Gemini API April 2026 Update: Flex/Priority Tiers, Deep Research with MCP, embedding-2 GA, gemini-3.1-flash-tts-preview

April 2026 was a substantive month for the Gemini API, with five changes worth pulling forward: Flex and Priority inference tiers (Apr 1), gemini-3.1-flash-tts-preview (Apr 15), a major Deep Research agent update with MCP server integration (Apr 21), gemini-embedding-2 GA (Apr 22), and the gemini-robotics-er-1.5-preview shutdown (Apr 30) in favor of 1.6-preview. Three of these — tiers, Deep Research + MCP, and embedding-2 — directly touch developer-tooling workflows.

TL;DR

• Apr 1 — Flex and Priority inference tiers introduced; pick per request to trade off cost vs latency.
• Apr 15 — gemini-3.1-flash-tts-preview released: cost-efficient, expressive, steerable TTS.
• Apr 21 — Deep Research agent gains collaborative planning, visualization, MCP server integration, and File Search; new model variants deep-research-preview-04-2026 (speed/streaming) and deep-research-max-preview-04-2026 (max research).
• Apr 22 — gemini-embedding-2 reaches GA.
• Apr 30 — gemini-robotics-er-1.5-preview shut down; migrate to gemini-robotics-er-1.6-preview.

Quick Answer

If you ship a Gemini-backed app, three of these are immediately actionable. (1) Flex / Priority lets you cost-tune per request without code surgery — useful if you have mixed batch and interactive traffic on the same key. (2) Deep Research + MCP integration means an agent can pull from your MCP-described tools natively inside the Gemini Deep Research workflow — a real lift for builders already on the MCP ecosystem. (3) gemini-embedding-2 GA means you can move embedding workloads off embedding-1 without preview-status caveats.

Flex and Priority Inference Tiers (Apr 1, 2026)

Google introduced Flex and Priority inference tiers across Gemini API on April 1. The model is the same; what changes is the serving SLA and the price point.

Per-tier pricing and exact latency targets are surfaced in the official changelog. The shape of this is similar to what AWS Bedrock has been doing with X-Amzn-Bedrock-Service-Tier — a sign the whole industry is converging on per-request tiering rather than only per-model price.

gemini-3.1-flash-tts-preview (Apr 15, 2026)

A new TTS preview rooted in the Gemini 3.1 Flash family. Google's framing: cost-efficient, expressive, and steerable. As a preview model:

• Expect the API surface to evolve before GA.
• Good fit for prototyping voice features, narrating agent output, or building accessibility flows on top of Gemini-driven coding tools.
• Avoid for compliance-bound production until it ships GA — preview models can change input/output shape without notice.

Deep Research Agent: Collaborative Planning, Visualization, MCP, File Search (Apr 21, 2026)

The headline coding-agent change of the month. The Deep Research agent picked up four capabilities:

1. Collaborative planning — multi-step plans that incorporate user feedback mid-flight rather than running to completion blind.
2. Visualization support — research outputs include rendered visualizations, not just text.
3. MCP server integration — Deep Research can call tools exposed via Model Context Protocol servers, the same standard surfaced in Claude Code and Cursor.
4. File Search — first-party document grounding inside the agent loop.

Two new model variants ship alongside:

• deep-research-preview-04-2026 — optimized for speed and client-side streaming. Use for interactive UX where you want the research to feel live.
• deep-research-max-preview-04-2026 — optimized for comprehensive automated research. Use when wall-clock time matters less than depth.

The MCP move is the structurally important one. Three of the major coding-agent stacks — Claude Code, Cursor, and now Gemini Deep Research — speak MCP, which means an MCP server you build (for your internal docs, your CI, your ticketing) becomes a portable plugin across vendors. We covered the MCP surface in detail in the complete guide to MCP servers.

gemini-embedding-2 Reaches GA (Apr 22, 2026)

gemini-embedding-2 graduated from preview to general availability on April 22, 2026. For developer workflows that depend on embeddings — RAG over codebases, semantic code search, doc-grounded chat — this lifts the production caveat that comes with preview models. If you've been holding embedding-1 traffic in production while testing -2 in staging, you can now flip the switch.

gemini-robotics-er-1.5-preview Shutdown (Apr 30, 2026)

Less relevant for AI coding tools, but worth noting for completeness: gemini-robotics-er-1.5-preview was retired on April 30, 2026. Migrate to gemini-robotics-er-1.6-preview, which is the supported successor.

How This Changes the Vendor Picture

For frontier-coding model context, see our OpenAI Codex GPT-5.5 update and Claude Opus 4.7 release post.

Sources

• Google AI for Developers — Gemini API changelog: ai.google.dev/gemini-api/docs/changelog
• Gemini 3.1 Flash coding context: /blog/gemini-3-flash-coding-guide
• MCP background: /blog/complete-guide-mcp-servers

For broader landscape context see our gpt-5-1-codex-max-vs-gemini-3-pro">Opus 4.5 vs GPT-5.1 Codex Max vs Gemini 3 Pro comparison and the competitive landscape of AI software development.