KDCube
← AI Industry News

Gemini 3.5 Pro Ships With 2M Context; MCP Cutover 11 Days Out

Gemini 3.5 Pro lands today with a reported 2M-token context window, reopening the retrieve-vs-stuff trade-off for RAG and memory-backed apps even as retrieval quality still gates cost. Meanwhile the stateless MCP 2026-07-28 cutover is 11 days out, and OpenAI's GPT-5.6 "Sol" keeps pushing subagents into the model runtim...

Highlights

  • Gemini 3.5 Pro ships today with a reported 2M-token context window, landing on the July 17 date flagged in yesterday's issue — the first frontier release to make million-plus retrieval-in-context a default rather than a preview tier (Google DeepMind).
  • The stateless MCP 2026-07-28 spec clock reads 11 days; the four beta SDKs (Python, TypeScript, Go, C#) are the last opt-in window before per-server session state stops being guaranteed (Model Context Protocol).
  • OpenAI's GPT-5.6 "Sol" continues to draw builder attention for baking parallel subagent spawning into the model itself, reframing "multi-agent" from an orchestration-layer job to a serving-layer one (OpenAI).
  • Long-context economics are back in the operator conversation: 2M windows change RAG cost math, but not RAG's need — retrieval quality still gates what you can afford to stuff in (Google DeepMind).

Key Signals

  1. Gemini 3.5 Pro goes GA with a 2M-token window — 2026-07-17

    The launch flagged as "tomorrow" in the July 16 issue is now live, pushing million-token context from experimental into a shipping default. For teams building memory-backed assistants and document-heavy RAG, this reopens the "retrieve-then-stuff vs. stuff-everything" trade-off — but the answer is still workload-dependent, since bigger windows raise per-call cost and latency even when they cut retrieval complexity (Google DeepMind).

  2. MCP stateless migration enters its final stretch — July 28 cutover, 11 days out

    With beta SDKs available in four languages, the remaining work is unglamorous and unavoidable: drop Mcp-Session-Id state assumptions, expose RFC 9728 resource metadata, and stand up OAuth 2.1. Because the new spec is opt-in, nothing breaks on July 17 — but teams that wait risk a rushed cutover against connectors they don't fully control (Model Context Protocol).

  3. Model-native subagents keep pressure on orchestration frameworks — ongoing

    GPT-5.6 Sol's in-model parallel subagents raise a live question for LangGraph, CrewAI, and AG2 users: how much orchestration logic belongs in your framework versus the model runtime? The answer shapes where governance, tracing, and budget controls have to live (OpenAI).

Why It Matters / What To Watch

  1. Long context does not retire your retrieval stack
    • Re-run cost/latency benchmarks before collapsing a RAG pipeline into a 2M-token prompt — the cheaper-per-token headline hides a higher per-call bill at scale (Google DeepMind).
    • Keep evaluation coverage on retrieval quality; a bigger window amplifies bad context as readily as good (Model Context Protocol).
  2. Treat the MCP deadline as a this-week task, not a next-week one
    • Inventory every MCP server and connector that assumes session stickiness, and test it against the beta SDK now while rollback is trivial (Model Context Protocol).
    • Confirm your identity provider path for OAuth 2.1 and RFC 9728 metadata before the 28th, not after (Model Context Protocol).
  3. Decide where your agent control plane lives
    • If subagents become a model feature, verify your observability and budget guardrails still capture spawned work you didn't explicitly orchestrate (OpenAI).

Quick Links