What Is Gemini 3.1 Pro?

Gemini 3.1 Pro is Google DeepMind's third-generation Pro foundation model, released on February 19, 2026. Built on the Gemini 3 Pro architecture, it is designed for complex multi-step reasoning, agentic coding, and long-context multimodal analysis.

If you are evaluating or implementing Gemini 3.1 Pro today, here is the operational baseline:

• Can you call it in the API today? Yes. The standard API identifier is gemini-3.1-pro-preview. Google also provides a dedicated variant, gemini-3.1-pro-preview-customtools, optimized for workflows combining bash shell execution and custom function calling. There is no gemini-3.1-pro model string; calling that name returns an error.
• Is it generally available (GA)? No. Gemini 3.1 Pro remains officially in Preview in the Gemini API and Google AI Studio.
• Is it deprecated or scheduled for shutdown? No. Google's API deprecations documentation explicitly lists "No shutdown date announced" and recommends no replacement. While Google DeepMind's Pro portal currently displays a banner stating "3.5 Pro coming soon," 3.1 Pro remains Google's active, callable Pro-tier model in the API.
• What does it cost? There is no API free tier (gemini-3.1-pro-preview is pay-as-you-go only; testing is available in Google AI Studio). Standard pricing has a sharp cliff at 200,000 input tokens: prompts up to 200k tokens cost $2.00 per million input tokens and $12.00 per million output tokens; prompts exceeding 200k tokens jump to $4.00 per million input and $18.00 per million output. Thinking tokens count as output tokens and bill at full output rates.
• What are the token limits? The model supports a 1,048,576-token context window (~1M tokens) and a 65,536-token maximum output limit (64k tokens, which includes internal thinking tokens).
• What modalities work? Input supports text, code, images, audio, video streams, and PDF documents. Output is strictly text. Audio generation, image generation, and the real-time Live API are unsupported.
• How does thinking work? Thinking is built-in and cannot be turned off. It supports three levels: low, medium, and high (defaulting to high). The minimal level is unsupported.

---

API Identifiers, Custom Tools, and Platform Availability

When integrating Gemini 3.1 Pro, product naming and API reality diverge. Marketing and research papers reference "Gemini 3.1 Pro," but passing that string to SDKs or HTTP gateways will fail.

Why the Custom Tools Variant Exists

The gemini-3.1-pro-preview-customtools endpoint addresses a known failure mode in autonomous coding agents: models frequently emit conversational text or generic explanations instead of executing custom tools when orchestrating bash commands.

Google fine-tuned this variant to prioritize custom tool execution over standard text replies in multi-turn tool loops. However, Google's documentation includes an explicit caveat: this endpoint may exhibit quality fluctuations on tasks that do not utilize bash or custom tools.

If your workload is conversational analysis, document summarization, or standard structured extraction, stick to gemini-3.1-pro-preview. Both endpoints are billed at identical rates.

Where Gemini 3.1 Pro Runs

Google provides access to Gemini 3.1 Pro across several developer and enterprise surfaces:

• Gemini API: Direct programmatic access via the google-genai SDK and REST endpoints.
• Google AI Studio: Web playground for interactive prompt design, system instruction testing, and API key creation.
• Google Cloud / Vertex AI: Enterprise deployment environment with customer-managed encryption keys (CMEK), VPC Service Controls, and enterprise SLAs.
• Google Antigravity: Supported foundation model within developer environments.
• Gemini Enterprise Agent Platform & Google AI Mode: Enterprise orchestration environments.
• Gemini App: Consumer web and mobile interface (available to Google One AI subscribers).

Note on Gemini Notebook: While Google's February 2026 model card listed NotebookLM among initial distribution channels, current October 2026 Google DeepMind product documentation does not list Gemini Notebook (the rebranded NotebookLM) among active Gemini 3.1 Pro surfaces.

---

API Pricing Structure and the 200k-Token Jump

Unlike Google's Flash-tier models, Gemini 3.1 Pro has no free tier in the Gemini API. Google's API pricing documentation marks the Free Tier as "Not available," noting that free interactive testing can be conducted inside Google AI Studio. Calling the API requires an active billing account.

