Rosetta Intel

Product and model developments at OpenAI / Anthropic / Google DeepMind. 🛡 = security-related.

OpenAI

GPT-5.6 (Sol / Terra / Luna)has been generally available across ChatGPT, the API, and Codex since 07-09, ending a 13-day preview. Terra is the default tier for standard paid users, Sol targets Pro (flagship, including the Ultra subagent mode and the Max reasoning tier), and Luna targets high-concurrency / budget-conscious workloads.
In early July the company unveiled its in-house Jalapeño inference chip, aimed at inference cost and supply independence.
As subagent autonomy grows, indirect prompt injection and agent privilege escalation remain the top security concerns.

Claude Sonnet 5 shipped on 07-01— positioned as the most "agentic" model to date: it can drive a browser and terminal autonomously, approaching Opus 4.8 quality at markedly lower cost.
Over the same period the U.S. government lifted the national-security restrictions on Fable 5 / Mythos 5.
Four senior Google researchers were reported to have moved to Anthropic recently, underscoring frontier talent mobility.
Tool permission boundaries and auditing under long-horizon agentic runs remain a focus.

Gemini 3.5 Pro GA slipped to 07-17— after enterprise testers reported reasoning and coding performance problems, the model underwent a full architectural rebuild (abandoning the 2.5 Pro base): a 2M-token context window, a Deep Think reasoning layer, and autonomous workflow features, positioned against GPT-5.6 and Fable 5.
At Cloud Next '26, Google expanded the Gemini Enterpriseportfolio: a unified end-to-end platform for building, orchestrating, and governing AI agents.

Both OpenAI and Anthropic are emphasizing injection mitigation within the agent-autonomy narrative; this week GitLost (GitHub Agentic Workflows) again demonstrated that agents plus untrusted input equal a data-exfiltration surface. See the "AI Security" section of the main brief.

Source configuration: see intel/sources.yaml