AI Frontier · Jul 2, 2026
Product and model updates from OpenAI / Anthropic / Google DeepMind (security-related items marked 🛡)
OpenAI
- GPT-5.6 Preview System Card(6/26) published, alongside a preview of the next-generation model GPT-5.6 "Sol".
- OpenAI × Broadcomannounced an inference chip optimized for LLMs(6/25), accelerating its in-house inference hardware strategy.
- Research: published "core dump epidemiology — fixing an 18-year-old bug" (6/30), an engineering/reliability-oriented piece.
- Earlier: OpenAI has already filed confidentially for an IPO.
Anthropic
- Claude Sonnet 5— the most agentic Sonnet to date: stronger reasoning, tool use, coding, and autonomous task handling; live across all plans, Claude Code, and the Claude Platform, with introductory pricing and higher rate limits.
- Seoul office opening + Korean AI ecosystem partnerships(6/17).
- Issued a statement (6/12) on the U.S. government directive to "suspend access to Fable 5 / Mythos 5"; the same day it published the first Anthropic Public Record results and announced a TCS × Anthropicpartnership to bring Claude into regulated industries.
Google DeepMind / AI
- Gemini 3.5 Flashreleased — frontier agentic/coding performance, beating Gemini 3.1 Pro on key benchmarks, leading on multimodal, and often at less than half the cost of comparable models; aimed at long-horizon agentic tasks.
- Pricing: the Ultra subscription drops from $250 to $200/month, with a new $100/month Developer tierfor engineering/professional users.
- Managed Agents on the Gemini API (public preview)— developers can build and deploy stateful autonomous agents running inside isolated, Google-hosted Linux sandboxes.
- DeepMind brought on 20+ Contextual AI researchers through a licensing deal worth roughly $80–90M.
Security Highlights This Issue
- Anthropic'sstatement on the government-ordered suspension of Fable 5 / Mythos 5 access touches on export/access governance for frontier models.
- Google'sisolated Linux sandbox design for Managed Agents is a productized expression of agent execution security (sandboxing, least privilege).
Source: the ai_frontier category in intel/sources.yaml