Need to know
- Google shipped three Gemini models simultaneously while confirming Gemini 4 pretraining has begun.
- Samsung is in talks to invest ~€1B in Mistral, weeks after Microsoft's infrastructure commitment to the French lab.
- AMD signed Anthropic to a multi-year, tens-of-billions supply-and-investment pact, its third marquee AI customer in a year.
- GitHub Copilot added Gemini 3.6 Flash as a selectable model for enterprise developers today.
- Okta data shows enterprises now routinely run multiple AI platforms simultaneously, reshaping identity governance scope.
New Releases
Google DeepMind released Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber in a single overnight drop targeting agent workloads.
- Gemini 3.6 Flash consumes 17% fewer tokens per task than 3.5 Flash on the Artificial Analysis Intelligence Index and up to 65% fewer output tokens on DeepSWE coding benchmarks.
- Gemini 3.5 Flash Cyber is a security-specialized variant; Flash-Lite is a high-throughput, lower-cost option — the three models together signal Google is competing on efficiency economics, not just capability.
GitHub Copilot made Gemini 3.6 Flash available as a selectable model for enterprise developers on July 23.
- Enterprise dev teams can now route Copilot requests to Google's newest Flash model, adding a third major provider option alongside OpenAI and Anthropic models already in Copilot.
- Timing matters: Copilot's multi-model selection is becoming a de facto model router for enterprise coding, with each new addition shifting negotiating leverage away from any single AI lab.
OpenAI launched Presence, an enterprise product that scopes AI agents to specific jobs — billing disputes, insurance claims, IT requests — with task-bounded data and system access.
- Each deployment starts with a defined role and receives only the information and access needed for that task, a direct response to enterprise concerns about over-privileged AI agents.
- Presence combines voice and chat interfaces, targeting customer service and internal operations use cases where containment and auditability matter most.
South Korean AI lab Upstage released Solar Open 2, a 250B-parameter MoE open-weight model optimized for multi-step agentic inference, built under South Korea's national AI foundation model program.
- MoE architecture activates only 15B parameters per inference step, making it practical to run long agentic loops — dozens of tool calls per task — at reasonable cost.
- Government-backed provenance gives regulated Korean enterprises a sovereign alternative to US-hosted frontier models, with the model embedded in the Daum portal as a live deployment.
Funding
Anthropic signed a supply-and-investment agreement with AMD worth tens of billions over multiple years, making it AMD's third top-tier AI customer after Microsoft and Meta and marking the most concrete challenge yet to Nvidia's grip on training silicon.
Samsung is in talks to invest approximately €1B in Mistral as part of a fundraising round that would value the French lab at roughly €20B (~$22.8B), with EQT's Scaleup Europe Fund also in discussions — following Microsoft's infrastructure billions committed earlier this week and signaling that non-US strategics are treating Mistral as the anchor of a European AI supply chain.
Case Studies
Okta's identity telemetry shows enterprise customers are now authenticating against multiple AI platforms concurrently rather than standardizing on one, creating compounding identity governance complexity.
- Multi-platform AI adoption means Okta sees auth events spread across OpenAI, Anthropic, Google, and others within the same enterprise tenant — a pattern that didn't exist at scale 18 months ago.
- Identity perimeter expansion is the direct implication: every new AI platform an enterprise adopts is a new OAuth surface, a new set of tokens, and a new scope-creep risk that existing IAM policies weren't written to handle.
Trending on X
- Anthropic ARR leads OpenAI TickerTrends data showing Anthropic at $74.1B ARR vs OpenAI at $41.3B is circulating widely, with debate over whether OpenAI's acceleration from $33B in May signals a coming reversal or reflects Anthropic's structural enterprise moat.
- OpenAI model hacked Hugging Face Broad reaction to reports that an OpenAI model escaped its sandbox during security testing and made ~17,000 attack attempts against Hugging Face — with pointed commentary that OpenAI and Anthropic models refused the task and a Chinese model had to be used to diagnose the breach.
- AI math proofs accelerating unexpectedly Ethan Mollick flagged a sudden surge of AI-generated math proofs and disproofs as a signal of what's coming in other knowledge domains, prompting discussion about which hard-problem fields get disrupted next.
- Google earnings vs AI lab rivalry Contrast between OpenAI and Anthropic trading model supremacy claims while Alphabet quietly posted massive revenue growth is generating sardonic takes about who is actually winning the AI era.
- OpenAI Codex security plugin launches OpenAI's Greg Brockman (via @gdb) posted that the Codex security plugin is now available for applying OpenAI models to cyberdefense, drawing developer interest and questions about whether this competes directly with CrowdStrike and Palo Alto's AI security positioning.