The 200,000-Token Pricing Cliff

Standard pay-as-you-go pricing is split into two prompt size tiers. Crossing the 200,000-token threshold triggers a steep price increase:

• Prompts $\le$ 200,000 tokens:
  • Input: $2.00 per 1,000,000 tokens
  • Cached Input: $0.20 per 1,000,000 tokens (90% discount)
  • Output (including thinking tokens): $12.00 per 1,000,000 tokens
• Prompts > 200,000 tokens:
  • Input: $4.00 per 1,000,000 tokens (100% increase)
  • Cached Input: $0.40 per 1,000,000 tokens
  • Output (including thinking tokens): $18.00 per 1,000,000 tokens (50% increase)
• Context Cache Storage: $4.50 per 1,000,000 tokens per hour.

For engineering teams ingesting large codebases or entire technical libraries, this threshold is critical. A prompt with 195,000 input tokens costs $0.39 to process. Expanding that prompt to 205,000 tokens doubles the input rate, costing $0.82—plus increasing the output rate on generated responses from $12.00/M to $18.00/M.

Pricing Across Inference Modes

Grounding Add-ons: Google Search Grounding and Google Maps Grounding each provide 5,000 free requests per month (shared across Gemini 3.x models), then cost $14.00 per 1,000 queries.

---

Thinking Controls, Thought Summaries, and Token Billing

Gemini 3.1 Pro integrates test-time compute directly into inference. Understanding how thinking is configured—and how it is metered—is essential to controlling latency and cost.

Thinking Cannot Be Turned Off

On gemini-3.1-pro-preview, thinking is mandatory. Google does not support disabling thinking on Gemini 3 Pro family models.

Supported configuration values are strictly:

• "low": Allocates minimal reasoning tokens to reduce latency on straightforward problems.
• "medium": Balanced reasoning for intermediate tasks.
• "high" (Default): Allocates maximum reasoning depth for complex multi-step deductions, coding, and mathematical proofs.

Thought Summaries vs. Raw Thoughts

When configuring the API to return thinking data (using parameters such as thinking_summaries or includeThoughts), the API returns summarized thought content.

Google documents that the model's raw internal reasoning process and underlying chain-of-thought tokens remain private and inaccessible. Thought summaries provide visibility into intermediate reasoning milestones without exposing unredacted model internals.

The Operational Cost of Thinking

Intermediate reasoning tokens generated during thinking count against the model's 65,536-token maximum output limit and are billed at full output token rates ($12.00/M or $18.00/M).

On deep reasoning evaluations, Artificial Analysis measured a median Time to First Token (TTFT) of 24.99 seconds for Gemini 3.1 Pro Preview, compared to a class median of 3.97 seconds across standard models. That delay reflects the time spent generating thinking tokens before emitting the first answer token. For interactive chat interfaces where sub-second latency matters, Gemini 3.1 Pro is often poorly suited compared to Flash models running at thinking_level="low".

SDK Configuration Example

from google import genai
from google.genai import types

client = genai.Client()

# Recommended configuration for complex engineering and reasoning tasks
response = client.models.generate_content(
    model="gemini-3.1-pro-preview",
    contents="Analyze this microservice topology for split-brain failure conditions.",
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(
            thinking_level="high"  # Supported: "low", "medium", "high"
        ),
        # Google advises keeping sampling temperature at default (1.0).
        # Clamping temperature can degrade reasoning quality during tool calls.
    ),
)

print(response.text)

---

Evaluation Evidence: Vendor Benchmarks vs. Independent Measurements

Evaluating Gemini 3.1 Pro requires distinguishing between Google's launch evaluations (February 2026) and current independent measurements from Artificial Analysis (October 2026).

The Reasoning Leap Over Gemini 3 Pro

Google's launch data established that Gemini 3.1 Pro is a major architectural improvement in abstract reasoning over the initial Gemini 3 Pro:

• ARC-AGI-2: Scored 77.1% in Thinking (High) mode, more than doubling Gemini 3 Pro's 31.1% (+46.0 percentage points) and outperforming Claude Opus 4.6 (68.8%) and GPT-5.2 (52.9%).
• GPQA Diamond: Scored 94.3% on expert graduate-level science reasoning (vs. 91.9% on Gemini 3 Pro and 92.4% on GPT-5.2). Artificial Analysis independently corroborated this with a 94.14% score.
• Humanity's Last Exam (HLE): Google reported 44.4% with no tools and 51.4% when equipped with search and code execution (where Claude Opus 4.6 led with 53.1%). Artificial Analysis independently measured 47.03% with reasoning enabled.
• MMMU-Pro: Scored 80.5% on Google's multimodal evaluation, with Artificial Analysis independently measuring 82.43%.

