Computer use in Gemini 3.5 Flash

The reaction was not excitement about a breakthrough. It was frustration that Google still seems behind on the surrounding product. People kept coming back to missing or weak pieces in the Gemini ecosystem: no MCP support in the main app, confusion around Gemini CLI versus Antigravity, weak repo-level coding workflows compared with Codex and Claude Code, and an app experience that several people called throttled, forgetful, or just bad. That shaped how the announcement landed. Many saw it less as a new platform and more as Google checking a box its competitors already checked. On the substance, the consensus was that computer use is a brute-force interface, not an elegant one. It is slow, costly, and easy to derail. But plenty of people still think it will be useful because the world runs on brittle UIs, proprietary internal tools, SSO-gated portals, and software with no usable API. In that setting, screenshot-and-click automation is ugly but real. The sharper framing was that computer use is a fallback for the long tail. If you control the stack, you should expose tools, APIs, accessibility hooks, or direct shell access instead. If you do not control the stack, computer use may be the only thing that works this week. People were also skeptical of Google’s benchmark presentation. The cited chart showed Gemini close to leading models on an OSWorld-style workload, but not actually ahead of the top Claude and OpenAI entries. Several readers said the more relevant story is price and latency. Flash may make sense as a front-line model where failures are tolerable and retries are cheap. That fit a broader pattern in the comments. Even users unimpressed by Gemini’s quality said they reach for Flash because it is fast and inexpensive. Reliability remained the biggest drag. Multiple firsthand reports described Gemini as hit-or-miss, quick to regress, prone to giving up, or oddly over-guardrailed. Others said they see the opposite and get solid results, especially through the API rather than the consumer app. The practical conclusion was not that Gemini is unusable. It was that Google still has a consistency problem across interfaces, regions, and task types. For computer use, that is a serious weakness because UI automation amplifies every small mistake into wasted time or broken state.