Gemini 2.5 Pro was a documented stable model with a one-million-class input context; Google’s Gemini 3.1 Pro launch on 19 February 2026 was a preview. That is more than a naming difference. A preview can have tighter usage terms or change ahead of a stable release, whereas the exact 2.5 Pro API ID is a specific previous-generation selection. Neither label tells you that a public availability claim applies to every region and billing tier.
What the model cards actually say
Google’s 3.1 Pro Preview reference specifies gemini-3.1-pro-preview, text/image/video/audio/PDF input and text output, with an input limit of 1,048,576 and output limit of 65,536 tokens. The 2.5 Pro card lists the stable ID gemini-2.5-pro and the same stated input/output token maxima. These are documented interface limits, not proof that every file format or long-context inference performs accurately.
Reasoning and tools
The 3.1 Preview reference documents thinking, function calling, structured outputs and caching. File search, grounding and code-execution availability can depend on the product surface. Test the exact combination of model, interface and tool rather than assuming an AI Studio feature is available identically from every Vertex AI deployment. A preview tag should trigger heightened regression monitoring.
What a fair evaluation needs
Ask each model to answer the same set of difficult document-grounded questions. Compare source attribution, omission rates, unsupported inference, output token consumption, timeout rate and schema validation failures. Long context must be tested with evidence distributed across many positions. To isolate model differences, use the same segmentation and image/PDF extraction pipeline.
Upgrade advice
For a new integration, confirm the 2026 service status from the cloud console. Do not blindly replace a stable identifier with a preview. Make rollout conditional on correctness gains for your actual cases and prepare a fallback when preview capacity is limited.
Verified capability snapshot
Specifications below describe the documented model reference as checked at the cutoff; they are not proof of your own regional or tier entitlement. For a comparison article, the values are associated with the preview model identified in the heading, not every model discussed.
| Field | Documented value | Official evidence |
|---|---|---|
| context window tokens | 1048576 | official model card |
| max output tokens | 65536 | official model card |
| modality inputs | [“text”, “image”, “audio”, “video”, “pdf”] | official model card |
| modality outputs | [“text”] | official model card |
What to verify before adopting this model or method
Use the exact release ID and API endpoint; confirm your geographic processing location, entitlement tier, model lifecycle, feature flags, context and output limits from the provider console. Run a reproducible evaluation on your own data and retain human oversight for consequential actions. Where a specification, price, commercial license or entitlement is not verified for your deployment, this article deliberately leaves it unasserted instead of filling a plausible number.
Sources & Further Reading
- blog.google — documentation/reference (supplier-provided reference; not independently checked in this pass)
- ai.google.dev — documentation/reference (supplier-provided reference; not independently checked in this pass)
- ai.google.dev — documentation/reference (supplier-provided reference; not independently checked in this pass)
Source documentation establishes vendor-published facts, not independently measured model quality, tenant access, contractual rights or current prices. Use a dated pricing ledger and an account-specific entitlement check for deployment decisions.