Software Engineering and Agentic Coding

In real-world software engineering, 3.1 Pro demonstrated solid gains over Gemini 3 Pro, though the margin is more incremental than in abstract reasoning:

• SWE-Bench Verified: Scored 80.6% on single-attempt GitHub issue resolution, improving on Gemini 3 Pro's 76.2% and matching frontier models like Opus 4.6 (80.8%) and GPT-5.2 (80.0%).
• SWE-Bench Pro: Scored 54.2% on harder enterprise tasks (vs. 43.3% for Gemini 3 Pro and 56.8% for GPT-5.3-Codex).
• SciCode: Solved 59% of research-paper scientific coding problems (vs. 56% on Gemini 3 Pro), closely tracking Artificial Analysis's independent result of 58.68%.
• MCP Atlas: Scored 69.2% on multi-step tool orchestration across Model Context Protocol servers, beating Sonnet 4.6 (61.3%), GPT-5.2 (60.6%), and Opus 4.6 (59.5%).

Understanding Benchmark Version Traps

When comparing benchmark numbers across sources, pay attention to exact version identifiers and test harnesses:

1. Terminal-Bench Versions: Google's model card reported 68.5% on Terminal-Bench 2.0 under the standardized Terminus-2 harness. On that identical harness, 3.1 Pro outperformed Opus 4.6 (65.4%) and GPT-5.3-Codex (64.7%). While OpenAI self-reported 77.3% for GPT-5.3-Codex, that figure came from OpenAI's proprietary Codex harness, not Terminus-2. Furthermore, Artificial Analysis evaluates newer iterations: Terminal-Bench 2.1 (73.78%), Terminal-Bench Hard (53.79%), and Terminal-Bench 4.0 (4.04%), illustrating the sharp drop-off frontier models experience as terminal tasks become longer-horizon.
2. GDPval-AA Elo Divergence: Google's February 2026 model card reported a GDPval-AA Elo of 1317. Artificial Analysis's current leaderboard lists GDPval-AA v2.1 with an Elo of 775.98 (normalized score of 0.138). These figures reflect updated versioning rubrics, prompt sets, and Elo anchoring—not a degradation in model capability.
3. Long-Context Needle Retrieval (MRCR v2): Gemini 3.1 Pro maintains strong multi-needle retrieval at 128k tokens (84.9% average across 8 needles). However, when evaluated pointwise across the full 1-million-token window, retrieval accuracy drops to 26.3%.

Independent Operational Metrics (Artificial Analysis)

Artificial Analysis measures production API performance across thousands of live requests:

• Intelligence Index (v4.3.2): 30 (exact: 29.72), above the model class median of 26.
• Output Generation Speed: 114.2 tokens/second (median), significantly faster than the class median of 69.7 tokens/second once generation begins.
• Time to First Token (TTFT): 24.99 seconds (median), compared to the class median of 3.97s, reflecting extensive test-time thinking before output emission.
• Weighted Task Cost: $0.67 per task across index evaluations.
• AA-Omniscience Index: 31.88 (Knowledge accuracy: 54.85%, Hallucination score: 50.87%).

---

The Lifecycle Dilemma: Building on 3.1 Pro with "3.5 Pro Coming Soon"

A common architectural question for engineering teams is whether to deploy production workloads on Gemini 3.1 Pro given its current lifecycle state.

Two facts create tension:

1. Gemini 3.1 Pro remains in "Preview": It launched in February 2026 and has never transitioned to General Availability (GA). Meanwhile, Google released newer models such as Gemini 3.8 Flash to full GA in September 2026.
2. Google DeepMind announces "3.5 Pro coming soon": Google's official Pro model portal displays an active banner advertising the arrival of Gemini 3.5 Pro.

