Google Finally Ships Gemini 4 — Closing Out a Wild Seven Weeks

Filed October 1, the day these developments broke — posting today.

After months of delay and a release date that had already slipped at least twice this year, Google used the first days of October to finally answer the question that's hung over the industry since mid-August: where is Gemini 4?

Gemini 4 arrives

Alphabet announced its new flagship model to anchor the Gemini 4 generation, explicitly positioned as a fresh bid to catch up to Anthropic and OpenAI in frontier capability. It's a release a long time coming — back in mid-August, reporting on Google's internal DeepMind restructuring noted Gemini's next release had already slipped by two months, and as recently as late September, industry trackers were still describing Gemini 4 as being "in post-training."

Whether it actually closes the gap with Claude Opus 5.5 and GPT-6 Astra, both of which have shipped and been iterated on multiple times since, is the question the next few weeks of independent benchmarking will have to answer — but after a year defined by Google's billion-user Gemini milestone in August and a rocky four-month stock slide that only broke in early September, landing Gemini 4 cleanly matters enormously for the company's competitive story heading into the end of the year.

The White House pact gets its first real test

Just a day after AI company leaders signed their voluntary safety pact at the White House, OpenAI quietly tightened its own security posture: starting today, Codex Security Cloud's cyber-forward capabilities require individual users to have an eligible hardware security key, while enterprise access remains unaffected.

It's a small, concrete step, but it's a notably fast turnaround from pledge to implementation — exactly the kind of fast-follow action that's been largely absent from the more symbolic commitments made earlier in this saga.

A stealth model keeps the open community guessing

OpenRouter extended free access to "Space Bunny Alpha," an anonymous multimodal stealth model with a 1-million-token context window and up to 524,288 tokens of output, through October 5, after further speed and reliability work.

Stealth-model releases like this have become a recurring feature of 2026's AI landscape — labs testing frontier capability anonymously before claiming it, letting the community's own benchmarking build hype ahead of an official reveal.

Medical AI gets a tougher, narrower test

Stanford-backed researchers opened the MMBU Challenge today, a competition specifically designed to test whether vision-language models actually recognize medical imaging modality, body part, specimen type, and stain — rather than simply guessing the right final clinical answer through pattern-matching shortcuts.

With more than $100,000 in compute and prizes at stake and the competition running through the end of the year, it's a useful corrective to a year of AI medical headlines that have often been long on promise and short on the kind of rigorous, mechanism-level evaluation this challenge is built around.

Closing out an extraordinary seven weeks

It's a fitting way to close out an extraordinary seven weeks. Since mid-August, the AI world has moved through billion-user milestones, a 744-billion-parameter open-weight release, a cyber-capability threshold crossed for the first time, an extinction-risk warning from inside one of the most safety-focused labs, a UN Security Council briefing, three separate countries reporting AI agents probing government systems, a training halt, a White House safety pact, and now, finally, the release everyone had been waiting on from Google.

If this pace holds, whatever comes next won't stay quiet for long — check back as the story keeps developing.