Skip to main content

Gemini 3.1 Pro Thinking Levels: Low vs High, Defaults, Gemini 4

Gemini 3.1 Pro takes thinking_level low, medium, or high and defaults to high; minimal errors. Gemini 4 Argon defaults to medium, the Gemini app to Low.

LaoZhang AI TeamPublishedUpdated 21 min read
On this page
Gemini 3.1 Pro thinking levels cover with LOW, MEDIUM and HIGH cards and their thinking_level settings

Gemini 3.1 Pro has three thinking levels: low, medium, and high. In the Gemini API you set them with thinkingLevel (REST and JavaScript) or thinking_level (Python and the Interactions API). If you don't set one, 3.1 Pro uses high, the level that spends the most thinking tokens, and every thinking token is billed at the output rate of $12 per million tokens for prompts up to 200k tokens. minimal is not a valid value on 3.1 Pro and returns a 400 error. Thinking can't be switched off.

The defaults differ by model and by product, as of October 22, 2026. Gemini 4 Argon accepts the same three levels but defaults to medium in the API. The Gemini app defaults to Low if you don't pick a level. Antigravity offers Gemini 3.1 Pro only at Low or High, and only on Enterprise plans. "Deep Think Mini" is not a Google setting: it was a reviewer's nickname for 3.1 Pro at high. Pick the level per task, set it explicitly in code, and check the thought-token count the API returns.

What Are Gemini 3.1 Pro Thinking Levels and Why They Matter

A thinking level sets how far Gemini 3.1 Pro may reason internally before it writes the answer. Google's Gemini 3 developer guide says thinking_level controls "the maximum depth" of that reasoning, and that the levels are "relative allowances for thinking rather than strict token guarantees." high is also dynamic: the model can still think briefly on an easy prompt. A level is a ceiling, not a fixed token count.

The level matters for three reasons:

  • Cost. Thinking tokens are billed as output tokens. On 3.1 Pro that is $12 per million for prompts up to 200k tokens and $18 per million above that, per Google's Gemini API pricing page. You pay for the full reasoning even if you only ask for a short summary of it.
  • Latency. Google describes high as the level where the model "may take significantly longer to reach a first (non thinking) output token."
  • Answer quality on hard problems. Google publishes its 3.1 Pro benchmark scores for the high level, listed in the Deep Think section below.

The default is what catches people. On 3.1 Pro, an API call without a level runs at high. A classifier or a translation job sent that way pays for the deepest reasoning on every request.

For the spec questions that come up next to this one: 3.1 Pro accepts up to 1M input tokens and returns up to 64k output tokens, with thinking counted inside the output limit. Google doesn't publish a parameter count.

Gemini thinking level by model: values, defaults, and what errors

Use this table to check a model before you set a level. It combines Google's thinking guide for the generateContent API, the thinking guide for the Interactions API, the Gemini 3 guide, and the Gemini API deprecations page, as of October 22, 2026.

Model IDLevels acceptedDefault if you send noneWhat fails or doesn't work
gemini-3.1-pro-previewlow, medium, highhighminimal returns 400; thinking can't be turned off
gemini-4-argonlow, medium, highmediumminimal returns an error; thinking can't be turned off; thinkingBudget is deprecated
gemini-3.8-flashlow, medium, highmediumminimal returns an error
gemini-3.6-flashminimal, low, medium, highmediumnone listed
gemini-3.5-flash-liteminimal, low, medium, highminimalminimal can still think a little on complex prompts
gemini-3-flash-previewminimal, low, medium, highhighthinking can't be fully turned off
gemini-3-pro-previewlow, highhighShut down March 9, 2026; move to gemini-3.1-pro-preview
gemini-2.5-prothinkingBudget 128–32,768dynamic (-1)Budget can't disable thinking; no thinkingLevel on generateContent
gemini-2.5-flashthinkingBudget 0–24,576dynamic (-1)0 turns thinking off; no thinkingLevel on generateContent

Three details in that table trip people up:

  • The 2.5 models depend on which API you call. The generateContent guide says 2.5 models don't support thinkingLevel and need thinkingBudget. The Interactions guide lists low, medium, and high for them. Google also limits 2.5 access to accounts that used those models before.
  • minimal is not "off." Even where it is accepted, Google says it "does not guarantee that thinking is off."
  • Sending thinkingLevel and thinkingBudget together returns 400 on Gemini 3 and Gemini 4 models.

