Philipp Schmid announced antigravity-preview-09-2026, a new version of the Gemini managed agent runtime. It runs on Gemini 3.8 Flash by default, provides a persistent Linux sandbox with bash execution, file tools, Google Search, and URL context, and compacts context automatically at about 135k tokens.
The update shifts managed agent runtimes toward persistent environments where files, secrets, and reusable configurations live outside the prompt. Higher cache hit rates cut repeated token spend across turns, while the Files and Credentials APIs move data and secrets through the platform instead of through model context. A new Files API supports uploading data with files.upload(), listing files with files.list(), and downloading artifacts. A Credentials API registers secrets with credentials.create(), supporting bearer_token, environment_variable, or oauth2, without the model ever seeing them.
This matters now because long-running coding and research agents accumulate state, and each turn that re-sends context or credentials adds cost and exposure. Schmid reports a 30% cost reduction on multi-turn coding, 17% on reasoning, 40% fewer tokens on file edits, and up to 8% higher task completion on SWE and research evals. The post does not publish the evaluation setup behind the cost and completion figures, so the gains cannot yet be checked against independent workloads.
Teams running agents on this runtime can keep credentials for MCP servers out of the model's view instead of injecting them into prompts. Published documentation of the SWE and research evals, and third-party cost measurements on multi-turn workloads, would confirm whether the stated savings hold.