AI Roundup May 2025: Claude 4 Lands, Google I/O Goes Full Video, OpenAI Buys Jony Ive

TL;DR

May was a heavyweight month. Anthropic shipped Claude Opus 4 and Sonnet 4, instantly reclaiming the coding crown. Google I/O turned into a video-generation showcase with Veo 3 and its native audio. OpenAI dropped $6.5 billion to absorb Jony Ive's hardware startup, betting the next interface is a device, not a chat box. And Mistral quietly handed builders an open-weight coding agent that fits on a single GPU.

Claude Opus 4 and Sonnet 4 Retake the Coding Crown

On May 22, Anthropic launched Claude Opus 4 and Claude Sonnet 4, and the headline number was hard to argue with: Opus 4 posted 72.5% on SWE-bench Verified, the best of any frontier model at the time. Both are hybrid models, meaning you get near-instant answers or an extended-thinking mode for harder problems, toggled per request. Pricing held steady at $15/$75 per million tokens for Opus and $3/$15 for Sonnet, so the gains came without a tax.

The pitch this time was endurance. Anthropic emphasized sustained performance on long, multi-step agent workflows, the kind of thing that falls apart after a few dozen tool calls in lesser models. For anyone building coding agents, this was the month the default backend changed. Sonnet 4 landed in GitHub Copilot's public preview within days, which tells you how fast the ecosystem moved.

Opus 4 is the model you reach for when the task is a real codebase, not a leetcode snippet. That distinction is the whole game now.

Google I/O Turns Into a Video-Generation Demo

Google I/O on May 20 was less a search keynote and more a generative-media flex. The star was Veo 3, the first major text-to-video model with native audio: dialogue, sound effects, and ambient noise generated in sync with the footage. That sounds incremental until you watch a clip and realize the lip-sync and the soundscape came out of one prompt, not a six-tool pipeline.

Around it, Google announced Imagen 4 for stills, Flow as a filmmaking front-end stitching Veo and Imagen together, and AI Mode rolling into Search. The catch for builders: Veo 3 launched gated behind the new Google AI Ultra subscription in the US, so the best toy was also the most expensive one. Still, the direction is clear. Prompt-to-finished-clip with audio is now a product, not a research demo.

OpenAI Pays $6.5 Billion for Jony Ive's io

On May 21, OpenAI announced it was acquiring io, the hardware startup co-founded by former Apple design chief Jony Ive, in an all-stock deal valued at roughly $6.5 billion. It is OpenAI's largest acquisition to date. Ive and his LoveFrom studio stay independent but take on design and creative leadership across OpenAI's products.

The subtext is a bet that the long-term home for AI is a dedicated device, not someone else's phone OS. OpenAI has the model and the user base but no hardware surface it controls. Buying the person who shaped the iPhone is an expensive way to fix that. Whether the eventual gadget justifies the price is a 2026 question, but the strategic message landed immediately: the interface war is heating up.

Mistral Ships Devstral, a Coding Agent That Fits on One GPU

For the local-AI crowd, the most useful drop of the month came on May 21. Mistral, working with All Hands AI, released Devstral, an open-weight agentic model built specifically for software engineering. It scored 46.8% on SWE-bench Verified, beating the prior open-source state of the art by more than six points, while being small enough to run on a single RTX 4090 or a 32GB Mac.

That combination is the point. Devstral is trained to operate inside agent scaffolds like OpenHands and SWE-Agent, resolving real GitHub issues rather than writing isolated functions. An open-weight coding agent you can self-host, with frontier-ish agent behavior, is exactly the kind of release that quietly changes what indie builders ship next quarter.

Key Takeaways

• Coding is the frontier battleground. Claude Opus 4 and Devstral both led with SWE-bench numbers, because agentic coding is now the benchmark that sells models.
• Native-audio video crossed a line. Veo 3 collapsed a multi-tool video pipeline into one prompt, which resets expectations for what "generate a clip" means.
• The interface is up for grabs. OpenAI's $6.5B io deal is a wager that AI's endgame is dedicated hardware, not an app on a phone you do not control.
• Open weights keep closing the gap. Devstral running on a single 4090 means serious coding agents no longer require a hyperscaler bill.
• Endurance beats peak IQ for agents. The month's model story was sustained multi-step performance, the trait that actually determines whether your agent finishes the job.