The parameter has a different name and location in each API:

Where you call GeminiSet the level hereRead the thought-token count hereGet thought summaries
generateContent, REST or JavaScriptgenerationConfig.thinkingConfig.thinkingLevelusageMetadata.thoughtsTokenCountincludeThoughts: true
generateContent, Python SDKtypes.ThinkingConfig(thinking_level="low")usage_metadata.thoughts_token_countinclude_thoughts=True
Interactions APIgeneration_config.thinking_levelusage.total_thought_tokensgeneration_config.thinking_summaries: "auto"
OpenAI-compatible endpointreasoning_effort, or thinking_config inside extra_bodynot listed on Google's compatibility pageinclude_thoughts inside extra_body

Google's samples use lowercase strings ("low"). The JavaScript SDK also exports a ThinkingLevel enum (ThinkingLevel.LOW).

Understanding LOW, MEDIUM, and HIGH: What Each Level Actually Does

Gemini 3.1 Pro LOW, MEDIUM and HIGH thinking levels comparison chart from February 2026

Chart from February 2026. Where it differs from the text, including the "Deep Think Mini" label, the token ranges, and the model list, the text is current as of October 22, 2026.

Google describes each level in one line in its thinking guides:

  • low: "Minimizes latency and cost." The Gemini 3 guide adds: "Best for simple instruction following, chat, or high-throughput applications." On 3.1 Pro, low still thinks. It is the lowest setting the model accepts.
  • medium: "Balanced thinking for most tasks." On 3.1 Pro you have to request it. It is the default on Gemini 4 Argon, Gemini 3.8 Flash, and Gemini 3.6 Flash.
  • high: "Maximizes reasoning depth." This is the default on 3.1 Pro, and the one Google's benchmark figures use.

None of the three is a different model or a different price per token. The level changes how many thinking tokens the model may spend, so it changes the bill and the wait. It doesn't change the rate.

Gemini 3.1 Pro High vs Low: the difference in one test

On a short reasoning prompt, high spent 20–30% more thinking tokens than low on 3.1 Pro: 1,200 and 1,297 in two high runs, against 995 at low. Every run gave the correct answer. That is a narrower gap than the "low barely thinks" description suggests. Google says the levels are allowances, not targets, so on an easy prompt high doesn't need to use its full room. Expect the gap to widen on hard multi-step problems, where high is allowed to keep going and low isn't. The full test is in the next section.

In the products, "High" and "Low" are labels on the same model:

  • Antigravity lists Gemini 3.1 Pro as Low or High, with no Medium.
  • The Gemini app lets you pick Low, Medium, or High on each model.
  • Google AI Studio shows a "Thinking level" dropdown in the run settings.

If you see "Gemini 3.1 Pro (High)" or "gemini-3.1-pro-low" in a model list, it is the same 3.1 Pro model with a thinking level attached.

Measured thought tokens per level: 3.1 Pro, Argon, and 3.8 Flash

On October 22, 2026, a set of single-sample calls went through the Gemini-native endpoint of laozhang.ai, a proxy in front of Google's API run by the same team as this blog. The calls checked two things: which parameter combinations the API accepts, and how many thinking tokens each level spends on one prompt. The prompt was a small purchase puzzle: pens come in packs of 6 for $4.50 and packs of 10 for $7.00, and you need at least 58 at the lowest price. Every model and level below answered it correctly ($41.50). Costs use Google's list prices: input plus thinking and answer tokens at the output rate. The whole run cost $0.088.

Disclosure: laozhang.ai is our own service. Its prices are platform pricing, not an independent comparison.

