# ChatGPT Images 2.0 Leads a Day of Coding-AI and Research-Agent Moves

*By AI News Digest • April 22, 2026*

OpenAI's new image model was the clearest launch of the day. Around it, a large Cursor-SpaceX deal, deeper research agents from Google and Hugging Face, and China's AI-in-education plan pointed to where the market is heading.

## OpenAI pushes image generation further into real work

### ChatGPT Images 2.0 arrives with stronger text, layout, and reasoning
OpenAI launched ChatGPT Images 2.0 as a state-of-the-art image model for complex visual tasks, emphasizing sharper editing, richer layouts, and thinking-level intelligence [^1]. The company says it can handle small text, UI elements, dense compositions, non-English text, flexible aspect ratios, and outputs up to 2K resolution; when used with a thinking model, it can also search the web, generate multiple distinct images, double-check outputs, and create functional QR codes [^2][^3][^4][^5].

It is available starting today to all ChatGPT and Codex users, with thinking features for Plus, Pro, and Business users, and the underlying `gpt-image-2` model is available in the API [^6]. Greg Brockman highlighted applications in education, slides, marketing materials, and diagrams for code documentation [^7].


[![Introducing ChatGPT Images 2.0](https://img.youtube.com/vi/sWkGomJ3TLI/hqdefault.jpg)](https://youtube.com/watch?v=sWkGomJ3TLI&t=577)
*Introducing ChatGPT Images 2.0 (9:37)*


**Why it matters:** OpenAI is framing image generation less as a novelty and more as a tool for design, documentation, and communication work [^8][^9][^7].

## Coding AI is starting to look like an infrastructure business

### SpaceX and Cursor sign an unusually large option-style deal
SpaceX said it is working closely with Cursor to build a coding and knowledge-work AI system, combining Cursor's distribution with its million H100-equivalent Colossus training supercomputer [^10]. The agreement gives SpaceX the right to acquire Cursor later this year for $60 billion, or pay $10 billion for the collaboration instead [^10].

Separately, Sam Altman said Codex reached 4 million active users less than two weeks after hitting 3 million, and OpenAI would reset rate limits amid demand [^11].

**Why it matters:** The coding-assistant market is showing both strategic value and real usage pressure, with large compute partnerships on one side and fast-rising user demand on the other [^10][^11].

## Research agents keep getting more production-oriented

### Google expands Deep Research in the Gemini API
Google launched Deep Research and Deep Research Max, autonomous research agents powered by Gemini 3.1 Pro that can navigate the web and custom data to create fully cited reports [^12]. The update adds better quality, MCP support, and native chart and infographic generation; Google says standard Deep Research is tuned for speed and efficiency, while Max uses more test-time compute for deeper context gathering and synthesis [^13][^14][^15][^13].

Google also highlighted planning mode, full tool support, multimodal inputs, real-time progress streaming, and availability through the Gemini API [^16][^14].

**Why it matters:** Research agents are being packaged more like configurable products for enterprise workflows, not just web-summarization demos [^12][^14].

### Hugging Face open-sources an agent for post-training research
Hugging Face released `ml-intern`, an open-source agent that researches papers and citation graphs, pulls datasets, runs training in GPU sandboxes, and iterates on model improvement across the HF ecosystem [^17]. In examples shared by the team, it raised a Qwen3-1.7B model from 10% to 32% on GPQA in under 10 hours, beat Claude Code's cited 22.99% result on the same prompt, generated 1,100 synthetic healthcare examples and beat Codex by 60% on HealthBench, and wrote plus iterated on a GRPO training script for competitive math [^17].

The tool is available as both a CLI and web/mobile app, and Hugging Face leaders framed the broader platform as a place for agents to use and build AI rather than just call APIs [^17][^18][^19].

**Why it matters:** This is an open-source attempt to automate part of the model-improvement loop itself, not only downstream agent tasks [^17].

## A notable policy signal came from education

### China rolls AI into teacher training and school curricula
A recently rolled-out national AI+Education action plan from China's Ministry of Education aims to weave AI into every stage of teaching and learning, before, during, and after class [^20]. The plan's stated goal is to establish by 2030 a comprehensive AI education system spanning all levels of schooling and the broader public [^20].

Reported implementation details include AI literacy requirements for teachers, AI content added to national teacher qualification exams, tiered training programs by subject, at least eight study hours per year for primary and secondary students in some regions, and a national smart education platform with more than 1,000 AI courses [^20]. The same reporting says teachers are increasingly being recast as learning designers and ethical guides, with AI expected to reduce time spent on standardized tasks such as question generation, grading, and student analysis [^20].

**Why it matters:** This is a national-scale effort to normalize AI literacy and AI-assisted teaching, not a small pilot or a single curriculum update [^20].

---

### Sources

[^1]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2046670977145372771)
[^2]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2046670980358266923)
[^3]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2046670983092990411)
[^4]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2046670987425669310)
[^5]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2046670989719924768)
[^6]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2046670994413322435)
[^7]: [𝕏 post by @gdb](https://x.com/gdb/status/2046694017287811390)
[^8]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2046670978890276918)
[^9]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2046670992123248802)
[^10]: [𝕏 post by @SpaceX](https://x.com/SpaceX/status/2046713419978453374)
[^11]: [𝕏 post by @sama](https://x.com/sama/status/2046604989527912590)
[^12]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2046627042335060342)
[^13]: [𝕏 post by @sundarpichai](https://x.com/sundarpichai/status/2046627545333080316)
[^14]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2046627049045967175)
[^15]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2046627044453134389)
[^16]: [𝕏 post by @OfficialLoganK](https://x.com/OfficialLoganK/status/2046628030777631000)
[^17]: [𝕏 post by @akseljoonas](https://x.com/akseljoonas/status/2046543093856412100)
[^18]: [𝕏 post by @ClementDelangue](https://x.com/ClementDelangue/status/2046598219853951346)
[^19]: [𝕏 post by @Thom_Wolf](https://x.com/Thom_Wolf/status/2046560319196070016)
[^20]: [How AI is redefining the teacher](https://www.youtube.com/watch?v=V3kp4wJxSqM)