Gemini 3 Pro Preview Reaches End of Life: Migration to Gemini 3.1 Pro Required

Gemini 3 Pro Preview Shutdown: What Changed on March 9, 2026

On March 9, 2026, Google officially decommissioned the gemini-3-pro-preview model on the Gemini API and Google AI Studio. The endpoint no longer runs the original model — any calls made with the gemini-3-pro-preview identifier are automatically redirected to gemini-3.1-pro-preview. For applications that had not yet migrated, this represented a significant silent model swap for some and a hard API failure for others, depending on how the model identifier was referenced.

What Actually Happened

The retirement unfolded in two stages. On March 6, 2026, Google moved the gemini-pro-latest alias from Gemini 3 Pro Preview to Gemini 3.1 Pro Preview. Then on March 9, the original gemini-3-pro-preview endpoint was fully shut down. Developers who had hardcoded the model identifier and not yet updated their integration saw API call failures beginning on that date — the shutdown was not graceful for those applications.

Vertex AI users operate on a different timeline. Google scheduled the Vertex AI deprecation of gemini-3-pro-preview for March 26, 2026, giving enterprise customers an additional two weeks beyond the standard Gemini API cutoff. Notably, some Vertex AI users received email notices referencing the later date, which caused confusion given the earlier AI Studio and API deadline.

Why Google Moved Fast

The rationale offered by Google engineering leads is straightforward: Gemini 3.1 Pro represents a substantial leap in capability that makes maintaining the older Preview infrastructure difficult to justify. On the ARC-AGI-2 reasoning benchmark, Gemini 3.1 Pro scored 77.1% — more than double the score achieved by Gemini 3 Pro. Google also framed the shutdown as a resource-management decision, consolidating GPU/TPU capacity onto newer models rather than running parallel inference stacks.

Preview models, by Google's own definition, are not guaranteed long-term stability. The Preview designation signals a model at the validation and feedback stage — not a production-stable release. That said, the February 26 announcement of the March 9 shutdown gave developers approximately 11 days to migrate, which some in the developer community flagged as shorter than the 14-day minimum typically associated with Google's preview model deprecation norms.

Developer Migration Path

For most integrations, migration is a single-line change:

# Before
model = "gemini-3-pro-preview"

# After
model = "gemini-3.1-pro-preview"

The model IDs are API-compatible — the request and response schema remains the same. The primary differences developers should account for are the significantly improved reasoning and agentic capabilities, and the practical reality that Gemini 3.1 Pro Preview was experiencing elevated latency at launch, with some community trackers recording response times exceeding 100 seconds during high-demand periods.

Google also introduced a separate gemini-3.1-pro-preview-customtools endpoint for applications that mix Google-native built-in tools (such as Bash execution or code interpreter) with custom function-calling schemas. Developers building multi-tool pipelines that had previously relied on the standard Pro endpoint should evaluate whether this variant better fits their architecture.

Implications for Gemini CLI Users

Gemini CLI users who had explicitly configured gemini-3-pro-preview via the -m flag or in their settings files needed to update their configuration. Users running with default settings were not directly affected — Gemini CLI's application-level defaults had already been set to the stable Gemini 3 model tier in earlier releases. The impacted group was specifically those who had pinned the old Preview model identifier for access to the Preview-tier capabilities.

Going forward, gemini-3.1-pro-preview is the recommended Pro-class model for developers building on the Gemini API and CLI. It carries the same pricing as its predecessor while delivering substantially improved performance across reasoning, agentic task completion, and coding benchmarks.