Your personal memory, across sessions, agents, and devices.

Gemini 3.1 Pro ARC-AGI-2 Score and Benchmark Results

memU Team memU Team

Google released Gemini 3.1 Pro on February 19, 2026, and its headline number was a verified 77.1% on ARC-AGI-2. Gemini 3 Pro had scored 31.1% on the same test. On the harder, interactive ARC-AGI-3 it scored under 1%. It costs $2 per million input tokens and is still in preview. In October, Google's Gemini 4 Argon beat it on both score and cost per task. If you build agents on Gemini, the last section explains what its reasoning keeps between conversations.

What is Gemini 3.1 Pro, in one sentence?

Gemini 3.1 Pro is Google's February 2026 reasoning model, still in preview as gemini-3.1-pro-preview, with a 1M-token input window and a verified 77.1% on ARC-AGI-2.

Google's launch post said it was "releasing 3.1 Pro in preview today" and would make it generally available soon. Developers got it through the Gemini API and Vertex AI. Consumers got it in the Gemini app. On the API model page the status still reads Preview, the input limit is 1,048,576 tokens, and the output limit is 65,536 tokens.

How Gemini 3.1 Pro differs from Gemini 3 Pro

All numbers below come from the table on Google DeepMind's Gemini Pro page, with both Gemini models at Thinking (High).

Benchmark Gemini 3 Pro Gemini 3.1 Pro Change
ARC-AGI-2 (ARC Prize verified) 31.1% 77.1% +46.0 points
Humanity's Last Exam, no tools 37.5% 44.4% +6.9 points
GPQA Diamond 91.9% 94.3% +2.4 points
SWE-Bench Verified, single attempt 76.2% 80.6% +4.4 points
Terminal-Bench 2.0 (Terminus-2 harness) 56.9% 68.5% +11.6 points
MMMU-Pro 81.0% 80.5% -0.5 points

MMMU-Pro, a multimodal test, is the one row where 3.1 Pro slipped, by half a point.

What did Gemini 3.1 Pro score on ARC-AGI-2?

ARC Prize lists Gemini 3.1 Pro (Preview) at 77.08% on the ARC-AGI-2 semi-private set, at $0.96 per task, in the data behind its leaderboard. Google rounds that to 77.1%. Its evaluation methodology PDF says the figure is sourced from the ARC Prize website.

ARC-AGI-2 is a set of grid puzzles. Each task gives a few input and output examples, and the model must infer the rule and apply it to a new grid. ARC Prize kept only tasks that at least two humans solved within two attempts. It also publishes cost per task next to every score, so a model cannot buy a high score with unlimited compute and hide the bill.

For comparison, Google's table lists Claude Opus 4.6 at 68.8%, Claude Sonnet 4.6 at 58.3% and GPT-5.2 at 52.9%. Those were the competitors' February scores at their maximum reasoning settings. On the ARC Prize leaderboard, Gemini 3 Deep Think scored 84.58%, but at $13.62 per task. That is about 14 times the cost of 3.1 Pro for 7.5 more points.

How does Gemini 3.1 Pro do on ARC-AGI-3?

Badly, like every model of its generation. ARC Prize launched ARC-AGI-3 on March 25, 2026. It replaced static puzzles with hundreds of interactive, game-style environments. The agent has to explore an unfamiliar game and work out its goal before it can act. At launch ARC Prize reported that humans score 100% and frontier AI 0.51%.

ARC Prize's ARC-AGI-3 data lists Gemini 3.1 Pro (Preview) at 0.42%. Google's newer Gemini 4 Argon is listed at 49.8% on its low setting. ARC-AGI-3 scores agents on what they learn while playing. Our read is that the February models could not carry that learning forward from step to step.

What does Gemini 3.1 Pro cost and how do you call it?

According to the Gemini API pricing page, 3.1 Pro costs $2 per million input tokens and $12 per million output tokens for prompts up to 200,000 tokens. Longer prompts cost $4 input and $18 output. Thinking tokens count as output. There is no free tier for this model.

The thinking docs list three thinking levels for 3.1 Pro (low, medium, high), with high as the default. The docs do not list an option to switch thinking off. Lower levels cut latency and the thinking tokens you pay for.

