AI News
All stories
Official

OpenClaw 2026.11.1 makes Interactions the default for Gemini Flash text models

The OpenClaw release notes say Gemini Flash text models now use Interactions by default on Google's official API-key endpoint, while custom proxies keep their existing default.

Products · · 3 min read

OpenClaw 2026.11.1, published November 5, changes the request path Gemini Flash text models take by default. When the google provider uses Google's official AI Studio endpoint with an API key, those models now go through the Interactions API instead of GenerateContent. The release pairs this with a second change: provider recovery now follows the retry timing a provider specifies, rather than the client's own backoff. The two changes came from pull requests #189395 and #180280, merged on November 1 and October 22.

The Interactions change is a routing default, not a new transport. The author of #189395, obviyus, wrote that OpenClaw already shipped the Interactions transport, but Flash models on the API-key provider still used GenerateContent. Discovery also dropped any Flash model that did not advertise generateContent. The stated reason is that Interactions is Google's newer Gemini API and Flash is the family most people run on API keys. The default covers google/gemini-*-flash* text models, including Flash-Lite, gemini-flash-latest and future Flash IDs, but only when the effective base URL is the official endpoint. Pro models, including the default google/gemini-4-argon, keep GenerateContent, as do custom baseUrl proxies that have no explicit api setting. An explicit api value always wins: google-generative-ai forces the old path, and a model-level google-interactions selects Interactions for any Gemini text model.

Two behavior notes matter for anyone with existing setups. Interactions uses implicit caching, so a configured cachedContent handle is ignored on that route instead of failing the turn. Native video input keeps working on Flash, now sent as an Interactions video input after payload hooks run.

The pull request's evidence comes from live calls to the Gemini API using an AI Studio key, with the branch built locally and the endpoint read from debug logs. A fake proxy that only supports GenerateContent recorded request paths for the proxy cases. The default official-endpoint call on gemini-3.5-flash-lite reached /v1beta/interactions, and a configured cachedContent handle that previously threw an error now returned a reply. gemini-3.8-flash returned a 503 high demand error on both APIs during testing, so the live proof used gemini-3.5-flash-lite. The pull request lists what was not tested live: native video through a channel attachment, music and video SDK calls, because those models are paid, and the proxy default without an api setting. That last case was covered only by unit tests, because a loopback proxy is classified as a custom non-Google endpoint.

Review history points to the compatibility concern. An automated ClawSweeper review labeled the pull request merge-risk: compatibility, and earlier review cycles blocked merging over two issues: custom Google endpoints staying on a supported API, and Gemini key rotation on the new route. The final review said the pull request was ready for maintainer review. The author's upgrade report, recorded in the same thread, says conversation history survived an upgrade from openclaw 2026.10.9 and that an explicit GenerateContent preference was kept. The release notes do not say how users who never set an api value will experience the switch in practice.

The rate-limit change, from #180280 by fuller-stack-dev, replaces a local counter. Before it, each rate-limit failure doubled a local backoff, from 30 seconds up to a 24-hour cap. The pull request says that under sustained tokens-per-minute pressure, OpenAI 429 responses that each said try again in 6s compounded into day-long lockouts. The cooldown now follows the timing the provider reports, from Retry-After headers, retry text in the error, or Google's RetryInfo field. An exhausted-bucket reset header is used only when the provider gives no explicit time. Previously OpenClaw took the larger of Retry-After and the full bucket refill, so a 6-second instruction cooled a profile for the whole refill. If a provider gives no wait time, the turn still fails over to another profile or model, and the profile remains usable for the next turn. Profiles locked by the old backoff are released on upgrade. Billing, permanent-auth, and the 30s, 1m and 5m overload and transient backoffs are unchanged.

What changed

Gemini Flash text models on Google's official API-key endpoint now go through the Interactions API by default instead of GenerateContent, while custom baseUrl proxies keep their existing default.

Type: Breaking change · Applies to: OpenClaw 2026.11.1, google/gemini-*-flash*, gemini-flash-latest, gemini-3.5-flash-lite

Why it matters

Teams that call Gemini Flash through OpenClaw may see different request behavior after upgrading, while setups behind custom proxies are unaffected by this default change.

What you can do

After upgrading to 2026.11.1, check OpenClaw configurations that call Gemini Flash models and test them against the new default.

Original source

Where we saw it

1 source

OpenClaw releases

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works