# OpenAI Moves on Ona, Gemini Tops Video, and the AI Price War Intensifies

*By AI High Signal Digest • June 12, 2026*

OpenAI moved to strengthen Codex with an acquisition, Gemini took a clear lead in video generation benchmarks, and the business fight between Anthropic and OpenAI is increasingly about pricing and margins. The brief also covers a sobering new agent benchmark, automated AI research, and key product and infrastructure launches.

## Top Stories

*Why it matters: today’s clearest signals were about agent infrastructure, multimodal model performance, and the economics of the frontier-model race.*

- **OpenAI moved to strengthen Codex’s agent stack** by reaching an agreement to acquire Ona. OpenAI said Ona’s secure cloud execution technology will help Codex handle longer-running work even when laptops are closed and help organizations deploy agents securely in production; Ona will join the Codex team after closing. [^1]
- **Gemini Omni Flash jumped to the top of video generation benchmarks.** It now ranks #1 in Video Arena for both text-to-video and image-to-video, with Arena reporting a +158 point gain over Veo 3.1 (1080p) in text-to-video, a +61 point lead over Seedance 2.0, a +77 point gain in image-to-video, and an 82% head-to-head win rate. DeepMind has framed Omni as its first step toward a model that can “create anything from anything,” starting with video. [^2][^3][^4][^5]
- **Competition is shifting from benchmarks to margins.** Anthropic is reportedly on track for its first profitable quarter, with revenue more than doubling to about $10.9B, while OpenAI is reportedly weighing further token price cuts to keep enterprises from moving to Claude. Separate estimates put $200 plans at roughly $8,000 of Claude Max 20x usage and $14,000 of ChatGPT Pro 20x usage, highlighting how aggressive current pricing already is. [^6][^7]

## Research & Innovation

*Why it matters: the strongest research updates sharpened the picture on agent limits, automated research, and training-data control.*

- **Agents’ Last Exam (ALE)** introduced a rolling benchmark built from more than 1,500 expert-sourced tasks across 55 occupations. The results were mixed: today’s frontier agents can solve a meaningful fraction of professional tasks, but every frontier system tested scored **0%** on the hardest tier requiring sustained reasoning, deep expertise, and long-horizon execution. Separate commentary said GPT-5.5 led the eval even when measured by tokens, cost, or wall-clock time. [^8][^9]

> "The age of useful agents is here. The age of truly job-ready agents is not." [^8]

- **Recursive** unveiled an early automated AI research system, “Eureka Machine” v0.1, that it says reached state-of-the-art results on NanoGPT speedrun, NanoChat, and NVIDIA’s Sol-ExecBench. The company says the code and ideas behind those results were generated by the AI system itself and is open-sourcing the discoveries for scrutiny. [^10]
- **Goodfire introduced “predictive data debugging,”** a method for estimating what DPO training data will amplify or suppress before training, reporting R²=0.9 against what models later learn. In examined datasets, it surfaced weaker safety guardrails, hallucinated links on sensitive topics, and localized sycophancy. [^11][^12][^13][^14]

## Products & Launches

*Why it matters: new releases kept pushing agents, open coding models, and media systems closer to practical deployment.*

- **Perplexity integrated Deep Research directly into Computer.** The feature is built on a “Search as Code” architecture in which the model writes code to assemble search, runs thousands of retrieval steps in parallel, connects to long-running sandboxes and tools, and is now available to Pro and Max subscribers. Perplexity says it outperforms its legacy Deep Research on every benchmark. [^15][^16]
- **Cohere released North Mini Code 1.0,** its first open-source coding model. It is a 30B-parameter MoE with 3B active parameters running at about 66 tok/s in BF16, with day-zero MLX support and local deployment paths through GGUF quants, llama.cpp, Ollama, and vLLM. [^17][^18][^19]
- **Ideogram 4.0 became Ideogram’s first open-weights release.** It debuted at #8 on the Open Weights Text-to-Image leaderboard and #31 overall, with 2K x 2K outputs, multilingual text rendering, bounding-box layout control, transparent backgrounds, and structured JSON prompts. [^20]

## Industry Moves

*Why it matters: capital and infrastructure decisions are increasingly defining who can build, train, and deploy the next generation of systems.*

