Starter at $19 a month gives 4M credits; extra credits are $6.50 per 1M.
A six-person agency team
Share one coding agent subscription across everyone
Team at $199 a month gives 47M credits with no charge per extra member. Staying on Lumen models rather than Claude or GPT stretches those credits about nine times further on the CLI's multipliers.
Before you commit
- Setup
- `brew install CosineAI/tap/cos`, then `cos login` and `cos start` in a repository; `cos init` writes cosine.toml and AGENTS.md. Desktop and Cloud use the same account.
- Integrations
- MCP servers, including GitHub, Jira, Linear, Slack, Figma, Stripe and Postgres; language servers for type and symbol information; Lumen plus GPT, Claude, Gemini, Kimi, Qwen, MiniMax and GLM models.
- Data and privacy
- Air-gapped installs keep source code, prompts and outputs inside the boundary you approve, with exact data flows confirmed in a deployment review. Cosine's pricing and air-gapped pages state no certifications and no policy on training with customer code.
- Cost at scale
- Enterprise is $999 a month for 240M credits per team; heavier use costs $4.50 per 1M credits. Private and air-gapped deployments are priced separately by sales.
- Limits
- Third-party frontier models burn credits more than 20 times faster than Lumen Outpost; documentation lags the product; Lumen Sovereign is planned for late 2026, not available.
What’s good
- Air-gapped and private deployments, with its own Lumen models
- Lumen Scout and Outpost cost a tenth and a quarter of a credit unit on the CLI
- Every agent turn is a git commit you can undo
- Priced per team, not per seat
- Runs GPT, Claude, Gemini, Kimi, Qwen and GLM models too
What’s not
- Only benchmark is Cosine's own Niche-Bench
- FAQ and billing docs contradict the pricing page
- Nothing between $19 and $199 a month
- No published certifications or air-gapped hardware requirements
- Lumen Sovereign not yet released
Best for
UK and European organisations in defence, finance or the public sector that need a coding agent and models they can run inside their own network.
Cosine alternatives
Cosine: common questions
- Is Cosine worth it?
- Cosine is the coding agent to watch if you need it to run inside your own walls, but outside that niche it is a well-made tool with thin evidence and confusing paperwork. We rate it 3.5 out of 5.
- How much does Cosine cost?
- Per team, not per seat: Starter $19/month (4M credits, extra credits $6.50 per 1M); Team $199/month (47M credits, $5.00 per 1M); Enterprise $999/month (240M credits, $4.50 per 1M). Private and air-gapped deployments quoted by sales. No free plan or trial on the current pricing page.
- What are the downsides of Cosine?
- Only benchmark is Cosine's own Niche-Bench; fAQ and billing docs contradict the pricing page; nothing between $19 and $199 a month; no published certifications or air-gapped hardware requirements; lumen Sovereign not yet released.
- Who is Cosine best for?
- UK and European organisations in defence, finance or the public sector that need a coding agent and models they can run inside their own network.
- How much does Cosine cost for a team or in production?
- Enterprise is $999 a month for 240M credits per team; heavier use costs $4.50 per 1M credits. Private and air-gapped deployments are priced separately by sales.
- Is Cosine safe for company data?
- Air-gapped installs keep source code, prompts and outputs inside the boundary you approve, with exact data flows confirmed in a deployment review. Cosine's pricing and air-gapped pages state no certifications and no policy on training with customer code.
- What are the best Cosine alternatives?
- The closest alternatives we have reviewed are Claude Code, Cursor, Devin (incl. Devin Desktop, formerly Windsurf), Factory (Droids) and Mistral Vibe (formerly Le Chat).
Answers checked Nov 2, 2026.