Model and levelThought tokensAnswer tokensCost per call at list price
3.1 Pro, no level set (runs at high)1,20045$0.0151
3.1 Pro, low99545$0.0126
3.1 Pro, medium1,12048$0.0141
3.1 Pro, high, with thought summaries on1,29748$0.0163
Argon, no level set (runs at medium)31657$0.0039
Argon, medium38449$0.0045
Argon, high91649$0.0098
3.8 Flash, no level set (runs at medium)1,05846$0.0044
3.8 Flash, low49946$0.0023

Prompt tokens were 61 on 3.1 Pro and 92 on Argon. The proxy reported 315 on 3.8 Flash for the same text, which it didn't explain. Argon costs use the price that runs through November 30, 2026.

Parameter checks, one call each, against 3.1 Pro unless noted:

RequestResult
thinkingLevel: "minimal"400, as the docs say
thinkingLevel: "low" plus thinkingBudget: 1024400, as the docs say
thinkingBudget: 0Accepted, but still 124 thought tokens: thinking stayed on
thinkingBudget: 1024 aloneAccepted (legacy parameter)
thinkingLevel: "LOW" in upper caseAccepted
Argon, thinkingBudget: 1024 aloneAccepted, as the docs describe for the deprecated budget

Argon at low and minimal couldn't be measured: the proxy's Argon pool returned rate-limit errors (429). These are single samples on one easy prompt, not a benchmark. A test of 12 prompts on the same proxy a week earlier found the same pattern on 3.1 Pro: it spent thinking tokens on every call, including at low (410–1,620 per call). Argon reported zero thinking tokens at default, low, and medium on those prompts, and 646–1,461 at high. Argon's medium thinking varies by prompt, so it isn't free in general. The Gemini 4 Argon vs 3.1 Pro cost test before and after Dec 1 has those 12 prompts in full.

How to Choose the Right Thinking Level for Your Task

Choose the lowest level that still passes on your own sample of real requests. Then set that level explicitly so the default never decides for you. Google's thinking guides sort tasks into three groups, and the levels follow from that:

  • Low: retrieval, classification, extraction, formatting, short rewrites, chat replies, and high-volume pipelines. Google's own examples here are "Where was DeepMind founded?" and "Is this email asking for a meeting or just providing information?" On 3.1 Pro, low still thinks. If cost per call is your main constraint, a Flash model at low is cheaper: 3.8 Flash at low cost $0.0023 on the test prompt, against $0.0126 for 3.1 Pro at low.
  • Medium: everyday code generation, writing, comparisons, and moderate analysis. Google's examples are comparing electric and hybrid cars, or drawing an analogy. On 3.1 Pro, medium is the setting to try first when low misses details. Argon already runs here by default.
  • High: multi-step math, hard debugging, planning, agentic coding, and problems where a wrong answer costs more than the tokens. Google recommends maximum thinking for "complex math problems or coding tasks." This is what 3.1 Pro does when you leave the level unset.

Two checks keep this honest:

  • Read the thought-token field (thoughtsTokenCount or total_thought_tokens) for a day of traffic before and after a change.
  • Move up a level only when answers actually fail. On short tasks, the savings from dropping 3.1 Pro to low may be small, as the test above showed.

When high on 3.1 Pro still isn't enough, change the model rather than the level. Gemini 4 Argon is Google's newer flagship reasoning model, and its API price doubles on December 1, 2026. For plan-by-plan access to it, see Gemini 4 Argon usage limits: Pro vs Ultra, resets, API credits.

Complete Code Examples: Python, JavaScript, and REST API

All examples use gemini-3.1-pro-preview and read the key from the GEMINI_API_KEY environment variable. Swap in gemini-4-argon or gemini-3.8-flash as needed, but check the model table first: minimal fails on all three.

