Google finally shipped, just not the model everyone was waiting for. On July 21 the company released three new Gemini models (3.6 Flash, 3.5 Flash-Lite, and a security-tuned 3.5 Flash Cyber) while confirming that Gemini 3.5 Pro is still not ready. In the same breath it announced it has begun pretraining Gemini 4, effectively asking the industry to look past the flagship it could not deliver. Meanwhile OpenAI added two heavyweight board members ahead of its IPO, Block launched a Slack competitor built for AI agents, and Claude Fable 5 disproved an 87-year-old math problem. Both things are true: the frontier is advancing, and the frontier is stalling. Progress is not linear.
Google Ships Three Flash Models While the Flagship Slips
Google released three new Gemini models on July 21: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber, a security-tuned variant restricted to governments and trusted partners. Conspicuously absent was Gemini 3.5 Pro, the flagship that has now missed its target multiple times. Google used the same announcement to confirm Pro is still not shipping. The stopgap reported last week turned out to be real.
Gemini 3.6 Flash cuts costs and tokens
Gemini 3.6 Flash uses roughly 17 percent fewer output tokens on the Artificial Analysis Index and takes fewer reasoning steps and tool calls to complete multi-step jobs. It is priced at $1.50 per million input tokens and $7.50 per million output, down from the $9 output price of the previous Flash generation. Its knowledge cutoff jumps from January 2025 to March 2026. For agentic workloads where a single task can involve dozens of tool calls, taking fewer steps to finish translates into real latency and cost improvements that benchmark scores do not capture. This is the most practically useful thing Google shipped this month.
Gemini 3.5 Flash Cyber takes on Anthropic’s security lead
Gemini 3.5 Flash Cyber is a security-tuned model restricted to governments and trusted partners, positioned directly against Anthropic’s Mythos-class security models. It joins Microsoft’s Project Perception in turning AI security into a three-way contest. The restricted access is the notable design choice: a model tuned to find software vulnerabilities is inherently dual-use, and this is becoming the established norm for security-capable models. Build them, but do not ship them broadly. It is one of the few places where the industry has converged on genuine restraint without being forced into it by regulation.
Google confirms Gemini 4 pretraining has begun
In the same update where it acknowledged Gemini 3.5 Pro is still not ready, Google announced it has begun what it calls its most ambitious pretraining run yet for Gemini 4. The move is understandable and risky in equal measure. Understandable because Google scrapped the original 3.5 Pro base model and restarted pretraining once already, and if that second attempt is still falling short, pouring more effort into fixing it may be worse than moving to a fundamentally better architecture. Risky because announcing the next generation while the current one is unshipped invites the obvious question of whether Gemini 3.5 Pro will ever arrive at all. Enterprises making platform decisions this quarter cannot buy a pretraining run.
OpenAI Adds Board Members Ahead of IPO
OpenAI appointed David Velez, founder and CEO of Nubank (Latin America’s largest digital bank), and Robin Vince, CEO of BNY, to both its nonprofit and for-profit boards on July 21. The appointments come as OpenAI marches closer to a prospective IPO, bringing external financial governance expertise into a structure that has been criticized for its unusual capped-profit setup. The WSJ reported the additions as independent board members, a signal that OpenAI is preparing for the scrutiny that comes with public markets. Chairman Bret Taylor separately told CNBC that companies will stop worrying about AI token costs within 12 months, as vendors and sector-specific AI companies take over the job of managing them. “I think if you fast-forward 12 months from now, IT departments will be really sophisticated about the industrial applications of AI,” Taylor said.
Block Launches Buzz, an Open Workspace for Humans and Agents
Block released Buzz on July 21, a free open-source collaboration platform built on the Nostr protocol where humans and AI agents work together in a shared workspace. Agents are full members with their own accounts, not chatbots bolted onto a chat app. The platform has channels, threads, direct messages, voice, media sharing, code repositories, and automated workflows. Jack Dorsey positioned it as a challenger to Slack and GitHub, built to reduce dependency on centralized platforms. Building on Nostr means the workspace is not owned or controlled by any single company. This is infrastructure play, not feature play: Block is betting that the next generation of collaboration tools will treat AI agents as first-class participants, and that decentralization matters enough to trade the polish of Slack for the sovereignty of an open protocol.
Claude Fable 5 Disproves an 87-Year-Old Math Conjecture
Anthropic researcher Levent Alpoge announced on July 20 that he used Claude Fable 5 to find a counterexample to the Jacobian conjecture, an open problem in mathematics dating back to 1939. The model produced a 216-character counterexample that was independently verified. The Jacobian conjecture has a reputation among mathematicians for looking deceptively approachable and then eating careers. It sits on Steve Smale’s famous list of the most important math problems for the 21st century, alongside the Riemann Hypothesis. This is not a benchmark score. This is original mathematical work, and it suggests frontier models are crossing from pattern reproduction into genuine research contribution.
Quick Hits
Google’s Frozen v2 chip (July 20-21) — Google is developing a server chip code-named Frozen v2, built around the Gemini architecture, that internal sources claim is 6 to 10 times more efficient than its current TPUs. If those numbers survive contact with production, it would be the largest single-generation efficiency jump in Google’s custom silicon program. Custom silicon is the one area where Google’s decade-long head start is undisputed. TechCrunch
Kimi K3 suspends new subscriptions (July 21) — Moonshot AI suspended new subscriptions for Kimi K3 after demand exceeded its serving capacity, days after the model launched and took the top spot on a major coding leaderboard. Running out of capacity is the good kind of problem. The July 27 weight release changes this dynamic entirely: once weights are public, capacity stops being Moonshot’s problem. TechTimes
Warren Buffett reveals $31B Alphabet stake as AI bet (July 19-20) — Warren Buffett disclosed he personally directed Berkshire Hathaway’s multibillion-dollar stake in Alphabet, calling it a direct bet on AI’s next phase. “I initiated it,” he told CNBC. The stake makes Alphabet Berkshire’s third-largest holding. Motley Fool
Meta says AI moderation beats humans, users disagree (July 21-22) — Meta reported its AI moderation system produces 13 percent fewer errors and finds 10 percent more policy violations than human moderators. At the same time, some Instagram and Facebook users say the system has incorrectly deleted their accounts. Both things can be true: aggregate accuracy is the metric Meta optimizes; individual catastrophic errors are what users experience. BuildFastWithAI
Substack adds AI detection through Pangram (July 21-22) — Substack partnered with AI-detection tool Pangram to let users scan text longer than 100 words for an estimate of how much appears AI-generated. The framing Substack chose (an estimate rather than a label) is the responsible approach. Detection is losing an asymmetric race: making a model imitate a writing style takes one line of prompting while detecting it is a hard statistical problem. The Verge
NVIDIA Cosmos 3 Edge launches (July 21) — NVIDIA released Cosmos 3 Edge, a 4-billion-parameter open world model for physical AI that reasons and generates robot actions on-device. It is the third and smallest member of the Cosmos 3 family, joining Cosmos 3 Nano (16B) and Cosmos 3 Super (64B). MarkTechPost
Anthropic opens rare disease research grants (July 20) — Anthropic opened a focused call for AI for Science rare disease research grants, offering up to $50,000 per project for researchers using Claude to study rare genetic disorders. Anthropic
Compiled July 22, 2026. Sources: Yellow, OpenAI, Google DeepMind, Google AI Blog, Anthropic, NVIDIA, Hugging Face, BuildFastWithAI, TechCrunch, CNBC, The Verge, SiliconAngle, MarkTechPost.