Cosine

We are proud to announce Cosine's selection for the UK Government's @UKSovereignAI initiative. As one of the first companies selected, we will train, own, and deploy the first sovereign AI model built entirely on UK soil. Read more below.

AI coding is now 55% of enterprise inference, but it’s blocked by three core fears: security, cost, and control. At Cosine, we solved them: → Extreme security, even air-gapped → Lumen Outpost: Our specialized model for max efficiency and lowest cost → Real-time traceability

AI is reshaping software & services, but execution is challenging. With the £500M Sovereign AI fund backing UK startups like @CosineAI the UK has a critical window to lead. Join our private breakfast panel on June 9th in London to discuss the future of UK AI leadership ↓

One year ago, I joined @CosineAI as an intern. Back in January, I took a full-time Product Marketing role mid-dissertation. I went from feeling behind to achieving more than expected in 12 months. Been a great ride so far. Thanks to the team – looking forward to what's next.

Cosine is partnering with @AntlerGlobal for their Deploy hackathon at Google London on 10 June. Build production-grade software with our cutting-edge coding agent. Cash prizes. Hosted by @GoogleStartupUK and credits by @Lovable – with tracks from @OvermindLab, @paid_ai, Ralio, + Listo. Limited spaces. Grab yours: luma.com/j6wkjmkf

Niche language codebases deserve better support. Lumen Outpost – our latest coding model built for engineers – fixed broken tests in a niche Elixir codebase and got the full test suite passing in under a minute. Try now with Cosine CLI: brew install CosineAI/tap/cos

Hackday at @RamenClubHQ next week👷🛠️ Builders, side projects, demos, and access to the latest Cosine agent for anyone who wants to ship something cool. Dalston, London 👇 luma.com/4f5ce8yu

Lumen Outpost benchmarks reveal how targeted post-training transforms model capability, rivaling frontier models. Highlights: → Niche-Bench: Leads with 53.9% pass rate (e.g., +11pp ABAP) → Slop-Bench: Highest quality score (25.4%) → Vibe-Bench: Near-perfect agent alignment (96.3% Action) → Cost-Efficiency: Strongest profile at $7.90 per successful task Read the full report: cosine.sh/blog/lumen-out…

This was a great first step for us in terms of focusing on a specific set of problems within agentic coding - performance across prolific programming languages and amount of sloppy code the model writes we focused on both of these during our RL post-training phase

Lumen Outpost is available now via API and in Cosine. These results show our post-training stack can push models toward the frontier, forming the blueprint for our upcoming Lumen Sovereign work. Full release: cosine.sh/blog/introduci…

Lumen Outpost delivers exceptional results across our key benchmarks. Outpost is 3x more cost-efficient than GPT-5.5, while also outperforming frontier models in niche languages, generating more maintainable code, and aligning with our preferred behavioral vibe signals.

Introducing Lumen Outpost. AI models usually try to be everything to everyone. We did the opposite. Every training decision we made for our new model was centered on a single objective: writing exceptional code. Outperforming frontier models. Available now.

I’m such a diagram nerd. Watching a swarm of agents visualise itself in real time is very satisfying!

What does “Sovereign AI” actually mean? As one of the founding partners of the UK Government’s @UKSovereignAI fund, our CRO shares what it means to us at Cosine.

If a motivated adversary targeted your infrastructure today, how far would they get? Introducing Cosine Red Team: our AI-native red team that maps your complete attack surface and safely simulates persistent breaches. Get prioritized, actionable fixes: cosine.sh/red-team

We got tired of opaque coding agents. Cosine gives developers: → Real-time monitoring → Context transparency → Full tool visibility Get started now: cosine.sh

Yesterday I built a macOS native app with @CosineAI in only a few minutes while working on copy for our website. Just one of the small tools I've built to make my day easier. 😎

We shipped a live swarm view👀 Click any subagent to see its task, files touched, tokens burned, and current tool call. Waves show which agents ran in parallel and which had to wait their turn. Turns out you can debug an orchestrator by looking at it.

Over lunch @AlistairPullen and a college were saying how cool it would be to travel back to a time in their music app and be locked just to that time frame, just like your time machine backups. Well about 30 minutes in swarm mode in @CosineAI :