Python (Google Gen AI SDK)

python
from google import genai
from google.genai import types

client = genai.Client()  # reads GEMINI_API_KEY from the environment


def ask(prompt: str, level: str) -> str:
    response = client.models.generate_content(
        model="gemini-3.1-pro-preview",
        contents=prompt,
        config=types.GenerateContentConfig(
            thinking_config=types.ThinkingConfig(thinking_level=level)
        ),
    )
    usage = response.usage_metadata
    print(f"{level}: {usage.thoughts_token_count} thought tokens, "
          f"{usage.candidates_token_count} answer tokens")
    return response.text


ask("Classify this email as billing, support, or sales: ...", "low")
ask("Write a Python function that merges two sorted lists in O(n).", "medium")
ask("Find the race condition in this code and explain the fix: ...", "high")

To see what the model reasoned about, turn on thought summaries. You are billed for the full reasoning, not the summary:

python
response = client.models.generate_content(
    model="gemini-3.1-pro-preview",
    contents="Prove that there are infinitely many prime numbers.",
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(
            thinking_level="high", include_thoughts=True
        )
    ),
)
for part in response.candidates[0].content.parts:
    if not part.text:
        continue
    print("Thought summary:" if part.thought else "Answer:", part.text)

JavaScript (Google Gen AI SDK)

javascript
import { GoogleGenAI, ThinkingLevel } from "@google/genai";

const ai = new GoogleGenAI({}); // reads GEMINI_API_KEY from the environment

const response = await ai.models.generateContent({
  model: "gemini-3.1-pro-preview",
  contents: "Explain the difference between REST and GraphQL with one example each.",
  config: {
    thinkingConfig: {
      thinkingLevel: ThinkingLevel.LOW,
    },
  },
});

console.log(response.text);
console.log("Thought tokens:", response.usageMetadata.thoughtsTokenCount);

The ThinkingLevel enum also has HIGH, MEDIUM, and MINIMAL. Use MINIMAL only on models that accept it, such as 3.6 Flash, Flash-Lite, and 3 Flash. On 3.1 Pro, Argon, and 3.8 Flash it fails.

REST API (cURL)

bash
curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-3.1-pro-preview:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -X POST \
  -d '{
    "contents": [{
      "parts": [{"text": "Extract every email address from this text: ..."}]
    }],
    "generationConfig": {
      "thinkingConfig": {
        "thinkingLevel": "low"
      }
    }
  }'

Send the key in the x-goog-api-key header rather than a ?key= query string, so it doesn't end up in logs and shell history. The response's usageMetadata.thoughtsTokenCount shows what the level spent.

Interactions API

Google's Gemini 3 guide now uses the Interactions API in its examples. Here the level sits directly in generation_config:

bash
curl -X POST "https://generativelanguage.googleapis.com/v1beta/interactions" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-3.1-pro-preview",
    "input": "Plan a migration of a cron job to a queue worker in five steps.",
    "generation_config": {
      "thinking_level": "medium",
      "thinking_summaries": "auto"
    }
  }'

In Python, the same request is client.interactions.create(model=..., input=..., generation_config={"thinking_level": "medium"}). Read the thinking spend from interaction.usage.total_thought_tokens.

OpenAI-compatible endpoint: reasoning_effort

If your code uses the OpenAI SDK against Gemini, set reasoning_effort. Google maps it to thinking_level on Gemini 3 models:

python
import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["GEMINI_API_KEY"],
    base_url="https://generativelanguage.googleapis.com/v1beta/openai/",
)
response = client.chat.completions.create(
    model="gemini-3.1-pro-preview",
    reasoning_effort="medium",
    messages=[{"role": "user", "content": "Review this SQL query for N+1 problems: ..."}],
)
print(response.choices[0].message.content)

Google's OpenAI compatibility guide publishes these mappings:

  • Gemini 4 Argon: minimal becomes low; low, medium, and high map one to one.
  • Gemini 3 Flash and 3.1 Flash-Lite: all four values map one to one.
  • Gemini 2.5: low is a budget of 1,024 tokens, medium 8,192, and high 24,576.

The table has no 3.1 Pro column, so stick to low, medium, or high there. You can't send reasoning_effort together with a thinking_config in extra_body. "none" turns thinking off only on 2.5 Flash models.

Deep Think Mini: What Happens When You Set HIGH

Setting high doesn't switch on a separate mode called Deep Think Mini, because there is no such Google mode or API value. The name comes from VentureBeat's February 2026 first-impressions headline, which called 3.1 Pro "a 'Deep Think Mini' with adjustable reasoning on demand." It was a reviewer's description of 3.1 Pro at its top thinking level. It doesn't appear in Google's model pages, thinking guides, or parameter tables.

