KDCube
← AI Industry News

MCP v2 Beta SDKs Land in Four Languages; Gemini 3.5 Pro Ships Tomorrow

The stateless MCP 2026-07-28 spec now has beta SDKs for Python, TypeScript, Go, and C#, so teams can test the rewrite before the July 28 cutover — and it's opt-in, so nothing breaks today. Meanwhile Gemini 3.5 Pro is reported to launch July 17 with a 2M-token window, and OpenAI's GPT-5.6 Sol bakes para...

Highlights

  • The 2026-07-28 MCP spec now has beta SDKs across all four Tier 1 languages — Python v2.0.0b1, TypeScript v2 (split packages), Go v1.7.0-pre.1, and C# 2.0.0-preview.1 — so teams can finally test the stateless rewrite against real workloads before the July 28 cutover, 12 days out. (Model Context Protocol Blog)
  • Google's fully rebuilt Gemini 3.5 Pro is reported to launch July 17 — tomorrow — with a 2M-token context window and a "Deep Think" reasoning layer, though every headline spec is still third-party reporting, not an official Google post. (BigGo Finance)
  • OpenAI's GPT-5.6 Sol went GA July 9 with an in-model "Ultra" mode that spawns parallel subagents on one problem — a multi-agent primitive baked into the base model, not the orchestration layer. (OpenAI)

Key Signals

  1. MCP's stateless spec is now testable, not just scheduled — RC + beta SDKs, June–July 2026
    The maintainers stress the migration is opt-in: "nothing breaks today, and nothing breaks on July 28 either." The RC removes the initialize handshake and protocol-level session, moves version/capability data into _meta on every request, and adds routable Mcp-Method/Mcp-Name headers so a plain round-robin load balancer can front an MCP server with no sticky sessions or shared session store. (Model Context Protocol Blog, Beta SDKs)
  2. Gemini 3.5 Pro's rebuild aims squarely at long-context agent work — reported launch July 17
    DeepMind reportedly scrapped the 2.5 Pro architecture for a full rebuild targeting a 2M-token window and autonomous workflow execution. Treat specs as unconfirmed until Google posts them, but the timing — one day out — makes this the frontier-model story to watch this morning. (BigGo Finance)
  3. GPT-5.6 pushes multi-agent into the model itself — GA July 9
    Sol's Ultra mode runs parallel subagents inside a single call (4 by default on Work/Enterprise tiers, a multi-agent beta on the Responses API), while Terra targets prior-flagship quality at lower cost and Luna serves the fast tier. Each subagent burns tokens independently, so Ultra is for genuinely parallelizable work — routine calls through it are expensive overkill. (OpenAI)

Why It Matters / What To Watch

  1. Migrate MCP servers on the betas now — don't wait for July 28.
    • Pin exact beta versions and validate stateless behavior in a branch; the SDKs are explicitly labeled beta and APIs can still shift. (Model Context Protocol Blog)
    • If your deployment relied on sticky sessions or gateway deep-packet inspection, re-test routing on Mcp-Method headers before the final spec locks. (Model Context Protocol Blog)
  2. In-model multi-agent changes your cost and eval math.
    • Meter Ultra-style parallel subagent calls separately — a single request can cost several times a standard call, so budget controls that assume one reasoning chain will under-count. (OpenAI)
    • Watch whether Gemini 3.5 Pro's 2M context and autonomous workflows actually land as reported tomorrow; hold procurement decisions until Google publishes official specs and pricing. (BigGo Finance)
  3. The framework layer is racing to keep up.
    • Microsoft's Agent Framework recently shipped progressive MCP tool discovery and session-isolation improvements across its Python and .NET builds — a signal that orchestrators are adapting to both stateless MCP and in-model subagents at once. (Microsoft / GitHub)

Quick Links