Google replaces Antigravity Agent 05-2026: what can break before the planned October 5 shutdown

What Google announced and which date needs attention

Gemini API documentation records the release of `antigravity-preview-09-2026` on September 17, 2026, and presents it as the replacement for `antigravity-preview-05-2026`, which is marked for retirement. The change matters especially to teams building agents with tool execution, rather than only to those requesting a text response from the model.

Google’s retirement table places October 5, 2026 as the earliest shutdown date for `antigravity-preview-05-2026` and recommends migrating to `antigravity-preview-09-2026`. That wording matters: the table does not necessarily confirm the specific time at which the version will cease to be available for every account or environment. Google says it will communicate the exact date in advance.

The operational reading, therefore, should not be that availability is guaranteed through the end of that day. Instead, teams should complete testing before that reference date. It is also advisable to check the current retirement documentation immediately before deployment, in case an update, postponement, or availability clarification has been issued that was not included in the information reviewed.

The replacement does not, by itself, demonstrate an improvement in quality, security, cost, latency, or autonomy. The sources describe a version change and tool-interface changes; they do not provide an experimental comparison between the two versions on those dimensions. Keeping that distinction prevents a compatibility migration from being presented as an unverified performance claim.

Two migration paths depending on how the response is consumed

Google explicitly identifies one straightforward transition case: a workflow running in a remote sandbox that consumes only `output_text` or `model_output`. In that scenario, the documentation says changing the agent identifier string may be sufficient. This is a narrow condition: it assumes the application does not itself interpret or execute tool calls that occur during the interaction.

The second case contains the integrations at greater risk of incompatibility: agents working in a local environment, applications that receive or process `function_call`, systems that reconstruct an action trace, or platforms that validate arguments before authorizing an operation. For those systems, the version identifier is only one part of the migration. The consumer must accept the new version’s tool contract.

There are intermediate situations as well. A service may use a remote environment while still retaining events for auditing, producing metrics by tool name, applying permission policies, or testing with simulated responses. If any of those components depend on prior names or arguments, the system should be treated as trace-sensitive even if the final product primarily displays text.

Before classifying a case as simple, the team should identify where `output_text`, `model_output`, function calls, and tool results are read. That review should include SDK adapters, event queues, logs, automated evaluators, and internal dashboards. The absence of errors in the conversational interface does not prove that supporting services have no dependencies.

Which contract changes can break an integration

The release note describes changes to filesystem tools. These include naming changes and a PascalCase convention for documented arguments. For a consumer that compares literals, deserializes against a strict schema, or derives permissions from field names, the change can produce rejections even when the request to the agent remains valid.

File editing is a specific breaking point. The new-version documentation describes replacements through line ranges, whereas the earlier version used full rewrites. A local executor expecting to receive the complete replacement content may not know how to apply a bounded edit. Conversely, converting a range edit into a full rewrite without safeguards can overwrite concurrent changes or alter line endings.

The update also adds file-search tools. Their presence does not require every application to use them, but it can affect tool allowlists, authorization mechanisms, and mocks that account only for operations available in the earlier version. A system that denies an unknown tool by default can stop a task that previously completed through a different sequence of actions.

Exact names and capitalization should not be inferred from older examples or normalized without a test. The current technical guide presents filesystem tools as function calls. In a local integration, the contract must be reviewed end to end: received event, validation, authorization, execution, result serialization, and interaction resumption.

Tests for finding failures that a textual answer will not reveal

The most useful test is not asking the agent whether it can perform a task, but running representative tasks and reviewing every step. A minimum suite should include file creation, content reading, directory listing, localized editing, and search. If the product permits modifications, add cases for insufficient permissions, invalid paths, missing files, and editing conflicts. Expected results should cover both user-facing output and events, as well as the final state of the environment.

Capture traces from an equivalent sample using both versions when the test environment permits it. The point is not to require an identical sequence: the new tools can cause that sequence to change. The objective is to verify that the adapter can interpret and execute every authorized call, that results return to the agent in the expected format, and that the task reaches a correct state.

Mocks deserve a separate review. They often encode the previous contract implicitly: function names, field casing, argument structure, or full file contents. A mock that still accepts only the old contract can hide defects; an overly permissive mock can approve calls that the real executor will reject. It is advisable to derive simulations from captured traces and retain negative cases.

Instrumentation should record the requested version, environment type, received calls, authorization decision, and execution result, while avoiding unnecessary storage of sensitive content. This information makes it possible to distinguish a contract issue from a permission denial, an environment failure, or an unexpected model response.

Cost, availability, and the limits of what can be concluded

Gemini Developer API pricing documentation states that Antigravity Agent bills inference, including intermediate tokens from agentic loops, under standard Gemini rates. It also states that environment compute is not billed during preview. This does not make it possible to calculate the cost of a particular workload: that will depend on models, token volume, duration, and the effective behavior of tasks.

This information should not be extrapolated to access channels that the supplied sources do not describe as equivalent. The information reviewed refers to Gemini API and does not demonstrate that availability, quotas, terms of use, or the schedule are identical in Vertex AI or other paths. Teams with a multichannel abstraction layer should confirm the actual provider for every deployment before applying this news item as a general rule.

The documentation analyzed does support the conclusion that there is a simple-change path for a very specific remote case and that there are meaningful interface modifications for tool consumers. It does not support the conclusion that all remote agents are unaffected, that a local migration is mechanical, or that the two versions provide functionally equivalent results for every task.

As a service measure, the priority is to treat this replacement as a contract migration. Changing the version name may be enough when only text output is consumed in the remote scenario described by Google. For any integration that observes, validates, or executes tools, the decision should be based on traces and regression tests, not on the apparent continuity of the conversation.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.