What high actually does is raise the ceiling on the same thinking process the other levels use. The model may reason for longer before it answers, and it decides how much of that room to use. In the test above, high spent 1,297 thinking tokens on an easy puzzle, not thousands. On hard problems it can spend far more, and all of it bills at the output rate.

Google's published scores for "Gemini 3.1 Pro Thinking (High)" on its Gemini 3.1 Pro model page:

BenchmarkGemini 3.1 Pro, Thinking (High)
ARC-AGI-2 (ARC Prize Verified)77.1%
GPQA Diamond94.3%
SWE-Bench Verified, single attempt80.6%
Humanity's Last Exam, no tools44.4%
Humanity's Last Exam, with search and code51.4%

Deep Think itself is something else. Gemini 3.1 Deep Think is a separate reasoning mode that Google built on top of 3.1 Pro in February 2026. Google reports 84.6% on ARC-AGI-2 and 48.4% on Humanity's Last Exam without tools for it. It lives in the Gemini app, not in the API: the thinking guides have no Deep Think parameter. Google's Deep Think help page says it requires Google AI Ultra, is experimental, and can take a few minutes per answer. The page also says "Deep Think is being replaced by Gemini 4 Argon." For the most reasoning you can get through the API today, use Argon at high.

Thinking level in the Gemini app, AI Studio, and Antigravity

The products use the same Low/Medium/High words, but each has its own default.

ProductWhere the setting isLevelsDefault
Gemini appModel menu under the chat boxLow, Medium, High, on every modelLow, unless you pick one
Google AI StudioRun settings, "Thinking level"Low, Medium, HighHigh on 3.1 Pro, Medium on Argon
AntigravityModel selector under the prompt box3.1 Pro: Low, High. Argon: Low, Medium, HighHigh on 3.1 Pro and on Argon
Gemini APIthinkingLevel / thinking_levelSee the model tablehigh on 3.1 Pro, medium on Argon

Gemini app: Low, Medium, High replace Standard and Extended

In the Gemini app, click the current model name at the bottom of the chat box, choose a model, then choose Low, Medium, or High. Google's help page on Gemini models describes the levels this way:

  • Low: "fast, concise answers for straightforward queries."
  • Medium: balances speed and reasoning depth "for everyday tasks."
  • High: "carefully process and reason through complex, multistep prompts."

If you don't select a level, it defaults to Low. Your choice holds for the rest of that conversation.

Higher levels and stronger models "require more of your usage." When a paid plan hits its limit, the chat continues on Flash-Lite. Google's help page on model access and limits calls the same setting an "effort level."

Which models you can pick depends on your plan:

  • No plan: Auto or Flash-Lite.
  • Google AI Plus: Flash-Lite, Flash, and Pro. Google is emailing Plus subscribers about changes to this.
  • Google AI Pro and Ultra: Argon as well.

If you are deciding which plan to buy, Stable & Cheap Gemini Pro Upgrade: Official Google AI Pro Deals compares the current offers.

The menu has changed twice this year. On May 17, 2026, 9to5Google reported a "Thinking level" menu with Standard and Extended options for Gemini 3 Flash and Gemini 3.1 Pro, on a limited rollout. On October 9, 2026, 9to5Google reported that Low, Medium, and High replaced "Extended", matching AI Studio and Antigravity. On the same date, accounts without a plan moved to Auto. Auto picks both the model and the thinking level for each prompt. To see what it chose, open More, then Show thinking steps.

Google AI Studio: the Thinking level dropdown

In Google AI Studio, open Run settings and use the "Thinking level" dropdown. As of October 22, 2026, it starts at High for Gemini 3.1 Pro Preview, which also shows a temperature slider set to 1. For Gemini 4 Argon it starts at Medium, with Low, Medium, and High available and no temperature slider. Argon in AI Studio needs a Google AI plan or a paid API key. 3.1 Pro is free to try in AI Studio, but gemini-3.1-pro-preview has no free tier in the API. Gemini API free tier 2026: limits, free models, and free API keys lists which API models are free.

