AI Frontier · Jul 21, 2026

Product and model developments at OpenAI / Anthropic / Google, with particular attention to security. No major new releases in the last 24 hours; below are the ongoing July storylines.

OpenAI

• The GPT-5.6 family (Sol / Terra / Luna) is publicly released(staged rollout from 7/9). The three tiers: Sol as the workhorse, Terra as the mid tier, Luna as the budget tier; Sol is claimed to be 54% more token-efficient on coding tasks. OpenAI calls 5.6 its "strongest cybersecurity model yet," reaching frontier performance with fewer tokens — worth watching from both the offensive and defensive automation sides.
  TechCrunch· CNBC
• GPT-Live voice model: it can listen and speak at the same time, making interaction closer to real conversation.
  Axios

Anthropic

• Claude Sonnet 5(released 6/30, the default model for Free/Pro from 7/1): billed as the most agentic Sonnet, approaching flagship Opus 4.8 across many tasks.
• Claude Code stability and security update: tightened permission validation, safer Bash / PowerShell handling, background session cleanup, telemetry and plugin reliability improvements, and a new EndConversation tool — directly relevant hardening for automation workflows driven by Claude Code (including this workbench).
  Releasebot — Claude Code
• Claude for Teachers: free for certified U.S. K-12 teachers; on the enterprise side, new self-service HIPAA configuration (BAA review plus one-click enablement) and an Admin API (Enterprise organization management, in beta).
  Chalkbeat

Google DeepMind / AI

• Gemini 3.5 Pro delayed: it missed the June window, with a rumored 7/17 target (not officially confirmed); the architecture was reportedly rebuilt from scratch, emphasizing mathematical reasoning, SVG scene generation, a 2M-token context window, and a Deep Think reasoning layer, positioned against GPT-5.6 and Anthropic Fable 5. The delay came alongside the departure of several core researchers, and Alphabet's market value fell by roughly $225B at one point.
  The Agent Report· BigGo

🛡 = security-related