Rosetta Intel

Latest developments at OpenAI / Anthropic / Google (automated collection roundup)

OpenAI

GPT-5.6 released (7/9)— three pricing tiers: Sol ($5 in / $30 out per million tokens), Terra ($2.5/$15), and Luna ($1/$6). Launched alongside it is ChatGPT Work: an agent product fusing Codex and ChatGPT that can execute complete tasks over long horizons across applications and files.
Axios· Nextgov
Custom instructions limit raised from 1,500 to 5,000 characters (Pro/Enterprise/Business/Edu).
Unified search: across chats / projects / images / documents (web + iOS + Android).
ChatGPT returns to WhatsApp in the EEA (+1-800-242-8478, no account required).
Releasebot

Claude for Teachers (7/14)— certified U.S. K-12 teachers get free access to advanced Claude tools, teaching skills, and curriculum connectors aligned to all 50 state standards.
Chalkbeat
Claude Science— a research workbench: 60+ scientific database connectors, prebuilt toolkits for genomics / protein structure / chemistry, running on existing Claude models (including Opus 4.8). The AI for Science grant (up to $30k in credits plus $2k of Modal compute) closed applications 7/15, with notification by 7/31 and a project period of 9/1–12/1.
TechCrunch· Endpoints
Self-service HIPAA configuration— Enterprise/API organizations can complete BAA review and one-click enablement themselves.
A major Claude Code stability update: streaming text output from subagents, improved permission and upload handling, background agent reporting, and faster terminal rendering.
Releasebot

Gemini 3.5 Pro— multiple outlets report a 7/17 target release, with the 2.5 Pro architecture abandoned for a ground-up rebuild: 2M-token context, a "Deep Think" reasoning layer, and autonomous workflows; Google has not officially confirmed any date or specification. The delay plus talent attrition (Noam Shazeer, John Jumper and others moving to OpenAI/Anthropic) was reported to have cost Alphabet roughly $225B in market value.
TechTimes· The Agent Report
Nano Banana 2 Lite— the fastest and cheapest Gemini image model is now live.
Gemini Omni Flash— the natively multimodal model enters public API preview (for enterprises and developers).
Google blog

Microsoft credited part of July's record 570 patches to an internal AI vulnerability-discovery system — AI bug hunting has reached production scale.
The "TuxBot v3 Evolution" IoT botnet shows traces of LLM-assisted development (the attacker-side counterpart).
A Microsoft Copilot command injection entered the CISA KEV catalog (remediation deadline 7/17) — AI products themselves are now targets of in-the-wild exploitation.

Sources: openai.com/news · anthropic.com/news · deepmind.google · blog.google (aggregated via web search)