Rosetta Intel

Daily movements at OpenAI / Anthropic / Google DeepMind (with a focus on security-relevant items, marked 🛡)

OpenAI

GPT-5.6 launching soon— expected in the early-July window, positioned around new benchmarks for speed, reliability, and output accuracy.
In-house inference chip "Jalapeño"— OpenAI announces a custom inference chip, continuing its path of cutting inference costs and reducing dependence on external compute.
Lockdown Mode (since 2026-02)— a hardened mode for ChatGPT; OpenAI publicly acknowledged that prompt injection in AI browsers "may never be fully fixed" — a significant industry statement on current agent security.

Claude Sonnet 5 released— performance approaching Opus 4.8, introductory pricing of $2/$10 per million tokens (through August 31), able to autonomously use tools such as the browser and terminal.
Security angle— models autonomously invoking browsers/terminals further widen the attack surface for prompt injection and tool abuse; combined with the Axis Intelligence Q2 report ("Claude Opus 4.7 has the strongest single-turn jailbreak resistance and the lowest aggregate vulnerability count"), Anthropic leads on single-turn robustness, but injection in agent scenarios remains an industry-wide unsolved problem.

Gemini 3.5 Pro delayed to July 17— abandoning the original 2.5 Pro architecture for a complete rebuild: 2M-token context, a "Deep Think" reasoning layer, and autonomous workflow capabilities; aimed at GPT-5.6 and Anthropic Fable 5, with a focus on mathematical reasoning, SVG scene generation, and image quality.
NanoBanana 2 Lite— image model, <4-second generation, from $0.034 per 1,000 images, with quality above the original NanoBanana.

🛡 = security-relevant. Full daily brief: intel/daily/2026-07-12.zh.md.