Antigravity: Gemini 3.1 Pro (High) and Gemini 3.1 Pro (Low)

Antigravity's models page lists Gemini 3.1 Pro with two levels, Low and High, defaulting to High. The plan table makes 3.1 Pro available only on Enterprise. Gemini 4 Argon offers Low, Medium, and High there and defaults to High, which is different from its medium default in the API. Argon is available on Google AI Pro (paid, not trial), Ultra, and Enterprise, in the Antigravity app, CLI, and IDE extensions, but not in the Antigravity IDE. The Gemini 3.8, 3.7, and 3.6 Flash models add a Fast option and default to Medium.

Thinking Token Costs and How to Optimize Your Spending

Every thinking token is billed as an output token, so the cost of a level is: thinking tokens × the model's output price. Google's thinking guides add three rules:

  • You pay for the full reasoning, even when the API returns only a summary of it.
  • max_output_tokens counts thinking and answer together. If a call hits the limit mid-reasoning, the answer comes back empty or cut off, and the thinking is still billed.
  • Thought signatures you send back in multi-turn chats count as input tokens.

Google's advice is to lower the thinking level rather than set a small max_output_tokens.

Paid-tier prices per million tokens, from Google's Gemini API pricing page as of October 22, 2026:

ModelInputOutput, thinking includedPrice change ahead
Gemini 3.1 Pro, prompts up to 200k$2.00$12.00No change announced
Gemini 3.1 Pro, prompts over 200k$4.00$18.00No change announced
Gemini 4 Argon$2.00$10.00$4.00 and $20.00 from December 1, 2026
Gemini 3.8 Flash and 3.6 Flash$0.75$3.75$1.50 and $7.50 from January 1, 2027

Neither 3.1 Pro nor Argon has a free API tier. Batch requests cost half: 3.1 Pro is $1 input and $6 output up to 200k. Cached input on 3.1 Pro is $0.20 per million up to 200k ($0.40 above), plus $4.50 per million tokens per hour of storage. Argon supports only implicit caching. The full rate card is in Gemini API pricing 2026: token rates, free tier, and key cost.

What the levels cost in practice:

  • Per 1,000 thinking tokens: $0.012 on 3.1 Pro. On Argon it is $0.010 until November 30, 2026, and $0.020 from December 1.
  • 3.1 Pro, high vs low, on the test prompt:
    • High: 61 input tokens × $2 per million, plus 1,345 thinking and answer tokens × $12 per million, comes to $0.0163 per call.
    • Low: 61 × $2 per million plus 1,040 × $12 per million comes to $0.0126.
    • The gap is $0.0037 a call, or about $3.70 per 1,000 calls.
  • Argon at December 1 prices, same prompt: high comes to 92 × $4 per million plus 965 × $20 per million, which is $0.0197 per call. Medium comes to $0.0090.

Treat these as one prompt's numbers, not a workload average. Measure your own traffic with the thought-token field before and after you change levels.

Ways to cut the bill without hurting the calls that need depth:

  1. Set the level on every 3.1 Pro request, so nothing runs at high by accident.
  2. Route easy, high-volume calls to low, or to a Flash model.
  3. Send work that can wait to the Batch API at half price.
  4. Cache long, repeated system prompts or documents.
  5. If you hit 429 errors after spreading calls across models, Gemini API rate limits 2026: free vs paid tiers and how to raise explains the tiers.

Migrating from thinking_budget to thinking_level

Chart from February 2026 mapping Gemini 2.5 thinking_budget values to Gemini 3.1 Pro thinking_level

Chart from February 2026. Google publishes no general budget-to-level conversion table, and medium is not the 3.1 Pro default. The text below is current as of October 22, 2026.

To migrate from a 2.5-era thinking_budget to Gemini 3.1 Pro:

  1. Remove thinking_budget.
  2. Add thinking_level.
  3. Delete any explicit temperature.
  4. Measure the thought tokens again.

On 3.1 Pro, thinking_budget still works for backward compatibility, and Google recommends migrating "for more predictable performance." Sending both parameters in one request returns 400. On Gemini 4 Argon, the budget is deprecated: sent alone, it maps to the nearest level and returns a warning, and sent with a level it returns 400.

