Gemini 3.1 Pro Preview: When the Pro Tier Still Earns Its Price

Gemini 3.1 Pro shipped on 19 February 2026 as a preview, and eight months later the model string still ends in -preview. In that time Google has released four Flash models, several of which beat it on agentic benchmarks at a fraction of its price. The obvious conclusion is that Pro no longer matters. That conclusion is only half right.

This guide covers what 3.1 Pro is, where its published numbers still lead, the pricing detail that catches people out on long prompts, and how to depend on a model that has never left preview.

What Google shipped in February

Google's model card describes 3.1 Pro as built on Gemini 3 Pro: a natively multimodal reasoning model that accepts text, images, audio, video and PDFs, with a context window of 1,048,576 tokens and up to 65,536 output tokens. Output is text only. The card gives no architecture details beyond that, so any parameter count you see quoted for it is a guess.

The developer docs list function calling, structured outputs, context caching, code execution, search grounding, URL context and Google Maps grounding as supported. It is available on the Batch API and on Google's Flex and Priority inference tiers. It does not do image generation, audio generation or the Live API.

Thinking is controlled with thinking_level. On 3.1 Pro the supported values are low, medium and high, with high the default. minimal is not supported. Google's guidance is to leave temperature at its default of 1.0, because lower values can cause looping on reasoning tasks.

The benchmark profile is reasoning first

Google's own launch table, run at high thinking, puts the emphasis on hard reasoning rather than agent loops:

• ARC-AGI-2 (ARC Prize verified): 77.1%, up from 31.1% for Gemini 3 Pro
• GPQA Diamond: 94.3%
• Humanity's Last Exam: 44.4% without tools, 51.4% with search and code
• SWE-Bench Verified: 80.6%; SWE-Bench Pro (public): 54.2%
• Terminal-Bench 2.0 (Terminus-2 harness): 68.5%
• BrowseComp: 85.9%; MCP Atlas: 69.2%
• MRCR v2 8-needle at 128K average: 84.9%

The ARC-AGI-2 jump is the figure that got attention, and it is the one most worth weighing: it is a test of novel abstract reasoning that is hard to train for directly. Two things argued against reading too much into the rest at launch. These are vendor-run numbers, and the model card itself shows Gemini 3 Pro still slightly ahead on MMMU-Pro (81.0% against 80.5%). See vendor-claimed vs independent benchmarks for how to discount a launch table.

The 200K price cliff

Google lists 3.1 Pro Preview at $2.00 per million input tokens and $12.00 per million output for prompts up to 200K tokens. Above 200K, the rates rise to $4.00 input and $18.00 output, and cached input goes from $0.20 to $0.40. Cache storage is $4.50 per million tokens per hour.

The tier is set by the size of the prompt, so the question to ask is not how much context you could send but how much you routinely do. A 1M context window is real, but a request that uses most of it is billed at the higher tier. If your workload habitually sends 250K-token prompts, your effective price is the higher tier, and a retrieval step that trims prompts below 200K pays for itself quickly. The token cost of long context walks through when filling a large window is worth it.

The :batch listing you may see on aggregators such as OpenRouter is the same model on Google's asynchronous Batch API, priced at half the standard rate: $1.00 input and $6.00 output under 200K. You trade latency for the discount, which suits evaluation runs and overnight document work rather than anything interactive.

The customtools endpoint

Alongside the main model, Google publishes gemini-3.1-pro-preview-customtools. It is the same generation tuned for agent harnesses that mix a bash tool with custom tools such as view_file or search_code. The problem it fixes is a model that ignores your purpose-built tools and shells out to grep and cat instead.

Google is explicit that quality may fluctuate on work that does not involve those tools. Treat it as a harness-specific variant: use it inside a coding agent with a defined tool set, and keep the plain endpoint for everything else.

The Flash models have caught up on agent work

This is the uncomfortable part for anyone paying Pro rates. Google's own Gemini 3.5 Flash model card, published in May, lists 3.5 Flash ahead of 3.1 Pro on Terminal-bench 2.1 (76.2% against 70.3%), MCP Atlas (83.6% against 78.2%) and OSWorld-Verified (78.4% against 76.2%). The 3.6 Flash card in July widened the gap on long-context retrieval, with MRCR v2 at 128K of 91.8% against Pro's 84.9%.

Where Pro still led in those same tables was pure reasoning: Humanity's Last Exam (44.4% against 40.2% for 3.5 Flash) and ARC-AGI-2 (77.1% against 72.1%). Those are the tasks where a model has to think rather than act.

Price makes the decision for most agent loops. Gemini 3.6 Flash is currently listed at an introductory $0.75 input and $3.75 output per million tokens, so Pro costs more than three times as much on output. For a tool-calling loop that makes eighty requests, that multiplier dominates. The Gemini 3.6 Flash guide covers the cheaper option in detail.

Living with a model that never left preview

Preview status means Google can change behaviour, pricing or availability with less notice than a stable model gets. The Gemini 3 developer guide, which documented 3.1 Pro's thinking levels, is now itself marked as deprecated and points readers to the Gemini 3.8 Flash guide. Gemini 3.5 Pro, announced at I/O in May, still does not appear on the Gemini API models page, which now lists Gemini 4 Argon as Google's flagship reasoning model. None of that tells you when 3.1 Pro will be retired, but it does tell you it is not where Google's attention is.

The practical response is the same as for any preview tier, and it comes down to three habits:

1. Route to 3.1 Pro only for the tasks where you have measured a reasoning advantage, and send everything else to a stable Flash model.
2. Keep a fallback configured so a deprecation notice becomes a config change, not an outage. Model routing and fallbacks has the pattern.
3. Re-run your own evaluation set whenever Google ships a new tier. Pro's lead is narrow and specific, and each Flash release has eaten into it.

The decision rule is short. If the job is hard reasoning with few tool calls, such as a difficult maths derivation or an algorithm design question, 3.1 Pro is still a defensible choice. If the job is an agent loop, the Flash line now does it for less.