Key takeaways
- Cosine has pivoted its branding from the Genie agent to the Lumen specialist coding model family (Scout, Outpost, Frontier) — the Genie name no longer appears on cosine.sh
- June 2026: Cosine assembled a blue-chip UK coalition (BT, HSBC, Lloyds, NatWest, BAE Systems, Babcock, LSEG, PwC, Thales UK, Leonardo UK, Telefónica Tech) to co-design Lumen Sovereign, trained on the Isambard-AI supercomputer under the UK's £500M Sovereign AI programme
- Disclosed funding remains tiny (~$3M across two rounds, including a $2.5M seed) for a company now positioning as a frontier-model lab — a striking mismatch with its sovereign-AI ambitions
FAQ
What is Genie by Cosine?
Genie was Cosine's autonomous AI software engineer. As of mid-2026, Cosine has rebranded around the Lumen coding model family and a unified agent system spanning CLI, Desktop, and Cloud; the Genie name has been retired from its website.
How does Cosine compare to other AI coding agents?
Cosine previously claimed 72% on SWE-Lancer. Its current marketing claims Lumen Outpost scores 59.3% on its own "Niche-Bench," ahead of GPT-5.5 and Gemini 3.1 Pro — a self-published benchmark that has drawn skepticism before.
Can Cosine's models run on-premise?
Yes. Cosine offers public cloud, managed single-tenant cloud, and fully air-gapped deployment. The forthcoming Lumen Sovereign model is designed to deploy entirely within customer infrastructure with no external data transfer.
What is Lumen Sovereign?
Britain's first sovereign frontier coding model, announced June 2026. Co-designed with UK banks, defense primes, and telecoms, trained entirely on UK soil on Isambard-AI, targeting delivery in late 2026.
Executive Summary
Genie was Cosine's fully autonomous AI software engineer, once claiming the highest score (72%) on the SWE-Lancer benchmark[1] for production-grade coding tasks — ahead of OpenAI and Anthropic models, per Cosine's own marketing. As of June 2026, Cosine has repositioned: the Genie name has been retired from its website in favor of the Lumen specialist coding model family and a unified agent system spanning CLI, Desktop, and Cloud.[2][3] The headline development is Lumen Sovereign — Britain's first sovereign frontier coding model, co-designed with a coalition of UK banks, defense primes, and telecoms, and trained on the Isambard-AI supercomputer under the UK Government's £500M Sovereign AI programme.[4][5]
| Attribute | Value |
|---|---|
| Company | Cosine |
| Founded | 2022 (Alistair Pullen, Yang Li)[6] |
| Funding | ~$3M across 2 rounds, incl. $2.5M seed (SOMA, Uphonest Capital; Lakestar, Focal participating)[7][6] |
| Employees | ~12[6] |
| Headquarters | London, UK |
Product Overview
Cosine now describes itself as "the AI software engineering system for professional engineering teams that need maintainable, visible, reviewable work and deployment control," delivered as a unified agent across CLI, Desktop, and Cloud interfaces.[3] The agent structures work through five phases — Research, Plan, Implement, Verify, Handoff — emphasizing legible, reviewable output over full autonomy.
