The Stack
· AUGUST 15, 2026 · Ace Workflow
In this issue
ChatGPT · GitHub · Cursor · Notion · Databricks · Vercel · DeepSeek · Anthropic
The scoreboard
● ChatGPT adds Model Context Protocol for custom connectors (Pro, Team, Enterprise, Edu), opening proprietary system access on August 15.
● GitHub ships Grok 4.6 and Gemini 3.7 Flash for Copilot across agentic coding workflows, plus Agent Plugins 1.0 standard on August 14 and 12.
● Cursor introduces Grok 4.6 for long-running agents and complex coding on August 12, starting Cloud Agents 3× faster with background builds on August 13.
● Notion ships a model picker with scorecards, favorites, and effort controls on August 14.
● Databricks adds Gemini 3.7 Flash to Model Serving on August 13 and expands ABAC GRANT policies to model services, provider services, MCP services, agent services, and skills on August 11.
● Vercel brings Gemini 3.7 Flash to AI Gateway with one API key, automatic fallbacks, and request traces on August 13.
● OpenAI releases GPT-5.6 with stronger agent performance at lower cost, plus Responses API controls and Ultrafast mode (up to 14× faster) on August 13.
● DeepSeek ships V4-Pro GA across app, web, and API with stronger agent performance and native OpenAI Responses API support on August 13.
● Anthropic adds watermarking to future Claude text outputs for EU AI Act compliance on August 14, with no visible impact on quality, speed, or pricing.
Fig 1 · Release velocity, last 3 months vs prior 3
The lead
The model-picker wars are heating up. ChatGPT added Model Context Protocol support for custom connectors, GitHub shipped three new models in a week (Grok 4.6, Gemini 3.7 Flash, MAI-Code-1.1-Flash), Cursor brought in Grok 4.6, and Notion built a full model picker with scorecards and favorites. Vercel and Databricks both added Gemini 3.7 Flash. The pattern is clear: vendors who want to own the agent layer are opening up model choice, betting that workflow lock-in matters more than model lock-in.
But not everyone is building the same way. OpenAI shipped GPT-5.6 with tighter Responses API controls and launched Ultrafast mode at 14× speed through Cerebras, optimizing for native workflows instead of interchangeable backends. DeepSeek followed the same path, shipping V4-Pro with native OpenAI API support but focusing on its own deployment. Anthropic added watermarking for compliance, a feature that only works if you stay on Claude.
Read the full issue →
The rest of this issue lives on the stack hub: what it replaces, and what to do about it.