Model Release
Google Launches Gemini 3 Deep Think Model
Google released Gemini 3 Deep Think on March 26, a frontier model designed for scientific and engineering tasks such as spotting logical flaws in math papers and optimizing fabrication methods. The model is live in the Gemini app for Ultra subscribers, with early API access now available. It represents Google's push into specialized reasoning models alongside its broader Gemini family. Source →
Policy
OpenAI Expands Bug Bounties to AI Safety
OpenAI announced a Safety Bug Bounty Program on March 26, extending its existing bounty framework to cover AI misuse and safety risks. The company also detailed its Model Spec framework, which outlines principles for model behavior, instruction-following, and conflict resolution. The move signals a more structured approach to external safety auditing as frontier models grow more capable. Source →
Enterprise
Amazon and OpenAI Build Stateful AI Runtime
Amazon partnered with OpenAI on March 26 to co-create a Stateful Runtime Environment on Amazon Bedrock, focused on memory and tool-use infrastructure for AI applications. The environment is designed to support persistent agent state across sessions, a key requirement for complex agentic workflows. Availability is planned for the coming months. Source →
Model Release
Google Ships Lyria 3 Music Generation Models
Alongside Gemini 3 Deep Think, Google released Lyria 3 and Lyria 3 Pro on March 26, its latest music generation models. The dual release reflects Google's strategy of shipping specialized models across modalities rather than concentrating on a single general-purpose system. Details on pricing and availability beyond the Gemini ecosystem have not yet been disclosed. Source →
Benchmark
NVIDIA Nemotron 3 Super Hits 0.8 GPQA
NVIDIA's Nemotron 3 Super, a 120-billion parameter model, posted a GPQA score of 0.8 following its release around March 11. The score places it among the strongest performers on the graduate-level reasoning benchmark this quarter. The model adds to NVIDIA's growing presence in the foundation model space beyond its dominant hardware business. Source →
Model Release
Mistral Small 4 Posts 0.7 GPQA Score
Mistral released Small 4 around March 17, achieving a GPQA score of 0.7 on the graduate-level reasoning benchmark. The model continues Mistral's strategy of shipping competitive smaller models that can run efficiently on constrained infrastructure. It arrived during a record-setting Q1 that saw 255 model releases across the industry. Source →
Policy
OpenAI Model Spec Details Behavior Principles
OpenAI's newly published Model Spec framework lays out explicit principles for how its models should handle instruction-following, behavioral boundaries, and conflict resolution between user requests and safety constraints. Released alongside the Safety Bug Bounty Program on March 26, the document represents one of the most detailed public accounts of how a frontier lab governs model behavior. The framework could set a precedent for industry-wide transparency on alignment decisions. Source →
Trend
Late March AI Release Pipeline Goes Quiet
No major new model releases were announced between March 27 and 31, following a burst of activity earlier in the month that included GPT-5.4, DeepSeek V4, and multiple Qwen3.5 variants. The pause comes after Q1 2026 hit a record 255 model releases. The lull may reflect labs shifting resources toward agentic frameworks and infrastructure rather than raw model launches. Source →
