Welcome to the agent platform research briefing for Sunday, June 29th, 2026.

**GPT‑5.6 Sol preview deepens — OpenAI publishes system card** — On June 28th, OpenAI published a detailed blog post titled "Previewing GPT‑5.6 Sol: a next-generation model," expanding on the limited preview that began June 26th. The post confirms three models: Sol is the flagship with a new "max reasoning" effort and an "ultra mode" that uses subagents to accelerate complex work beyond what a single agent can do. Terra matches GPT‑5.5 at roughly half the price — $2.50 per million input tokens and $15 output. Luna is the cheapest tier at $1 and $6. Sol scored 91.9% on Terminal‑Bench 2.1, beating Claude Mythos 5. On ExploitBench‑2, Sol is competitive with Mythos using only about one-third the output tokens. The cyber safety assessment notes Sol did NOT cross the Cyber Critical threshold — it found bugs and exploitation primitives in Chromium and Firefox but did not produce a functional full-chain exploit autonomously. OpenAI says it plans broad GA "in the coming weeks" after the limited government-coordinated preview with trusted partners.

**Google restricts Meta's Gemini access — compute capacity crunch** — Bloomberg reported June 28th, citing the Financial Times, that Google has placed limits on Meta Platforms' use of its Gemini AI models. The reason: Google could not supply the computing capacity Meta was demanding for the partnership, and this constraint is now delaying some of Meta's AI projects. This is significant because the Google-Meta AI deal was announced as a major strategic partnership earlier this year — Meta gaining access to Gemini for its messaging and social apps, Google getting distribution. The compute constraint signals that even hyperscaler partnerships are hitting infrastructure walls. For the AI platform landscape, it means Meta may need to accelerate its own Llama development or seek additional compute partners — opening the door for AWS, Oracle, or xAI to compete for Meta's AI compute needs.

**Austria proposes EU host Anthropic — jurisdiction shopping for the kill switch** — Also on June 28th, Austria's State Secretary for Digitalization Alexander Pröll sent a formal letter to EU Tech Commissioner Henna Virkkunen urging the bloc to explore establishing Anthropic within the European Union. The letter offers "legal certainty, market access, capital and a set of values that suits this company." This is a structural response to the US export control order that silenced Fable 5 and Mythos 5 for foreign users on June 12th. Austria isn't trying to buy a model — it's trying to relocate the permission layer. If Anthropic were headquartered in the EU, US export controls would not directly apply. The letter acknowledges there would be skepticism, but Pröll frames this as sovereignty, not industrial policy. The move signals that every government outside the US now understands frontier AI access is a sovereign dependency that can be revoked without warning.

**Claude Code adds sandbox credential blocking** — Anthropic's latest Claude Code release introduces a sandbox.credentials setting that blocks sandboxed commands from reading credential files and secret environment variables — a direct response to the growing class of MCP token hijacking and credential leakage attacks. The release also adds organization-configured model restrictions to the model picker, command-line flags, and environment variables, with a "restricted by your organization's settings" message. Other fixes include improved resume handling, structured output stability, remote MCP hang resolution, and VS Code responsiveness improvements. This is part of an accelerating security hardening cycle for agent coding tools — following MITiga's five-step Claude Code MCP token hijack demo and Tenet Security's Agentjacking attack demonstration from earlier this month.

That's the briefing for today.