OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom-designed inference processor, built in under a year using OpenAI's own AI models.

• Designed for LLM inference at scale: Jalapeño is purpose-built for OpenAI's serving stack, not general compute, which is the architectural distinction that matters for cost and latency.
• Still in testing, preliminary results promising: OpenAI reports meaningful efficiency gains but has not published final benchmark numbers, leaving the full performance story open.

Google launched Gemini 3.5 Flash with computer use natively integrated, allowing developers to build agents that see and act across desktop and mobile environments without a separate model.

• Previously a standalone model, now baked in: The computer use capability was a separate Gemini 2.5 endpoint; folding it into the main Flash model removes an integration step and improves agentic task performance.
• Significant price increase over predecessor: Gemini 3.5 Flash costs 5x more on input tokens than Gemini 2.5 Flash, making workload economics the primary adoption decision.

Notion shipped external agent support today, letting users tag Claude or Cursor directly in docs, threads, and databases to execute tasks end-to-end without leaving Notion.

• Cursor SDK enabled a two-week integration: Notion's engineering team used the Cursor SDK to embed coding agents into their product, demonstrating how third-party platforms can add agent capability without building the stack themselves.
• Agents plan, build, test, and open PRs: The integration goes beyond autocomplete — Cursor takes assigned tasks through a full development loop before surfacing results to the user.

Slack shipped an MCP client inside Slackbot, enabling AI agents from any MCP-compatible server to participate in conversations and take actions directly within Slack channels.

• Multiplayer agent work in existing channels: Agents join where the work already happens rather than requiring users to context-switch to a separate AI interface.
• Extends Salesforce's agentic Slack strategy: Combined with the existing Agentforce integration, this positions Slack as the coordination layer for multi-agent enterprise workflows across the Salesforce platform.

GitHub reported that Copilot usage reached a record high in June following its shift to pay-as-you-go pricing.

• Consumption model unlocked latent demand: The move away from flat seat pricing allowed organizations to expand usage across more developers without committing to per-seat costs for infrequent users.
• Simultaneous free-tier tightening: GitHub also removed model selection from the free Copilot tier today, pushing users who want model choice toward paid plans and sharpening the upsell path.