The Lumen model family powers the system:[2]
| Model | Description |
|---|---|
| Lumen Scout | 8B parameters, runs efficiently on-device |
| Lumen Outpost | High-quality model for production tasks |
| Lumen Frontier | Complex reasoning — "coming soon" |
| Lumen Sovereign | UK-sovereign frontier model, target late 2026[4] |
Cosine previously described itself as a "Human Reasoning Lab" — studying how humans perform tasks, then teaching AI to replicate that performance.[8] The Genie-branded autonomous agent (drafting PRs end-to-end with GitHub, Jira, and Slack integration) was the company's prior flagship; those integrations are no longer prominent in current product messaging.[3]
Launches Since Early 2025
| Launch | Date |
|---|---|
| Cosine CLI | September 2025[6] |
| VS Code extension | October 2025[6] |
| Lumen model family | 2025–2026[2] |
| Lumen Sovereign coalition | June 8, 2026[5] |
Technical Architecture
Cosine offers three deployment tiers to meet different enterprise security requirements:[2]
Deployment Options
| Option | Description |
|---|---|
| Public Cloud | Fully managed SaaS |
| Managed Single-Tenant Cloud | Private environment in a dedicated tenancy |
| Fully Air-Gapped | On-premise, no data egress, for strict security requirements |
Lumen Sovereign
- Trained entirely on UK soil on Isambard-AI, one of Europe's most powerful supercomputers[4]
- Built by continual pre-training and expert expansion of a 256k-context sparse Mixture-of-Experts open-weight base, targeting a long-horizon agentic coding model[5]
- Deployable entirely within customer infrastructure, including fully air-gapped environments, with no external data transfer[4]
- Coalition co-designers: BT, HSBC, Lloyds, NatWest, BAE Systems, Babcock, LSEG, PwC, Thales UK, Leonardo UK, Telefónica Tech[5]
Key Technical Details
| Aspect | Detail |
|---|---|
| Deployment | Public cloud, single-tenant cloud, or air-gapped |
| Models | Lumen Scout / Outpost / Frontier / Sovereign (proprietary) |
| Training services | Post-training on customer engineering data; legacy/enterprise languages |
| Open Source | No |
Strengths
- Sovereign-AI positioning — the only coding-model startup anchoring the UK's £500M Sovereign AI programme, with backing from BT, HSBC, BAE Systems, and other national institutions[5]
- Enterprise security — air-gapped deployment with no external data transfer; prior marketing cited SOC 2 attestation and ISO 27001 alignment
- Specialist-model thesis — Lumen Outpost claims 59.3% on Cosine's "Niche-Bench" vs. GPT-5.5 (48.3%) and Gemini 3.1 Pro (44.9%)[2]
- Legacy system support — post-training on internal codebases and enterprise languages (COBOL, Fortran)
- Visibility-first workflow — Research/Plan/Implement/Verify/Handoff phases keep agent work reviewable[3]
Cautions
- Funding mismatch — ~$3M in disclosed funding[6] is extraordinarily thin for a company promising a frontier model by late 2026; the sovereign-compute allocation offsets training cost but not operating runway
- Benchmark churn — the 72% SWE-Lancer claim that anchored earlier marketing has been replaced by a self-published "Niche-Bench"; Cosine's Genie-era SWE-Bench claims were never listed on the official leaderboard[9][10]
- Brand whiplash — Genie, the product this profile covers, has been quietly retired from the website; buyers evaluating "Genie" are now buying something materially different[3]
- Enterprise-only opacity — pricing, customer counts, and revenue remain undisclosed
- Delivery risk — Lumen Sovereign is a late-2026 target built by upcycling an open-weight base model; coalition co-design letters are not purchase commitments[5]
What Developers Say
Community discussion of Cosine peaked around the August 2024 Genie launch and has been thin since; no substantial Hacker News or Reddit threads on the Lumen rebrand or the June 2026 Sovereign announcement were found as of June 11, 2026. The Genie launch thread was notably skeptical:[9]
"I'm skeptical. Partially because if you go to swebench.com, you can see this company underreported results from their competitors like Amazon Q Developer. I've also seen plenty of other projects claim they've reached 30%+ on SWE-bench without verifying or posting their results on this site." — potatoman22, Hacker News, August 2024[9]
"Any external verification of the benchmark results?" — Y_Y, Hacker News, August 2024[9]
The absence of recent practitioner reviews — positive or negative — is itself a signal: Cosine's traction story currently rests on institutional coalition letters rather than developer word-of-mouth.
