Built for engineering teams buried in tickets rather than solo coders wanting autocomplete, Cosine replaces manual triage with an agent that clears GitHub and Jira backlogs unattended. Backed by the UK government's 500 million pound Sovereign AI programme, it suits regulated teams needing on-premises or air-gapped code generation, not casual hobbyists.
Cosine is a London AI lab whose coding agent scored 30% on SWE-bench in 2024, then rebuilt its models into the Lumen family for 2026. It reads a GitHub, Jira or Linear ticket, plans a fix, writes and tests the code, then opens a pull request, deployable cloud-only or fully air-gapped.
Maker: Cosine · Autonomy: semi autonomous · Maturity: GA
Underlying models: Custom (proprietary): Lumen family (Scout, Outpost, Sovereign)
About Cosine
Cosine is a London based AI lab founded in 2022 by Alistair Pullen and Yang Li. Pullen previously built and sold Fancy, a YC-backed marketplace acquired by GoPuff; Li spent his career scaling Luno, Ziglu and Mobike through IPOs and acquisitions before earning a Forbes 30 Under 30 spot. The pair started Cosine to fix a narrower problem than general purpose chat assistants: getting AI to write production grade code for tickets a team has not gotten to yet, not just autocomplete lines in an editor. Cosine has raised $15M in total, according to the Financial Times, backed by investors including Lakestar, SOMA Capital and Gaingels, and now by the UK government's Sovereign AI programme. Cosine's coding agent runs on its in-house Lumen model family (Scout, Outpost, and an upcoming Sovereign variant), trained on real production code rather than general web text, which is how it handles COBOL, Fortran, Verilog and complex SQL that general purpose coding models often get wrong. A ticket arrives from GitHub, Jira, Linear, or Slack; the agent retrieves the relevant parts of the codebase, plans a fix, writes and runs the code, then opens a pull request for review. It can run unattended or with a developer checking in at each step, and Cosine says it optimizes for readable, reviewable diffs rather than raw volume of generated code. The clearest fit is engineering teams with a backlog bigger than their headcount, using Cosine to close small feature requests, bug tickets, and refactors a senior engineer would otherwise have to triage one by one. It also suits organizations maintaining legacy systems in COBOL or Fortran, where few human engineers are available and general purpose coding assistants are weak. A newer audience is regulated industries piloting Cosine's air-gapped deployment, where no code or data leaves the customer's own network, a requirement mainstream coding assistants rarely meet. Cosine sells monthly plans priced on a usage-credit system rather than per token, scaling from solo developers up to enterprise teams with custom air-gapped contracts; exact tiers and limits are covered in the pricing FAQ below. The agent is reachable from a command-line install, a VS Code extension (currently macOS and Linux only, with Windows support still pending), a desktop app, and a cloud dashboard, so teams can assign it work from whichever surface they already use. The most significant recent shift is strategic rather than a single feature release: in April 2026 the UK government named Cosine part of its first cohort of Sovereign AI startups, and by June 2026 Cosine had assembled a coalition of major UK banks, telecoms and defense primes to co-design Lumen Sovereign, a frontier coding model trained entirely within the UK and targeted for deployment by the end of 2026. That puts Cosine's roadmap increasingly at the intersection of AI coding tools and national AI infrastructure policy, a positioning few coding-agent competitors share.
Pricing
No free tier as of 2026. Starter $19/mo (4M credits), Team $199/mo (47M credits), Enterprise from $999/mo (240M credits); extra credits run $6.50 down to $4.50 per 1M depending on tier. Fully air-gapped or single-tenant enterprise deployments are quoted separately.
Key Features
- Ticket-to-PR pipeline: Picks up a ticket assigned from GitHub, Jira, Linear, or Slack, retrieves the relevant code, writes and tests a fix, then opens a pull request without further prompting.
- Production-code-trained models: The Lumen model family trains on production codebases instead of general web text, extending support to niche languages including COBOL, Fortran, and Verilog.
