Google's Gemini API changelog for September shows a platform pivoting hard toward agentic workloads and real-time multimodal interaction. The month brought four generally available model releases, a managed agent update that rewrote its file operation interface, and a restriction on older 2.5 models that signals where Google wants developers to focus next.
The 2.5 Models Are Not Dead, But Google Wants You Off Them
On September 18, Google announced it is limiting access to the Gemini 2.5 models to users who have actively used them in the past. The models are not deprecated and will continue to be served, but new projects should use Gemini 3.5 Flash-Lite or 3.8 Flash instead. The practical effect is that teams still running on 2.5 need to plan a migration. Google is reserving capacity for current workflows while steering new development toward the 3.x line.
This is a capacity management move, not a technical sunset. The 2.5 models still work for existing integrations, but Google is making it clear that the investment in new features, performance improvements, and agent capabilities is going into the newer models. Teams that have not yet migrated will eventually hit a wall as the older models receive fewer updates and less optimization.
Gemini 3.8 Flash and the Agent-First Model
Gemini 3.8 Flash, generally available since September 2, is described as Google's most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows. The framing matters: Google is not positioning this as a general-purpose chatbot model. It is explicitly targeting the kind of sustained, multi-step work that agents perform.
Alongside it, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking reached general availability on September 15. These are audio-to-audio models for real-time voice applications. The standard version supports interleaved reasoning and asynchronous function calling for low-latency voice agents. The Extended Thinking variant adds background reasoning during live audio interactions, recommended when higher reasoning depth is needed mid-conversation.
The combination of a Flash model optimized for agent workflows and a Live model optimized for real-time voice suggests Google is building the infrastructure for voice-controlled agents that can execute multi-step tasks. The Live API integration means developers can build voice interfaces that call functions, manage files, and run code without leaving the audio session.
Antigravity Agent Rewrites Its Tool Interface
The September 17 update to the Antigravity Agent is the most developer-facing change in the month's changelog. The new version, antigravity-preview-09-2026, replaces the May release and changes how the agent interacts with the filesystem. Parameters switched from snake_case to PascalCase. File edits now use line-range replacements instead of full rewrites. File reading changed from byte offsets to line ranges. Two entirely new tools were added: find_by_name for file and code search, and grep_search for searching file contents.
For teams running the agent locally with local_environment, this is a breaking change. The built-in tools have different signatures, and any code that parses function_call steps needs to be updated. For teams running on a remote sandbox reading only output_text or model_output steps, the migration is just updating the agent string. The May version shuts down on October 5.
The tool changes are significant because they reflect how the agent's capabilities are evolving. Adding file search and code search as first-class tools means the agent no longer needs to shell out for these operations. Line-range edits instead of full file rewrites reduce token consumption and make partial modifications more precise. These are the kinds of changes that matter when agents are operating on large codebases where efficiency directly affects cost.
Agentic Video Understanding Cuts Token Usage by 88%
Released September 1, agentic video understanding for Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite takes a different approach to processing long-form video. Instead of feeding the entire video to the model as a static input, the model dynamically navigates the video timeline, requesting transcripts, frames, or audio tracks on demand. Google reports up to 88% fewer tokens for long-form content compared to static processing.
This is a practical cost reduction for any application that processes video at scale. The agentic approach means the model only pulls the data it needs rather than consuming the full video token budget upfront. For developers building video analysis pipelines, content moderation systems, or media indexing tools, the token savings translate directly to lower inference costs.
The Broader Pattern: Agents Everywhere
The September changelog sits within a longer arc that started in May with the launch of Managed Agents and the Antigravity Agent in public preview. Managed Agents let developers build autonomous, stateful agents that run in isolated Google-hosted Linux sandboxes. The Antigravity Agent can plan, reason, write and execute code, manage files, and browse the web inside its container.
Computer Use launched in public preview in June for Gemini 3.5 Flash, adding browser, mobile, and desktop environment support with configurable safety policies. The Gemini Omni Flash model, generally available in August, added video extension, interpolation between frames, and resolution control up to 4K. Gemini 3.5 Transcribe brought dedicated speech-to-text with speaker diarization and custom vocabulary biasing across 85-plus languages.
The pattern is clear: Google is building a full agent stack. Models optimized for sustained agent work, managed sandboxes for running agents, file and code search as first-class tools, agentic video processing that reduces cost, voice interfaces for real-time interaction, and computer use for interacting with external systems. Each release adds another piece, and the September updates fill in gaps that were obvious from the earlier launches.
What Developers Should Act On
Teams still on Gemini 2.5 models should start planning migration to 3.5 Flash-Lite or 3.8 Flash. The models are not being shut down, but access is tightening and the feature gap will widen. Teams using the Antigravity Agent locally need to update their function_call parsing before October 5. Teams building voice agents should evaluate the 3.8 Live models for real-time audio workflows. And anyone processing video at scale should test agentic video understanding against their current pipeline to measure the token savings.
The deprecations are also worth tracking. The gemini-3.1-flash-image-preview and gemini-3-pro-image-preview models shut down on June 25. Older Veo video models shut down on June 30. Imagen 4 and Gemini 3 Image models shut down on August 17. Google is consolidating its model surface, and the churn is real for teams that have not kept up with the migration paths.