- **Jeff Bezos raised $12B for Prometheus** at a $41B valuation. The company’s pitch is an “artificial general engineer” that compresses the design-to-build loop by 10x or more, alongside a reported $100B vehicle to acquire industrial companies and the manufacturing data they generate. [^21]
- **xAI is building a 500 MW data center in Saudi Arabia** with HUMAIN and NVIDIA, which would make it xAI’s largest facility outside the U.S.; for comparison, Colossus-1 in Memphis is around 300 MW. [^22]
- **Google DeepMind launched a $10M research fund** with partners including Schmidt Sciences, Cooperative AI, and ARIA to study how AI systems behave as a group. [^23]

## Policy & Regulation

*Why it matters: governments are starting to define how AI connects to national infrastructure, not just consumer software.*

- **China’s MIIT issued an AI+ICT implementation plan for 2026-2028** that ties together the “East Data West Compute” project, autonomous networks, agentic AI, embodied intelligence, and the national compute-network strategy. [^24]

## Quick Takes

*Why it matters: these smaller updates still add signal on deployment speed, tool quality, and inference economics.*

- Trajectory Labs says it post-trained NVIDIA Nemotron 3 Ultra on Harvey Legal Agent Bench in under 24 hours, putting an open model into the same performance band as leading closed legal models at lower cost. [^25]
- Cursor made Auto-review the default for new users; its classifier subagent reviews actions in context and reportedly hits 97% accuracy. [^26]
- Baseten and Inception launched Mercury 2 in production, citing 1,000+ tok/s on standard NVIDIA GPUs and early reports of 82% lower latency and 90% cost savings. [^27]
- OpenAI added developer mode for Codex browser use in Chrome and the in-app browser, with Chrome DevTools Protocol support for profiling JavaScript, console output, network traffic, and page state. [^28]

---

### Sources

[^1]: [𝕏 post by @OpenAINewsroom](https://x.com/OpenAINewsroom/status/2065088002335158753)
[^2]: [𝕏 post by @arena](https://x.com/arena/status/2065112147093545333)
[^3]: [𝕏 post by @arena](https://x.com/arena/status/2065112152659386646)
[^4]: [𝕏 post by @arena](https://x.com/arena/status/2065112150914527459)
[^5]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2056786446636212467)
[^6]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2065077145664192739)
[^7]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2064987311402537184)
[^8]: [𝕏 post by @dawnsongtweets](https://x.com/dawnsongtweets/status/2065095757988868190)
[^9]: [𝕏 post by @polynoamial](https://x.com/polynoamial/status/2065125807585149136)
[^10]: [𝕏 post by @RichardSocher](https://x.com/RichardSocher/status/2065094362774876232)
[^11]: [𝕏 post by @GoodfireAI](https://x.com/GoodfireAI/status/2065118209209258013)
[^12]: [𝕏 post by @GoodfireAI](https://x.com/GoodfireAI/status/2065118244583956908)
[^13]: [𝕏 post by @GoodfireAI](https://x.com/GoodfireAI/status/2065118262002888886)
[^14]: [𝕏 post by @GoodfireAI](https://x.com/GoodfireAI/status/2065118278025097405)
[^15]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2065124930463916317)
[^16]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2065124948793028691)
[^17]: [𝕏 post by @Prince_Canuma](https://x.com/Prince_Canuma/status/2064437722689962242)
[^18]: [𝕏 post by @cohere](https://x.com/cohere/status/2065184560875176437)
[^19]: [𝕏 post by @cohere](https://x.com/cohere/status/2065184564406854142)
[^20]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2065135515171709056)
[^21]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2065097988859744503)
[^22]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2065092418215543186)
[^23]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2065031279213441309)
[^24]: [𝕏 post by @pstAsiatech](https://x.com/pstAsiatech/status/2065111824802988084)
[^25]: [𝕏 post by @trajectorylabs](https://x.com/trajectorylabs/status/2064756005507072339)
[^26]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2065137803084857845)
[^27]: [𝕏 post by @baseten](https://x.com/baseten/status/2065100012934095171)
[^28]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2065226355495895521)