Gemini 2.5 Flash

Gemini 2.5 Flash, Google's own language model, turns up mostly as the cheap working model inside someone else's architecture. Dimitri Mazmanov, Principal Product Manager at Spotify, published on 3 September 2026 an account of two public AiKA Modes running on it: bulk-reader, which reads a batch of files and answers in structured bullets, and code-writer, which generates code from a specification plus a reference file whose conventions it must reproduce. Routed through a Claude Code plugin called shunt, the arrangement averaged roughly 90% token savings across four scenarios tested against a Java monorepo. The model is not load-bearing in that design: any model configured in the Portal instance can take its place.

The same substitutability appears in Google's AP2 repository, where the sample scenarios use the Agent Development Kit and Gemini 2.5 Flash while the documentation states that the Agent Payments Protocol itself does not depend on either.

Where it is measured directly, the results split. In Microsoft's Magentic Marketplace, Gemini 2.5 Flash and GPT-5 were the exceptions to the pattern of agents accepting the first "good enough" offer after talking to few vendors. Mistral AI reports Voxtral Small surpassing it on speech tasks. Wharton's persona-prompting study found one of its few non-null results here: a "Young Child" prompt lifted GPQA Diamond accuracy (RD = 0.098), while Gemini Flash models sometimes refused to answer at all under an out-of-domain expert persona. Mazmanov hit a harder limit: the model missed a thread-safety bug that Claude Code caught immediately.