Google doesn't publish a conversion from token budgets to levels for 3.1 Pro. The closest reference is the OpenAI-compatibility mapping: medium means thinking_level: "medium" on Gemini 3 and Gemini 4 models and a budget of 8,192 on 2.5. Use it as a starting point, then compare thoughtsTokenCount before and after. A budget of 0 has no equivalent on 3.1 Pro: the test call with thinkingBudget: 0 still spent 124 thinking tokens.

The temperature step answers the "remove custom sampling parameters" part of the migration. For all Gemini 3 models, Google strongly recommends keeping temperature at its default of 1.0. It warns that lower values "may lead to unexpected behavior, such as looping or degraded performance." Its migration notes say to remove an explicit temperature you set for deterministic output on 2.5. AI Studio no longer shows a temperature slider for Argon.

python
from google.genai import types

# BEFORE: Gemini 2.5 Pro with a token budget and a low temperature
response = client.models.generate_content(
    model="gemini-2.5-pro",
    contents="Solve this scheduling problem: ...",
    config=types.GenerateContentConfig(
        temperature=0.2,
        thinking_config=types.ThinkingConfig(thinking_budget=8192),
    ),
)

# AFTER: Gemini 3.1 Pro with a thinking level and the default temperature
response = client.models.generate_content(
    model="gemini-3.1-pro-preview",
    contents="Solve this scheduling problem: ...",
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(thinking_level="medium"),
    ),
)
print(response.usage_metadata.thoughts_token_count)

If your code still calls gemini-3-pro-preview, change the model ID first. That model shut down on March 9, 2026, and it only ever accepted low and high.

Common Mistakes and Troubleshooting

The bill is higher than expected. 3.1 Pro runs at high whenever the request carries no level, and that includes calls built by wrappers and frameworks that never set one. Add the level in one place, then check thoughtsTokenCount on a few logged responses.

400 error after adding minimal. minimal is rejected on 3.1 Pro, Gemini 4 Argon, and Gemini 3.8 Flash. Use low. If you were copying a 3 Flash or Flash-Lite example, minimal is valid only on those models and on 3.6 Flash.

400 error with both parameters. A request that carries both thinkingLevel and thinkingBudget fails on Gemini 3 and Gemini 4 models. This often happens when a default config is merged with a per-request override. Build the thinking config in one helper:

python
from google.genai import types

NO_MINIMAL = {"gemini-3.1-pro-preview", "gemini-4-argon", "gemini-3.8-flash"}


def thinking_config(model: str, level: str) -> types.ThinkingConfig:
    """Return a level-only config; never add thinking_budget next to it."""
    if level == "minimal" and model in NO_MINIMAL:
        level = "low"  # these models return an error for minimal
    return types.ThinkingConfig(thinking_level=level)

thinkingLevel rejected on a 2.5 model. On generateContent, 2.5 models take thinkingBudget only. On the OpenAI-compatible endpoint, reasoning_effort plus a thinking_config in extra_body is also a conflict. Pick one.

Empty or cut-off answer with finish_reason: MAX_TOKENS. The thinking used up max_output_tokens. Raise the limit or lower the level. The thinking spent before the cut-off is billed either way.

Looping or repetitive answers. Check for a temperature below 1.0 left over from older code.

Argon cuts off long non-streaming answers. Argon's generateContent returns up to 32,768 tokens per non-streaming request and 65,536 when streaming. For longer output, use the continuation or long-decode options in Google's thinking guide.

Gemini 3.1 Pro thinking level FAQ

What is the difference between Gemini 3.1 Pro High and Low?

They are the same model at the same price per token. High lets 3.1 Pro reason longer before answering. Low "minimizes latency and cost" but still thinks. In a one-prompt test, high spent 1,297 thinking tokens and low 995, and both got the right answer. Expect a bigger gap on hard multi-step problems. In Antigravity, "Gemini 3.1 Pro (High)" and "(Low)" are these two settings.

What is the default thinking level for Gemini 3.1 Pro?

