# OpenAI Expands Into Deployment and Cyber Defense as Real-Time Models Debut

*By AI News Digest • May 12, 2026*

OpenAI made the day’s clearest enterprise moves with a new Deployment Company and the Daybreak cyber-defense initiative. Thinking Machines pushed native real-time interaction, Cognition posted a notable business signal for coding agents, and research and policy updates pointed to training and governance as the next bottlenecks.

## AI moves closer to live workflows

### OpenAI creates a deployment arm with capital, partners, and field engineers

OpenAI launched the OpenAI Deployment Company to help businesses build and deploy AI to production. The company says it is majority-owned and controlled by OpenAI, brings together 19 investment firms, consultancies, and system integrators, starts with $4 billion of initial investment, and will add 150 Forward Deployed Engineers and Deployment Specialists through its agreed acquisition of Tomoro [^1][^2][^3].

*Why it matters:* OpenAI is building a formal implementation layer around its models, not just selling access to them [^1][^2].

Details: [Deployment Company](https://openai.com/index/openai-launches-the-deployment-company/) [^1]

### Daybreak turns OpenAI’s latest models toward cyber defense

OpenAI also launched Daybreak, described as frontier AI for cyber defenders. It combines OpenAI’s models, Codex, and security partners to help teams find and fix vulnerabilities earlier, cut through security backlogs, and automate detection, validation, and response; Sam Altman said OpenAI wants to work with as many companies as possible on continuous security now [^4][^5][^6][^7][^8].

*Why it matters:* This is one of OpenAI’s clearest attempts to package frontier models into a specific, high-stakes enterprise workflow [^4].

Details: [OpenAI Daybreak](https://openai.com/daybreak/) [^9]

### Thinking Machines makes the case for native real-time interaction

Thinking Machines introduced interaction models, a new class trained from scratch for real-time interaction rather than adapted from turn-based systems [^10]. The company said the models are built to talk, listen, watch, think, and collaborate simultaneously, and Soumith Chintala framed this as step one in increasing human-AI bandwidth; a Horace demo showed model and user speaking at once, which Nathan Lambert called “genuinely different” [^11][^12][^13][^14].

> “People talk, listen, watch, think, and collaborate at the same time, in real time. We’ve designed an AI that works with people the same way.” [^11]

*Why it matters:* The interface race is moving beyond turn-taking chat toward systems built for live collaboration [^10][^12].

Details: [Thinking Machines blog](https://thinkingmachines.ai/blog/interaction-models) [^11]

### A notable commercial signal arrives for AI coding agents

A Colossus profile reported that Cognition’s Devin reached a $445 million revenue run rate in its first 18 months, with usage doubling every eight weeks; customers cited include the U.S. Army, Goldman Sachs, and Mercedes-Benz, and the company is reportedly raising at around a $25 billion valuation [^15]. The same profile says Scott Wu founded Cognition in November 2023 and shipped Devin in March 2024 after an initially rough reception [^15].

*Why it matters:* Reported revenue at this scale suggests AI software agents are moving from demo category to material enterprise spend [^15].

## The next bottlenecks: training recipes and state capacity

### Better pre-training recipes are still producing large gains

A Stanford CS25 lecture described three levers: a two-phase curriculum that improved results 17% over random ordering and 3.4% over an optimal blend without curriculum, front-loading reasoning data so gains persist through SFT and RL, and “reinforcement as pre-training” (RLP), where models generate reasoning traces before predicting the next token [^16]. Combined, the strategies yielded up to 60% relative improvement over baselines using the same data, and related datasets were open-sourced on Hugging Face [^16].

*Why it matters:* The frontier is still moving through training method, not only more data and more compute [^16].

### AI policy discussion gets more operational

Big Technology noted that the U.S. government gained early access to models from Microsoft, Google, and xAI for national security testing [^17]. Separately, Import AI highlighted the Institute for Law & AI’s “radical optionality” approach: avoid overregulation in the short term while building institutions, information channels, legal authorities, model-security measures, assessments, and technical talent for a range of future scenarios [^18]. Jack Clark argued that ideas like these can start paying off quickly by generating information and building state capacity around advanced technology [^19][^20].

*Why it matters:* The policy conversation is inching away from abstract principles and toward concrete testing, staffing, and oversight mechanisms [^17][^20].

---

### Sources

[^1]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2053824997777457651)
[^2]: [𝕏 post by @gdb](https://x.com/gdb/status/2053884619695730745)
[^3]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2053824999736410415)
[^4]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2053939702110269822)
[^5]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2053939703473430658)
[^6]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2053939705146945887)
[^7]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2053939706468139481)
[^8]: [𝕏 post by @sama](https://x.com/sama/status/2053951874408276193)
[^9]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2053939707818692907)
[^10]: [𝕏 post by @miramurati](https://x.com/miramurati/status/2053939069890298321)
[^11]: [𝕏 post by @thinkymachines](https://x.com/thinkymachines/status/2053938892152435174)
[^12]: [𝕏 post by @soumithchintala](https://x.com/soumithchintala/status/2053940215505645938)
[^13]: [𝕏 post by @thinkymachines](https://x.com/thinkymachines/status/2053938899735724254)
[^14]: [𝕏 post by @natolambert](https://x.com/natolambert/status/2053972090110562548)
[^15]: [𝕏 post by @colossusmag](https://x.com/colossusmag/status/2053801052571312414)
[^16]: [Stanford CS25: Transformers United V6 I From Next-Token Prediction to Next-Generation Intelligence](https://www.youtube.com/watch?v=e_H_tkpCAK4)
[^17]: [Satya, Sam To Take The Stand This Week + Highlights From Musk v. Altman](https://www.bigtechnology.com/p/satya-sam-to-take-the-stand-this)
[^18]: [Import AI 456: RSI and economic growth; radical optionality for AI regulation; and a neural computer](https://importai.substack.com/p/import-ai-456-rsi-and-economic-growth)
[^19]: [𝕏 post by @jackclarkSF](https://x.com/jackclarkSF/status/2053846259077877814)
[^20]: [𝕏 post by @jackclarkSF](https://x.com/jackclarkSF/status/2053846447989416289)