Google also publishes a second ID, gemini-3.1-pro-preview-customtools, on the same model page. The deprecations page shows that the Gemini 3 Pro preview was shut down on March 9, 2026, with 3.1 Pro as the replacement. No shutdown date has been announced for 3.1 Pro itself.

Is Gemini 3.1 Pro still worth using in October 2026?

For new projects, probably not as a default. On October 13, 2026, Google made Gemini 4 Argon available to developers through the API. Its introductory price is $2 input and $10 output per million tokens until November 30, 2026, then $4 and $20. The models list marks it stable with a 1M-token context window. ARC Prize lists Argon at 91.25% on ARC-AGI-2 for $0.32 per task., about a third of 3.1 Pro's cost per task.

3.1 Pro never left preview. The DeepMind page still shows its status as Preview and says Gemini 3.5 Pro is coming soon. If you already run 3.1 Pro in production and it works, there is no announced deadline to move. Keep in mind that preview models can change or be retired on short notice, as 3 Pro was.

Read the fine print

Some of the launch numbers are stronger than others. ARC-AGI-2 is verified by ARC Prize. Google's PDF says its Humanity's Last Exam and GPQA scores for 3.1 Pro are self-computed. It also says the scores for non-Gemini models come from each provider's self-reported numbers.

Claude Opus 4.6 is slightly ahead on SWE-Bench Verified (80.8% against 80.6%). It also leads Humanity's Last Exam with search and code (53.1% against 51.4%). On GDPval-AA, a test of office-style knowledge work, Sonnet 4.6 scores an Elo of 1633 against 3.1 Pro's 1317. Our read: 3.1 Pro was a large step in abstract reasoning and a modest one elsewhere.

The bottom line

Gemini 3.1 Pro's 77.1% on ARC-AGI-2 is real and independently verified, at under a dollar per task. Eight months later, Gemini 4 Argon beats it on the same test at a third of the cost per task. Use 3.1 Pro where it already works, and test Argon before starting anything new.

Does Gemini remember its reasoning after a conversation ends?

Within one conversation, partly. The Gemini API returns thought signatures, which its docs describe as encrypted representations of the model's internal reasoning. You send them back on the next turn, unchanged, so the model can continue its line of thought. Nothing in the docs says they carry over to a new conversation. For agents that is a real gap. An agent that has worked out how your codebase or data pipeline behaves starts from zero in the next session unless someone writes the conclusions down.

memU is a place to write them down. It keeps a shared wiki of readable Markdown that several agents can use, across sessions and devices. The agent decides what is worth remembering and writes it as memory or skill files. memU's memory service makes no LLM calls of its own. It stores and embeds what the agent wrote, then returns it on request.

A standing instruction in files like CLAUDE.md or AGENTS.md tells the agent to retrieve before answering. The notes are plain Markdown you can open and edit, and a different agent can read them later. More on that design is in why memU leaves the thinking to the agent.

To try the hosted version, get an API key from memu.so and send your agent one message:

Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU, API Key is YOUR_MEMU_API_KEY.

The list of supported agents and the self-hosted setup are in the memU GitHub repository.

FAQ

What is Gemini 3.1 Pro's ARC-AGI-2 score?

77.1%, verified by ARC Prize on the semi-private set at about $0.96 per task. Gemini 3 Pro scored 31.1% on the same test.

How does Gemini 3.1 Pro compare with Claude Opus 4.6?

In Google's February table, 3.1 Pro leads on ARC-AGI-2 (77.1% against 68.8%) and on Humanity's Last Exam without tools. Opus 4.6 is slightly ahead on SWE-Bench Verified and on Humanity's Last Exam with search and code. The Claude figures are Anthropic's self-reported numbers.

What is Gemini 3.1 Pro's context window?

The API accepts up to 1,048,576 input tokens and returns up to 65,536 output tokens per request.

How much does the Gemini 3.1 Pro API cost?

For prompts up to 200,000 tokens, it costs $2 per million input tokens and $12 per million output tokens. Longer prompts cost $4 input and $18 output. Thinking tokens are billed as output, and there is no free tier.

Is Gemini 3.1 Pro generally available?

No. As of October 2026 the API still lists it as a preview model under the ID gemini-3.1-pro-preview, with no shutdown date announced.