- Air-gapped and single-tenant deployment: Runs in Cosine's cloud, a dedicated single-tenant instance, or a fully air-gapped environment with no data leaving the customer's network, the deployment tier used by its 2026 bank and defense design partners.
- Multi-surface access: Works from a terminal install, a VS Code plugin, a native desktop client, or the web dashboard, letting a team assign the same ticket from whatever surface is already open.
- Usage-credit billing: Bills on a credit system tied to agent work, model calls, and cloud execution time rather than per token, so cost scales with task complexity instead of raw tokens generated.
Strengths
- Publicly benchmarked at 30% on SWE-bench in 2024, a score that beat Cognition's Devin and Factory's public results at the time, though Cosine has not published an equivalent figure for its newer Lumen models.
- One of the few coding agents banks and defense contractors will actually pilot: its air-gapped deployment option is why HSBC, Lloyds, NatWest and BAE Systems signed on as Lumen Sovereign design partners.
- Optimizes for readable, reviewable diffs over raw output volume, which reviewers cite as a reason it produces less throwaway code to clean up than some rivals.
Weaknesses
- Cosine has not published a SWE-bench or comparable benchmark score for Lumen, the model family that replaced its Genie-era coding agent in 2026, so buyers must judge current performance on an older result from a different model generation.
- The VS Code extension supports macOS and Linux only as of 2026; Windows support is still listed as coming rather than shipped.
- Reviewers report Cosine costs more and includes less general-purpose editor tooling than IDE-first competitors like Cursor, so teams that mainly want inline autocomplete may be better served elsewhere.
Frequently Asked Questions
How much does Cosine cost in 2026?
Cosine sells three monthly plans billed on usage credits rather than per token: Starter at $19 a month with 4 million credits, Team at $199 a month with 47 million credits, and Enterprise from $999 a month with 240 million credits. Extra credits can be purchased on any plan, at rates from $6.50 per million on Starter down to $4.50 per million on Enterprise. Fully air-gapped or single-tenant deployments for regulated industries are priced outside these tiers through a custom agreement.
Is Cosine free to use?
No. As of 2026 Cosine has no free tier or free trial; every plan is a paid monthly subscription starting at the Starter tier. Some third-party directories still list an older 80-task free plan, but that offer no longer appears on Cosine's current pricing page.
What are the best alternatives to Cosine?
Claude Code suits teams already inside Anthropic's ecosystem who want a terminal-first agent with file and shell access. Cursor is a better fit for developers who want fast in-editor autocomplete and chat rather than an asynchronous ticket-to-PR workflow. Cognition's Devin is the closest positioning to Cosine on paper, though it is not yet listed on HokAI, and independent tests put its public benchmark figure below Cosine's result.
How does Cosine compare to Cursor in 2026?
Cursor is an editor-first AI pair programmer: a developer stays in the loop for every suggestion, and pricing is built around per-seat access to fast autocomplete and chat inside a familiar IDE. Cosine works the opposite direction, taking a ticket and returning a finished pull request with no developer present for the middle of the work, and it additionally offers air-gapped deployment that Cursor does not. Teams that want to close a backlog unattended should lean toward Cosine; teams that want a faster typing experience for developers already writing the code themselves should stay with Cursor.
How do you get started with Cosine?
Sign up at cosine.sh/login, then connect a GitHub, Jira, or Linear account so Cosine can see your ticket queue. Install the CLI with a single Homebrew command or add the VS Code extension (macOS or Linux) to work from an editor instead. Assigning a ticket to Cosine for the first time, typically a small well-scoped bug, is the fastest way to see a real pull request within the same working session.
Top Alternatives
- Claude Code: Pick Claude Code if you want a terminal-first agent tied to Anthropic's models; pick Cosine if you want ticket-to-PR automation with air-gapped deployment options.
- Cursor: Pick Cursor if you want fast in-editor autocomplete with a human driving every change; pick Cosine if you want tickets resolved into pull requests unattended.