Good morning Rich. This is your daily agent platform research briefing for Thursday, June 26th, 2026.

**OpenAI GPT-5.6 delayed by Trump administration request** — The Information reported on June 25 that OpenAI has agreed to stagger the release of GPT-5.6 after formal requests from the Office of the National Cyber Director and the Office of Science and Technology Policy. Commerce Secretary Howard Lutnick advised against a broad launch without cross-agency approvals. GPT-5.6 will initially ship only to government-approved enterprise partners, with a broader public rollout expected "a couple weeks" later. Sam Altman told employees this is not OpenAI's "preferred long-term model" for releases. The administration views GPT-5.6 as on par with Claude Fable 5 and Mythos 5 in cybersecurity capability — the same rationale behind the Anthropic export control directive on June 12. This marks the second major government intervention in a frontier model release in two weeks, signaling a new era of federal oversight over when and how advanced AI reaches the public.

**OpenAI unveils Jalapeño — its first custom AI inference chip, built with Broadcom** — OpenAI and Broadcom jointly announced a purpose-built inference processor called Jalapeño. Engineering samples were delivered to Sam Altman and Greg Brockman at OpenAI's San Francisco headquarters on June 24. The development timeline is remarkable: nine months from initial design to tape-out, compared to the typical 18-to-36 months for custom chip development. Early testing shows roughly 50 percent cost savings per inference token versus current-generation GPUs, according to Broadcom CEO Hock Tan. The chip was manufactured by TSMC and enables gigawatt-scale data center deployments with Microsoft and other partners beginning this year. This makes OpenAI the latest frontier lab to build its own silicon — joining Meta, Amazon, Google, and reportedly Anthropic, which is weighing a chip of its own. But OpenAI's unique position — designing both the model and the inference hardware — gives it optimization latitude that GPU vendors cannot match across thousands of different customers. If the 50 percent cost reduction holds, this could be the single development that most directly changes what AI services cost to use.

**Anthropic Fable 5 remains offline on Day 14, Commerce deadline is today** — No restoration announcement from Anthropic as of Wednesday June 25. Anthropic staff confirmed zero Fable 5 or Mythos 5 traffic. The bipartisan House letter deadline for Commerce Secretary Lutnick to provide written justification of the export controls was June 26 — that's today. No public response has been issued. Polymarket prices for early restoration have collapsed while longer-term options surged. Meanwhile, Anthropic continues to ship normal products — Claude Tag for Slack launched June 23 — keeping business momentum while the Fable 5 standoff resolves at the highest levels of government. The structural timeline, unchanged: July 8 and August 1 as potential resolution windows.

**Salesforce launches Agentforce Help Agent with pay-per-resolution pricing** — On June 25, Salesforce unveiled a pre-built AI customer service agent on Agentforce that companies can connect to their knowledge base in minutes. The headline is the pricing model: $2 per autonomously resolved issue. If the agent escalates to a human or the customer abandons the chat, Salesforce doesn't charge. Salesforce's own help portal has handled 4.3 million inquiries with a 70 percent resolution rate. The agent supports chat, portals, web, text, and voice channels from a single configuration screen. This is a direct signal that the agentic enterprise is shifting from seat-based subscriptions to outcome-based pricing — a model that could reshape how organizations budget for AI.

**Google makes Interactions API the default for Gemini agents** — Google DeepMind's Interactions API is now the primary interface for Gemini models, replacing the old generateContent endpoint. The shift moves Google from flat text outputs to a step-based execution timeline, closer to an agent runtime trace. New capabilities include Managed Agents with Linux sandboxes, background execution for long-running tasks, and tool chaining with Google Search and Maps. Developers also get a choice between Flex mode, which cuts costs 50 percent, and Priority mode for speed-critical operations. The old API still works, but all new agent features will ship exclusively through Interactions going forward. Google has published a migration guide.

That's the briefing for today. Five stories: GPT-5.6 gets a government gatekeeper, OpenAI rolls its own silicon, Fable 5 waits on Congress, Salesforce prices AI by the resolution, and Google's API bet is all-in on agents. See you tomorrow.