Summary

Today's ten signals cluster around three platform vendors reshaping how developers reach AI models. GitHub is narrowing its surface: it retired GitHub Models entirely (playground, catalog, inference API, and BYOK) and deprecated older Gemini models in Copilot, funneling developers toward Microsoft Foundry and Copilot. Vercel is doing the opposite, using AI Gateway as a fast-refreshing model hub with updated DeepSeek V4 Flash weights, a new Qwen 3.8 Max, and a 10x capacity bump for Poolside's Laguna S 2.1, while extending its Chat SDK on Teams and adding richer workflow-run search in Observability. Cloudflare split the difference on capacity, defaulting Workers Builds to Node.js 24, shipping a Browser Run Playground, but moving three heavy Workers AI models behind its paid plan. The through-line: capacity and agent-workload tuning are becoming the real differentiators among model gateways.

Key themes

  • GitHub consolidates its AI surface: retiring GitHub Models and deprecating older Gemini versions in Copilot pushes developers toward Microsoft Foundry for raw model access and Copilot for in-workflow AI, trading a neutral multi-model playground for a tighter Microsoft-aligned funnel.
  • Vercel positions AI Gateway as a self-refreshing model hub: updated DeepSeek V4 Flash weights, the new Alibaba Qwen 3.8 Max with automatic fallbacks, and a 10x Laguna S 2.1 capacity increase all let teams pick up provider improvements and higher throughput without changing integrations.
  • Capacity is emerging as the key constraint for agentic workloads, and vendors are diverging: Vercel expands headroom (10x on Laguna S 2.1) while Cloudflare gates its most resource-intensive Workers AI models behind a paid plan to manage load.
  • Agent- and workflow-oriented tooling is deepening across the stack, from DeepSeek weights tuned for agent applications and Teams-native Chat SDK interaction primitives to richer, shareable workflow-run search in Vercel Observability.
  • Cloudflare balances developer experience with capacity management: Node.js 24 becomes the default for Workers Builds and a Browser Run Playground speeds testing, even as heavy inference models move behind the paid tier.

Notable items

  • GitHub retired GitHub Models as of July 30, 2026, shutting down the playground, model catalog, inference API, and BYOK for all customers, and named Microsoft Foundry and Copilot as successors (medium impact).
  • GitHub Copilot deprecated Gemini 2.5 Pro and Gemini 3 Flash across Chat, inline edits, ask/agent modes, and completions, replacing them with Gemini 3.1 Pro (Preview) and Gemini 3.6 Flash; enterprise admins must opt in via policy.
  • Cloudflare moved Kimi K2.6, Kimi K2.7-Code, and GLM-5.2 behind the Workers Paid plan (from $5/month) to cut 429/out-of-capacity errors, while many smaller models like GLM-4.7-Flash and Gemma-4-26B stay free (medium impact).
  • Vercel raised Laguna S 2.1 usage limits on AI Gateway roughly 10x across free and paid tiers for high-volume agentic coding and long-running tasks.
  • Vercel added Alibaba's Qwen 3.8 Max to AI Gateway with a unified API, automatic provider fallbacks, and per-request cost and token tracking; it also shipped updated DeepSeek V4 Flash weights tuned for agent workloads.
  • Cloudflare set Node.js 24.18.0 as the default runtime for Workers Builds (with 22.23.2 preinstalled) and added a Browser Run dashboard Playground for testing Quick Actions against live browsers before deploying.
  • Vercel deepened its enterprise agent surface with Chat SDK reactions and ephemeral messages on Microsoft Teams, and expanded Observability with workflow-run search by workflow, environment, deployment ID, region, or custom attributes plus shareable query URLs.

Source coverage

Source rows used: 10