Google updates Gemini Interactions API with steps schema and Api-Revision 2026-05-26

• Google is replacing the Gemini Interactions API's older outputs plus roles structure with a steps schema, where user_input, thought, function_call, search calls, and model_output each become first-class step types, according to GoogleAIStudio's API update and Phil Schmid's migration summary.
• The migration is a real breaking change, not a cosmetic rename, as Phil Schmid says SDK users need Python or JS 2.0.0+, while older clients can opt in early with the Api-Revision: 2026-05-26 header.
• Google framed the redesign around multi-step agent workflows, and GoogleAIStudio's thread paired it with Gemini 3.1 Flash-Lite, which it described as its most cost-efficient model for high-volume agentic tasks.
• The docs rollout already includes a dedicated breaking changes guide and a gemini-interactions-api skill, both linked by GoogleAIStudio and Phil Schmid.

Google's own slide literally labels the shift GoogleAIStudio's slide "From Turns to Steps." The migration guide is already live at ai.google.dev, and Phil Schmid's thread adds a companion skill that can rewrite prompts and code for the new shape. A separate Wes Roth post suggests the API work is landing alongside a broader Gemini push into scheduled tasks, reusable skills, and an app-level Agent Mode.

Google's example replaces conversational turn structure with an explicit sequence of typed actions. The point is less chat history, more execution trace.

The new steps array shown by GoogleAIStudio's slide and restated in Phil Schmid's summary includes:

• user_input
• thought
• function_call
• function_call_result
• google_search_call
• google_search_result
• model_output

That makes hidden reasoning, tool calls, tool results, and final output look like sibling records in one interaction log, instead of being squeezed into user and model roles.

The migration knobs are unusually explicit for a beta API change.

• Upgrade to Python SDK >= 2.0.0
• Upgrade to JS SDK >= 2.0.0
• Or send Api-Revision: 2026-05-26 to opt into the new behavior before the default flips
• Use the docs toggle between Interactions API and generateContent
• Use the /gemini-interactions-api skill from the gemini-skills repo to automate migration work

Ivan Leo called it a breaking change outright, which is the useful framing here. Any tool or wrapper that assumed outputs plus role-based turns now has a schema boundary to cross before GA.

Google bundled the API update into the Gemini 3.1 Flash-Lite announcement. In that same thread, GoogleAIStudio called Flash-Lite its most cost-efficient model for high-volume agentic tasks, translation, and simple data processing.

That pairing helps explain why the schema changed now. If Google wants developers to run more multi-step agent loops cheaply, representing thinking and tool traffic as discrete steps is cleaner than pretending the whole workflow is a back-and-forth chat.

A separate set of Gemini app screenshots from Wes Roth shows where this is probably heading on the product side: an "Agent" view with task history, reusable "Skills," and "Schedules" that can run on repeat, respond to events, or monitor in real time.

Those screenshots are not the API announcement, but they line up with it. The app UI includes recommended skills like email triage and meeting prep, plus schedule templates like Inbox Zero Assistant and News Digest, which is a much more agent-shaped product surface than a normal chatbot tab.