Daily Briefing
June 25, 2026 · 5 items (site) · 7 items (base)
🔥 Headlines
01
Cognition AI raises $1B at $26B: Devin now writes 89% of its own code
The team behind Devin, the AI agent that codes for you, just closed a Series D north of $1 billion at a $26 billion valuation, led by Lux Capital, General Catalyst and 8VC. The headline number: 89% of the code that gets shipped inside Cognition itself is now written by Devin. The remaining 11% comes from sibling agents inside the Windsurf editor. Annualized revenue is at $492M, enterprise usage has multiplied by 10 since January, and big names like Mercedes-Benz, Goldman Sachs, Citi, Dell and the U.S. Army are paying customers. One case study: Mercedes-Benz turned an 8-month legacy migration into an 8-day project. The agent lab is dogfooding its own agent to ship software — and the market is rewarding it like a top-tier AI company.
02
OpenAI buys Ona: Codex now keeps working after you close your laptop
OpenAI announced the acquisition of Ona, a startup that builds secure, reproducible cloud environments for developers — used today by 2 million people. The bet: Codex should keep doing useful work for hours or days inside the customer's cloud, even when your laptop is shut. The deal answers the most common pushback against agentic coding — "what happens when I close the tab?" — by giving the agent a persistent home, with the right credentials, the right logs, the right isolation. Codex already has 5 million weekly users, up 4× since January. With Ona, it stops being a smart autocomplete that forgets everything when you log off, and starts being a colleague that finishes the job while you sleep. For enterprises, this is the missing piece that turns an interesting demo into a production system.
03
Cursor Bugbot 3× faster and 22% cheaper — the AI code reviewer now runs before you push
The team behind the Cursor editor upgraded Bugbot, the AI agent that reviews your code. Backed by the new Composer 2.5 model, Bugbot now finishes a review in about 90 seconds (down from roughly 5 minutes), catches 10% more bugs per run, and costs 22% less to operate. The new /review command folds Bugbot and the Security Review agent into a single click that runs before you open a pull request — and it remembers what was already checked, so GitHub and GitLab don't get the same review twice. In plain English: cheaper, faster, earlier, fewer duplicate alerts — the boring optimizations that decide whether an agent gets used in real teams.
04
OpenAI launches Daybreak: GPT-5.5-Cyber, Codex Security, and Patch the Planet for open-source
OpenAI pushed three security products at once under the Daybreak umbrella. GPT-5.5-Cyber scores 85.6% on the CyberGym benchmark (vs. 81.8% for the regular GPT-5.5) and is being handed to vetted defenders only. Codex Security is a new Codex plugin that builds a threat model of your repo, finds the vulnerabilities, checks which ones can actually be exploited, drafts a fix, then verifies the fix. And Patch the Planet is a program with Trail of Bits, HackerOne and 28 launch partners to ship automated patches to famous open-source projects (cURL, Go, Python, Sigstore, pyca/cryptography) — 30,000 repos and 30 million commits already scanned. OpenAI is betting the real cyber bottleneck isn't finding bugs, it's shipping the fix.
05
OpenAI unveils Jalapeño, its first inference chip, built with Broadcom in 9 months
OpenAI pulled back the curtain on Jalapeño, its first chip built specifically to run large language models at scale. Co-developed with Broadcom (silicon) and Celestica (boards and racks) in just 9 months, the chip is already powering GPT-5.3-Codex-Spark in OpenAI's labs. Mass deployment at gigawatt scale is planned for late 2026 across Microsoft and other partners. Unlike a general-purpose GPU, Jalapeño is tuned for the exact memory and traffic patterns of ChatGPT, Codex and the API. For a normal user, it means faster, cheaper AI responses; for the industry, it means OpenAI is no longer renting all of its horsepower from Nvidia.
📡 To Watch
DeepSeek V4.1 Flash and Pro: 15% cheaper, multimodal, and built for agentic coding
DeepSeek shipped a refresh of its V4 family: same 1M-token window, V4.1 Flash is about 15% cheaper per token, and both versions are tuned for longer multi-step agent tasks and tool use. V4.1 also adds image and audio inputs (which V4 did not) and the MCP standard. V4.1 Flash stays open-weight (MIT license) and is already integrated into Cursor, Continue, Aider and Cline. The funding round last week means they can keep undercutting US pricing — keep an eye on V5, likely Q3.
Cerebras vs OpenAI Jalapeño: the battle of purpose-built inference chips begins
OpenAI's Jalapeño is the second purpose-built inference chip to surface in a month, after Cerebras' announcements. Broadcom is said to be pushing Microsoft to commit to 40% of Jalapeño's first production — but Microsoft is also building its own Maia 2 and Trainium chips. Expect a three-way fight over the next 18 months that will shape how much margin sits below every AI agent call.
Agent infrastructure is being absorbed: Ona (OpenAI), Windsurf (Cognition), Tavus rumored
Within a few weeks, three agent-infrastructure plays have been acquired or folded: Ona into OpenAI, Windsurf into Cognition, Tavus rumored to be in talks. The labs are buying the layers that let agents actually run — the cloud workspaces, the editor surfaces, the human-realistic video. Expect more consolidation in agent "plumbing" before the model layer gets consolidated too.
Patch the Planet vs Anthropic Glasswing: two labs, one open-source cyber war
OpenAI's Patch the Planet and Anthropic's Glasswing both want to fund automated patching of critical open-source projects, with the maintainers as the key constituency. If both labs commit seriously, the result is a meaningful acceleration of fixes for the software the entire internet depends on. If only one follows through, the other gets the open-source goodwill monopoly.
📊 Trend
June 25, 2026 looks like the day the agent economy crossed a threshold. Four things converged at once. (1) Agents run their own company: Cognition's Devin writes 89% of Cognition's own code, and the market valued the lab at $26B on the strength of that. (2) Agents stop being session-bound: OpenAI buying Ona turns Codex into a coworker that keeps working while you sleep, with proper credentials, logs, and isolation. (3) Agents get cheaper and earlier in the loop: Cursor's Bugbot at 90 seconds, 22% cheaper, and now running before code leaves your machine. (4) Agents get the missing chip: OpenAI's Jalapeño gives the model lab its own inference engine at gigawatt scale. The story isn't a single breakthrough — it's the simultaneous closing of the four biggest gaps that kept agents out of production: trust (89% dogfooded), persistence (Ona), cost (Bugbot, Engram earlier this week), and infrastructure (Jalapeño). Once those four are closed, the only remaining question is which labs get to the next billion users first.