At the same time, Google's API deprecations documentation lists gemini-3.1-pro-preview with "No shutdown date announced" and leaves the recommended replacement blank. The model is active, fully supported, and not facing imminent deprecation.

Practical Engineering Guidance

---

Frequently Asked Questions

What is the exact API model identifier for Gemini 3.1 Pro?

The standard callable API model ID is gemini-3.1-pro-preview. For agentic workflows combining bash execution and custom tools, Google provides gemini-3.1-pro-preview-customtools. Passing gemini-3.1-pro is not supported.

Why does Gemini 3.1 Pro remain in "Preview" status?

Google has maintained the "Preview" designation on gemini-3.1-pro-preview in both the Gemini API and AI Studio since its February 2026 launch. While subsequent models like Gemini 3.8 Flash launched directly into GA, Google has not transitioned 3.1 Pro to GA. However, it is fully callable and has no announced shutdown date.

Does the Gemini API offer a free tier for Gemini 3.1 Pro?

No. Google's API pricing documentation marks the Free Tier as "Not available" for Gemini 3.1 Pro Preview. Free interactive testing is supported inside Google AI Studio, but API calls require pay-as-you-go billing.

When does Gemini 3.1 Pro pricing increase from $2/$12 to $4/$18?

Pricing jumps when a prompt contains more than 200,000 input tokens. Prompts up to 200k tokens cost $2.00 per million input tokens and $12.00 per million output tokens. Prompts exceeding 200k tokens double the input cost to $4.00 per million tokens and increase output cost to $18.00 per million tokens.

Can I turn off thinking or use thinking_level="minimal"?

No. Thinking is mandatory on gemini-3.1-pro-preview and cannot be disabled. The model supports only "low", "medium", and "high" (default "high"). Passing "minimal" returns an INVALID_ARGUMENT (400) error.

What is the difference between gemini-3.1-pro-preview and gemini-3.1-pro-preview-customtools?

The customtools variant is fine-tuned to prioritize executing custom tools over emitting explanatory text when orchestrating bash commands in autonomous agent workflows. Google cautions that its output quality may fluctuate on non-tool tasks. Both endpoints carry identical pricing.

Has Google announced a shutdown date for Gemini 3.1 Pro?

No. Google's official Gemini API deprecations table lists the shutdown date as "No shutdown date announced" and provides no recommended replacement. While Google DeepMind displays "3.5 Pro coming soon" on its Pro model portal, no calendar date has been released for either 3.5 Pro's launch or 3.1 Pro's retirement.

Where should I read next?

• To evaluate model architectures across reasoning depth, latency budgets, and unit economics, read How to Choose an AI Modelconcept↗.
• For Google's high-speed agentic model with flat context pricing and Antigravity integration, read Gemini 3.8 Flash Overviewentity↗.
• For Google's restricted frontier reasoning model featuring 1M output tokens, read Gemini 4 Argon Overviewentity↗.
• To understand how AI Model Waddle validates claims and separates facts from inference, visit How We Handle Evidenceconcept↗.

Sources & Documentation

1. Google DeepMind: Gemini Pro Model Portal — Model capabilities, Preview status, availability, and '3.5 Pro coming soon' banner. Verified October 1, 2026.
2. Google DeepMind: Gemini 3.1 Pro Model Card — February 19, 2026 release details, architecture foundations, vendor benchmark matrix, and Frontier Safety Framework evaluations.
3. Google Gemini API: Pricing Documentation — Standard 200k token tiers, lack of API free tier, Batch, Flex, Priority, and Grounding add-on rates. Updated October 1, 2026.
4. Google Gemini API: Models Documentation — Model status catalog and Preview category documentation. Updated October 1, 2026.
5. Google Gemini API: Gemini 3.1 Pro Preview Documentation — Token limits (1M/64k), tool support matrix, and dedicated customtools endpoint documentation. Verified October 1, 2026.
6. Google Gemini API: Deprecations Documentation — Lifecycle tracking confirming no shutdown date announced. Verified October 1, 2026.
7. Artificial Analysis: Gemini 3.1 Pro Preview Evaluation — Independent Intelligence Index, speed (114.2 t/s), latency TTFT (24.99s), task cost ($0.67), and benchmark runs. Verified October 1, 2026.