Pricing & Licensing
Pricing is not publicly available. Enterprise-focused with custom quotes based on deployment model and scale.[2]
Expected cost: Likely $500+/seat/month based on competitive positioning vs. Devin; air-gapped and sovereign deployments will be custom-priced.
Licensing model: Commercial, enterprise contracts
Competitive Positioning
Direct Competitors
| Competitor | Differentiation |
|---|---|
| Devin (Cognition) | Both autonomous engineers; Cosine now competes on proprietary specialist models and sovereign/air-gapped deployment |
| Factory | Both enterprise-focused; Cosine trains its own models, Factory uses third-party frontier models |
| Mistral / sovereign-AI labs | Lumen Sovereign moves Cosine into national-champion model territory, not just coding agents |
| Tembo | Tembo orchestrates multiple agents; Cosine is a single vertically-integrated agent + model stack |
When to Choose Cosine Over Alternatives
- Choose Cosine when: You need air-gapped or UK-sovereign deployment, have strict security/compliance requirements, or work with legacy codebases
- Choose Devin when: You want the established market leader with proven enterprise deployments
- Choose Tembo when: You need agent orchestration across multiple tools rather than a single autonomous agent
Ideal Customer Profile
Best fit:
- UK enterprises and public-sector bodies with data-sovereignty mandates
- Organizations needing fully air-gapped AI deployment (defense, financial services, healthcare)
- Teams with legacy codebases (COBOL, Fortran, proprietary languages)
- Companies requiring SOC 2, ISO 27001, or regulatory compliance
Poor fit:
- Individual developers or small teams
- Organizations comfortable with cloud-only solutions
- Budget-constrained teams seeking transparent pricing
- Buyers needing a shipping product today rather than a late-2026 model roadmap
Viability Assessment
| Factor | Assessment |
|---|---|
| Financial Health | Fragile — ~$3M disclosed funding against frontier-model ambitions[6] |
| Market Position | Distinctive — sole UK sovereign coding-model play, blue-chip coalition[5] |
| Innovation Pace | Active — CLI, VS Code extension, Lumen family, Sovereign program inside 12 months[6] |
| Community/Ecosystem | Weak — no open source, minimal developer discussion since 2024[9] |
| Long-term Outlook | High-variance — government-backed compute is a moat, but delivery and funding risk are real |
Bottom Line
Cosine has traded its Genie autonomous-agent story for a specialist-model story, capped by the Lumen Sovereign coalition — easily its strongest credibility signal to date, with BT, HSBC, BAE Systems, and the UK government's Isambard-AI compute behind it.[5] But the disclosed funding is seed-scale, the flagship model is a late-2026 promise, and the company's benchmark claims have shifted from SWE-Lancer to a self-published eval.
Recommended for: UK enterprises with data-sovereignty mandates, and security-sensitive organizations that need air-gapped coding AI and can tolerate roadmap risk.
Not recommended for: Individual developers, small teams, or any buyer who needs transparent pricing, community support, or a product whose identity won't shift underneath them.
Outlook: If Lumen Sovereign ships on schedule and converts coalition co-designers into paying customers, Cosine becomes the UK's national-champion coding lab. If it slips, a ~12-person team with ~$3M disclosed funding has very little cushion. Watch for a funding announcement — the current trajectory demands one.
Research by Ry Walker Research • methodology
Sources
- [1] SWE-Lancer Benchmark (OpenAI, arXiv)
- [2] Cosine Website
- [3] Cosine Product Page
- [4] Building Lumen Sovereign: Cosine Forms Coalition with UK Industry Leaders
- [5] Cosine secures industry backing for Britain's first sovereign frontier model (Tech.eu)
- [6] Cosine Company Profile (Tracxn)
- [7] Cosine Unveils "World's Best AI Software Engineer," Secures $2.5M Funding (Maginative)
- [8] Cosine About Page (archived)
- [9] Hacker News: Genie — Best AI Software Engineer
- [10] SWE-Bench Leaderboard
- [11] Cosine Jobs Page