On February 19, 2026, Google released Gemini 3.1 Pro, which it described as a step forward in core reasoning built on the Gemini 3 series and a noticeably smarter baseline for complex problem-solving. It was the first ".1" release in Gemini's history, and Google shipped it in preview across its developer and enterprise channels at once. The headline number was reasoning: Google reported 77.1% on the ARC-AGI-2 benchmark, well over double the score of Gemini 3 Pro three months earlier, at the same list price.

Key Facts#

  • Release date: February 19, 2026. API model ID gemini-3.1-pro-preview.
  • Availability: in preview on Vertex AI and Gemini Enterprise, and for developers through the Gemini API in Google AI Studio, Android Studio, Google Antigravity and the Gemini CLI. Google also brought it to its consumer Gemini subscriptions.
  • Pricing per million tokens: $2 input and $12 output for prompts up to 200,000 tokens, and $4 and $18 above that, unchanged from Gemini 3 Pro.
  • Vendor-reported results: 77.1% on ARC-AGI-2, 80.6% on SWE-bench Verified and 94.3% on GPQA Diamond, up from 31.1%, 76.2% and 91.9% for Gemini 3 Pro.
  • Context and output: about one million input tokens and up to roughly 65,000 output tokens.
  • New controls: a medium thinking level between low and high, streaming of function-call arguments, and function responses that can include images and other multimodal content.
  • Partner results: JetBrains reported up to 15% improvement over its best Gemini 3 Pro Preview runs with fewer output tokens, and Databricks reported best-in-class results on its OfficeQA benchmark.

What Happened#

Google framed Gemini 3.1 Pro as an upgrade to the model's fundamentals rather than a new product. Its developer notes list four areas of improvement: stability and grounding, with better factuality and fewer repetitive responses; coding and agentic performance; efficiency, with better token usage across reasoning tasks; and general quality in reasoning, instruction following and writing.

The ARC-AGI-2 result drew the most attention. The benchmark is designed to stop models from memorizing their way to a high score. Grok 4's 15.9% had counted as a record in July 2025, and Gemini 3 Pro reached 31.1% in November. Moving from 31.1% with Gemini 3 Pro to 77.1% in three months was an unusually large jump, though, like all launch-day numbers, it was reported by Google itself.

For developers, the practical changes were in the controls. Gemini 3 had introduced thinking levels in place of numeric thinking budgets, with low and high settings. Gemini 3.1 Pro added a medium level for a better balance of depth and latency. Thinking cannot be turned off for the model, and the default remains high. Google also documented how its OpenAI-compatible Chat Completions interface maps onto the new levels: a reasoning_effort of medium corresponds to the high thinking level.

Tool use improved in two ways. The API can now stream the arguments of a function call as they are generated, which reduces perceived latency in agent UIs, and function results can include multimodal content such as images, so a tool can hand the model a chart or a scanned document to reason about. Partner feedback emphasized efficiency: JetBrains said the model was stronger and faster while requiring fewer output tokens.

Background#

Gemini 3.1 Pro arrived three months after Gemini 3, and two weeks after Anthropic's Claude Opus 4.6. OpenAI's GPT-5.2 had been on the market since December, and GPT-5.4 followed in early March. In that environment, a point release that kept prices flat while posting large reasoning gains let Google stay competitive without waiting for a full generation change.

The release also continued a pattern of shipping in preview first. Gemini 3 Pro had stayed in preview through its lifetime, and Gemini 3.1 Pro launched the same way, meaning developers got early access but no guarantee that the model would remain unchanged.

Why It Matters for Developers#

First, if you adopted Gemini 3 Pro, the upgrade path is simple on price and more involved on behavior. Pricing is identical, but efficiency changes mean output lengths and tool-call patterns can shift. Re-run your evaluation suite rather than swapping the model ID blindly; our guide to evaluating AI applications in .NET covers how to structure one.

Second, the new medium thinking level is useful for cost control. Many workloads, such as structured extraction with some reasoning or code explanation, do not need maximum depth. Treat thinking level as a per-workload setting in configuration, and measure the latency and token impact before choosing.

Third, streaming function arguments and multimodal function responses matter for agent builders. If your tools return screenshots, diagrams or scanned documents, the model can now reason over them directly. In .NET, the Google.GenAI SDK implements IChatClient, so these models plug into Microsoft.Extensions.AI pipelines; test multi-turn function calling end to end, because Gemini 3 models validate thought signatures strictly and return an error if they are missing. For image-heavy tools, see multimodal AI in .NET.

Finally, remember the preview label. Gemini 3 Pro Preview was eventually shut down, and 3.1 Pro began life with the same status. Keep the model ID in configuration and plan for migrations measured in months, not years.

What's Next#

At Google I/O in May 2026, Google introduced Gemini 3.5 Flash and said it outperforms Gemini 3.1 Pro on key benchmarks, while Gemini 3.5 Pro was still in testing. That makes Gemini 3.1 Pro a transitional flagship: the strongest Pro-tier Gemini for several months, but quickly challenged by the next generation's faster models.

For teams on Google Cloud, the open question is when Pro-tier models will again reach general availability with long-term support. Until then, pin preview model IDs deliberately, watch Google's model lifecycle documentation, and keep a second provider behind the same abstraction as a fallback.

Sources#