Google Antigravity May Preview Retirement: Gemini API Migration Checklist

Google lists October 5, 2026 as the shutdown date for antigravity-preview-05-2026 and directs Gemini API builders to antigravity-preview-09-2026. This affects workflows using the May managed agent, not the whole Antigravity product. Google’s retirement notice names the replacement explicitly.

The replacement and migration instructions were announced on September 17; October 5 is the scheduled retirement, not a new launch. Check the agent ID, file tools and task budgets before promoting the replacement.

The live cutoff remains unverified. Google’s lifecycle page describes shutdown and qualifies its table as earliest possible retirement dates, while its gemini-skills migration reference says old-ID requests redirect after October 5. Neither a confirmed outage nor guaranteed automatic migration follows from those conflicting statements. No cutoff hour or timezone is specified.

1. Find the dependency, including custom agents

Start with a filename-only search of your repository; review matches before editing:

rg -l --hidden -F -g '!.git/**' -g '!node_modules/**' 'antigravity-preview-05-2026' .

Also inspect deployed configuration, CI variables, gateways, saved request templates and queued work. Separate active dependencies from historical documentation and fixtures; a blanket replacement can erase useful migration evidence.

A clean application search is not enough. Custom managed agents have a base_agent behind their own invocation ID. Inspect that saved definition and its deployment template, then update affected references to the September preview through your existing configuration process.

Record the underlying model separately. Google’s agent overview currently identifies Gemini 3.8 Flash as the foundation, with model configuration through agent_config. The managed agent remains a preview with free-tier quota and paid usage. Confirm access in the project you actually deploy.

For historical context, the Gemini 3.7 Flash guide explains why an agent ID and its underlying model are different dependencies. Its May-preview examples describe the earlier configuration, not the migration target today.

2. Check whether your integration consumes tool events

Google’s September migration note separates two cases:

• Remote, output-only clients: if you read only output_text or model_output, Google says changing the agent string is sufficient.

• Local executors or event consumers: if you execute tools locally or parse function_call, adapt the built-in tool interface too.

Practical implication: include audit-log parsers, permission rules and UI event renderers in this review. Google executing the tool remotely does not make your integration output-only if another service still interprets its calls.

The documented mapping is below. File editing now uses line-range replacement, and the changed tool arguments use PascalCase. Source: Google’s tool migration table.

Shell execution and web search are unchanged in that table. Apply the casing changes to these tool adapters—not indiscriminately to SDK request fields or custom tools. Check the applicable schema for line and offset semantics rather than guessing inclusive or exclusive boundaries.

3. Bound the whole task, including continuations

Use agent_config.max_total_tokens, not max_output_tokens, which the agent rejects. Google’s budget documentation says the limit counts input, output and thinking tokens but excludes cached tokens. Enforcement can overshoot slightly, and reaching the budget returns incomplete.

Illustrative request fields—not a complete request or a tested recommendation:

{
  "agent": "antigravity-preview-09-2026",
  "agent_config": {
    "type": "antigravity",
    "max_total_tokens": 50000
  }
}

The 50,000-token value is a trial budget to tune against your task, not a universal default. Merge these fields with your existing input, environment and permissions.

Google documents continuation using previous_interaction_id and the saved environment ID. Each continuation gets another interaction budget. That does not establish compatibility for resuming a May-preview session on September.

Implementation recommendation: track cumulative usage and continuation count under one application task ID. Stop automatic retries when the task allowance is exhausted, and handle incomplete explicitly. An interaction token cap is neither a task-wide cap nor an exact dollar ceiling.

4. Reprice the complete agent loop

Google’s pricing page bills intermediate reasoning and input across the loop; sandbox compute is unbilled during preview. For Gemini 3.8 Flash Standard paid usage, USD rates per million tokens through December 31, 2026 are $0.75 uncached input, $0.075 cached input and $3.75 output including thinking. The listed January 1, 2027 rates double each figure.

Illustrative calculation, not a benchmark: assume 1 million aggregate input tokens, 60% billed as cached input, plus 50,000 output/thinking tokens.

0.4 × $0.75 + 0.6 × $0.075 + 0.05 × $3.75 = $0.5325

At the scheduled January rates, identical usage costs $1.065. Both figures exclude tools, cache storage, taxes and retries and assume the stated cached-input rate applies. Recalculate for your configured model and billing tier. These numbers do not measure May-versus-September savings. Rate-card basis.

5. Make artifacts and control behavior the acceptance gate

The following are proposed migration checks, not tests performed by RohitAI. Run them against disposable inputs and retain the resulting files and event traces:

• File correctness: create a numbered multiline file, replace its middle section and assert exact output bytes with surrounding lines unchanged. Read a selected region, list the directory, locate the filename and search for a known sentinel. Include Unicode, empty files and names with spaces where relevant.

• Event and policy compatibility: replay synthetic allowed and denied operations through your adapters, approval rules, logs and UI. Require correct decisions and intelligible records, not merely a plausible final answer.

• Lifecycle handling: retrieve a background run through a terminal status; cancel a disposable run and verify it reaches cancelled. Google documents a possible cleanup delay and distinguishes cancellation from deleting the stored record. Background execution reference.

• Continuation and retries: interrupt a trial through its token budget, then verify an intentional continuation uses the expected environment and preserves artifacts. Exercise pre-existing sessions separately. Prevent retries from repeating external side effects.

Add a long-session fixture if the workflow relies on earlier instructions surviving compaction. The September model page lists a 1,048,576-token input window but compaction around 135,000 tokens. Do not equate nominal capacity with an uncompressed conversation. Google also describes caching and file-tool improvements; this guide has no measured speed, cost or quality comparison.

For each acceptance run, record the agent or base agent, configured and resolved model where available, SDK/API revision, environment, permissions, token categories, tool usage, final status and artifact checks. If May is unavailable, compare against saved fixtures rather than treating the retired ID as a rollback route.

Suggested release gate: promote a workflow only after required artifacts, event consumers, authorization decisions and task-level limits pass. If a critical check fails, pause that workflow or use an independently validated alternative or manual path.

Methodology: This AI-assisted guide uses Google’s published documentation, rechecked on October 5, 2026. The checklist and cost example are analysis, not hands-on results. No authenticated endpoint test established shutdown behavior, automatic redirection or cross-version session compatibility.