Which SDKs and tools have added gemini-3.5-flash support (a running tracker)

This is a running list, not a one-time roundup — every entry below is a real, verifiable PR from a framework or tool adding gemini-3.5-flash support. The interesting part is not just "which tools support it now," it is *what each PR actually had to touch* to do it correctly — because the diffs disagree on how much work that is, which tells you something about how much your own integration might be missing.

github.com/OpenHands/software-agent-sdk#3315 — opened 2026-05-20, closed 2026-06-03 without merging. Worth listing anyway because the PR itself is instructive: adding the model ID (gemini-3.5-flash routed through litellm_proxy) was the easy part. The bulk of the diff (465 additions) went into a is_prompt_cache_too_small classifier and retry-without-caching logic, because Vertex AI enforces a minimum prompt-cache size of 4096 tokens and grounded/short-context calls under that threshold need a fallback path or they fail outright. The PR author also had to update ADDINGMODEL.md to warn future contributors not to add models to "verified" lists without maintainer sign-off — a process scar from models getting added prematurely. Even though this PR did not merge as-is, the cache-size gotcha it surfaces is real and applies to anyone hand-rolling Vertex prompt caching, not just this SDK.

github.com/langgenius/dify-official-plugins#3165 — opened and merged same-day, 2026-05-20. This one adds the gemini-3.5-flash model definition (tool calls, vision, audio, video, structured output) to Dify's Gemini plugin, but the more interesting half of the diff is a refactor of _set_service_tier: it replaces a hardcoded FLEX_INFERENCE_SUPPORTED_MODEL_NAMES list with an explicit service_tier parameter (standard, flex, priority) that still stays backward-compatible with the old flex_inference flag. That hardcoded-model-list pattern is exactly the kind of thing that quietly breaks every time a new model ships — this PR is a reasonable template if your own integration has a similar allowlist that needs a new entry every model release.

github.com/coder/mux#3334 — opened 2026-05-19, merged the next day. Mux repoints its stable "Gemini Flash" alias from the older gemini-3-flash-preview to gemini-3.5-flash, while keeping the old alias available separately rather than deleting it outright. The part worth copying: the PR adds a dedicated "Gemini Flash thinking-policy helper" shared between the model-selection policy and the Google provider options, specifically so the alias swap does not silently change thinking_level defaults out from under existing users — the exact silent-regression failure mode this whole site is about (see /thinking-level). If you maintain a "stable alias -> latest model" mapping anywhere in your own stack, decoupling the alias swap from the thinking-config defaults, the way this PR does, is the right pattern.

None of these three PRs treat "add the model ID" as the whole job. Two of the three (OpenHands, Mux) needed a real behavioral change alongside the model swap — a cache-size fallback and a thinking-policy decoupling, respectively — not just a config entry. If your own gemini-3.5-flash integration is just a string swap from gemini-3-flash-preview with nothing else touched, these three diffs are a reasonable checklist of what else might be missing: prompt-cache minimum handling if you use caching, and an explicit (not inherited) thinking_level/thinking_budget setting if you route through any kind of alias or default.

This tracker will get updated as more frameworks ship their own gemini-3.5-flash support — check back, or see /migrate for the config-level checklist these PRs are all working around.