High, in the API and in Google AI Studio. Antigravity also starts 3.1 Pro at High. The Gemini app is different: it uses Low if you don't pick a level. Gemini 4 Argon defaults to medium in the API and AI Studio, and to High in Antigravity.

Is Gemini 3.1 Pro the same as Extended thinking?

No. Gemini 3.1 Pro is a model. Extended was a thinking level in the Gemini app, offered from May 2026 on Gemini 3 Flash and 3.1 Pro next to Standard. In October 2026, the app replaced it with Low, Medium, and High. On any model, High is now the setting that allows the most reasoning.

How do I make Gemini think longer?

In the Gemini app, open the model menu under the chat box, pick Pro (or Argon on Google AI Pro and Ultra), and set the thinking level to High. It stays on High for that conversation and uses more of your limit. In the API, set thinkingLevel to "high", which is already the 3.1 Pro default, or switch to gemini-4-argon with "high".

What does Auto do with low, medium, and high in the Gemini app?

Auto chooses the model and the thinking level for each prompt. Accounts without a plan have used Auto since October 9, 2026, with most prompts going to Flash-Lite. To see which model answered, open More, then Show thinking steps.

Gemini Thinking vs Pro: which should I pick?

In the current Gemini app, thinking is a level you set on each model, not a separate model. Flash at High is still Flash, and Pro at Low is still Pro. Pick the model for capability and the level for depth. For hard reasoning, use Pro or Argon at High. For quick everyday questions, use Flash or Pro at Low.

Can I use Gemini 3.1 Pro Deep Think in the API?

No. Deep Think is a Gemini app mode for Google AI Ultra subscribers, and Google says Gemini 4 Argon is replacing it. The thinking guides have no Deep Think parameter. Setting high on 3.1 Pro is not Deep Think.

Can I turn thinking off on Gemini 3.1 Pro?

No. low is the minimum, minimal returns 400, and thinkingBudget: 0 still thinks. If you need a model that can run with almost no thinking, use a Flash or Flash-Lite model that accepts minimal. To turn thinking fully off, use Gemini 2.5 Flash with a budget of 0. If you are moving to the newest Flash, Gemini 3.8 Flash: pricing, model ID, and a safe 3.7 migration covers the switch.

Sources15

External pages this guide links to, in the order they appear. Last updated Oct 22, 2026.

  1. 1.Gemini 3 developer guideai.google.dev/gemini-api/docs/gemini-3
  2. 2.Google's Gemini API pricing pageai.google.dev/gemini-api/docs/pricing
  3. 3.thinking guide for the generateContent APIai.google.dev/gemini-api/docs/generate-content/thinking
  4. 4.thinking guide for the Interactions APIai.google.dev/gemini-api/docs/thinking
  5. 5.Gemini API deprecations pageai.google.dev/gemini-api/docs/deprecations
  6. 6.OpenAI compatibility guideai.google.dev/gemini-api/docs/openai
  7. 7.Gemini 3.1 Pro model pagedeepmind.google/models/gemini/pro
  8. 8.Gemini 3.1 Deep Thinkdeepmind.google/models/gemini/deep-think
  9. 9.Deep Think help pagesupport.google.com/gemini/answer/16345172
  10. 10.help page on Gemini modelssupport.google.com/gemini/answer/18648722
  11. 11.model access and limitssupport.google.com/gemini/answer/17004136
  12. 12.reported a "Thinking level" menu9to5google.com/2026/05/17/gemini-app-thinking-level
  13. 13.reported that Low, Medium, and High replaced "Extended"9to5google.com/2026/10/09/gemini-auto-free-users
  14. 14.Google AI Studioaistudio.google.com
  15. 15.Antigravity's models pageantigravity.google/docs/models
Gemini 3.1 Pro Customtools complete guide for AI agent development
Gemini & Google AI

Gemini 3.1 Pro Customtools: The Complete Guide to Building Reliable AI Agents (2026)

Gemini 3.1 Pro Customtools is Google's specialized model variant that prioritizes your custom tools over bash commands. This guide covers everything from the standard vs customtools behavioral difference to building production-ready AI agents with working code examples, dynamic routing strategies, and verified pricing data.

16 min