Rosetta Intel

Product and model developments at the three frontier labs; 🛡 marks security-related items.

OpenAI

The GPT-5.6 family (Sol / Terra / Luna)— three size tiers, a 1M-token context window, a knowledge cutoff of February 2026, and input pricing of $1–$5 per million tokens.
GPT-Live— voice evolves from "taking turns" to "listening and speaking at once," able to handle interruptions and real-time translation.

Strategic partnership with AMD (July 22)— deploying up to 2 GW of AMD Instinct MI450 series GPUs (the AMD Helios rack design), with the first gigawatt expected to come online in the first half of 2027; AMD plans a strategic equity investment in Anthropic of up to $5 billion.
Fable 5 / Mythos 5 restored worldwide (July 1)— back after U.S. export controls were lifted, with Anthropic describing the added cybersecurity classifier as its "strongest protection."
Voice mode expanded— covering Claude Opus/Sonnet/Haiku, able to connect to tools such as Gmail and Slack, with support for more languages.
Anthropic Economic Index connector— AI usage data (occupations, tasks, trends) can now be queried directly within a conversation.

Gemini 3.5 Pro (July 17)— the 2.5 Pro architecture was abandoned and rebuilt from scratch; a 2 million token context window, a Deep Think reasoning layer, and autonomous workflow capabilities.
Gemini 3.6 Flash (July 21)— a lightweight, high-speed tier released.

Anthropic emphasized the strengthened cybersecurity classifier when Fable 5 / Mythos 5 returned — echoing the trend of frontier models being used on both the offensive and defensive sides (see the Unit 42 report in the main brief's AI Security section).
Agent and tool integration capabilities in frontier models keep expanding (Anthropic voice connecting to Gmail/Slack, Gemini autonomous workflows, GPT-Live), and the indirect prompt injection attack surface expands with them — OWASP LLM01 still has no root fix.

For sources, see the AI Frontier and AI Security section links in the main brief 2026-07-25.en.md / .zh.md.