Codex goes mobile and AI reinvents Chinese drama

Today’s most important AI news

Friday arrived with concrete moves in developer tools: OpenAI expanded Codex to mobile while a top-tier Asian company deploys it in production, signaling that agentic coding is no longer an experiment but everyday operation. In parallel, an MIT Technology Review investigation exposes how China is using generative AI to industrialize the production of audiovisual content at massive scale. And from the labs, a finding about “committees” of weak models challenges the logic that you always need the biggest model to get the best results.

Codex leaves the desktop: OpenAI puts agentic coding in your pocket

OpenAI announced that Codex is now available from the ChatGPT mobile app, letting you monitor, redirect and approve programming tasks in real time from any device. The move looks incremental, but it isn’t: it means agentic coding workflows, where the model works autonomously on long tasks, no longer require the developer to be in front of a computer. You can check what your agent is doing from the airport or approve a pull request while waiting for your coffee.

What gives this launch weight is the backing of a real case. Sea Limited, the Asian tech conglomerate that owns Shopee and Garena, published through its CPO David Chen how it’s deploying Codex across its engineering teams to speed up software development. It’s not a pilot: it’s adoption at scale at a company with tens of millions of active users in Southeast Asia. Sea’s bet is that Codex lets it compete on development speed without scaling its engineering headcount linearly.

Meanwhile, Cat Wu, product lead for Claude Code at Anthropic, gave an interview to Ars Technica in which they acknowledged that the product doesn’t have “a big master plan” and that its intentionally loosely coupled design, what they call the “lean harness”, is precisely what lets them iterate quickly. The transparency about usage limits and the lack of a rigid roadmap contrasts with OpenAI’s polished launch tone, but both companies are aiming at the same market: the developer who wants to delegate real work to the model, not just autocomplete lines.

The practical implication for product and development teams is clear: if you’re still evaluating whether to adopt agentic coding tools, the window of “exploration without pressure” is closing. Companies like Sea are already measuring delivery speed with these tools, which is starting to create a benchmark for comparison across the industry.

China turns AI into a short-drama factory

MIT Technology Review today published a detailed investigation into how China’s short-drama industry — those 1-to-3-minute melodramatic videos designed for smartphone consumption — is being produced end-to-end with generative AI. Complete scenes, dialogue, visual effects and narrative continuity generated by AI at a fraction of the cost of traditional production.

The phenomenon is industrial in scale. Chinese short dramas are already a market worth several billion dollars a year, and the incorporation of AI isn’t an artistic experiment but an economic decision: it cuts production time from weeks to days and lets content be iterated based on engagement signals almost in real time. The result is hyper-personalized content, optimized for the scroll and produced in volumes no human team could match.

For professionals in media, entertainment and product, this is a signal of where the content economy is heading. Generative AI isn’t competing with Hollywood for Oscars, it’s competing for the attention minutes of 500 million mobile users with content that’s good enough, produced fast enough. The copyright, cultural identity and content moderation questions this raises are real, but the market has clearly already made its decision.

Weak models working as a team can beat a strong model alone

A paper published today on arXiv poses a provocative question: can a committee of weak reasoning models match the performance of much more powerful models? The answer, according to the authors, is yes under certain conditions. The mechanism isn’t that “more agents always help”, but that multiple calls to simpler models expose latent correct solutions that a verifier can recover, even without directly accessing the correct answer during the process.

The finding has direct cost implications. If a system of five calls to a small model can match one call to a frontier model, the arithmetic of cost per task changes radically. For companies evaluating which model to use in production, this work suggests the decision isn’t just “which model is better” but “which inference architecture is most efficient for my specific use case”.

The important limitation the paper points out is that the system requires a reliable verifier: something that can identify when the committee arrived at a correct answer. In domains where verification is cheap and clear, like executable code or math, the approach is directly applicable. In ambiguous domains, the advantage shrinks. For product managers designing AI pipelines, that’s the usage guide: committees of weak models where you can verify the output, frontier models where you can’t.

In a nutshell

What ties today’s news together is a quiet maturation: AI is no longer the topic of the strategy meeting but the operating infrastructure for decisions already made. Sea isn’t evaluating Codex, it’s using it. China isn’t debating whether to produce content with AI, it’s industrializing it. Researchers aren’t asking whether agents are useful, but how to optimize their cost. The debate has shifted from “whether” to “how” and “how much it costs to do it well”, and that completely changes the conversations worth having within teams.

---

Sources used: https://openai.com/index/work-with-codex-from-anywhere, https://openai.com/index/sea-david-chen, https://arstechnica.com/ai/2026/05/claude-codes-product-lead-talks-usage-limits-transparency-and-the-lean-harness/, https://www.technologyreview.com/2026/05/15/1137326/chinese-short-dramas-ai/, https://arxiv.org/abs/2605.14163