Google has introduced the Gemini 3.1 Pro preview, and the headline number is its 77.1% score on ARC-AGI-2, a benchmark focused on solving unfamiliar problems through reasoning. According to Google, that is more than double the result of the previous Gemini 3 Pro. Across 16 benchmarks listed by the company, Gemini 3.1 Pro placed first in 13.
Google also published several other benchmark results: 94.3% on GPQA Diamond, 80.6% on SWE-Bench Verified, 44.4% on Humanity’s Last Exam, and 92.6% on MMMLU. On MCP Atlas, which measures multi-step tool-use workflows, Gemini 3.1 Pro reached 69.2%, putting it nearly 10 percentage points ahead of Claude and GPT-5.2 in the comparison cited by the report.
Developers Can Now Choose the Model’s Reasoning Depth
A central addition in this release is the thinking level system. Developers can switch between low, medium, and high reasoning settings, adjusting how much inference budget the model uses for a task. Lighter API calls can be run with lower latency and lower cost, while harder debugging or problem-solving jobs can be pushed to a deeper setting. The idea is simple: let developers decide how much “thinking” they want to pay for.
At the high setting, the report says Gemini 3.1 Pro behaves like a smaller version of Google’s dedicated reasoning model, Gemini Deep Think. VentureBeat described it as an “on-demand Deep Think Mini.” On BrowseComp, a benchmark for autonomous web search by AI agents, the model jumped from 59.2% in the prior generation to 85.9%, pointing to a sharp improvement in web-based agent tasks.
Performance Rises While Token Prices Stay Flat
Google kept API pricing unchanged from Gemini 3 Pro: $2 per million input tokens and $12 per million output tokens. Based on the comparison in the source material, Gemini 3.1 Pro’s input cost is 60% lower than Claude Opus 4.6, while output cost is 52% lower.
The model retains a 1 million-token context window, described in the report as five times Claude’s and 2.5 times GPT-5’s. Output limits have been expanded to 65,000 tokens. Google also raised the single API upload cap from 20MB to 100MB, and added support for passing in a YouTube URL so the model can process video content directly. The report links this pricing stance to Google’s structural cost advantage from in-house TPU chips and its cloud infrastructure.
Competition Remains Split Across Reasoning, Safety, and Coding
Gemini 3.1 Pro does not lead every category. The report notes that Claude Sonnet 4.6 in Thinking Max mode matches Gemini 3.1 Pro on the MRCR v2 long-context memory test. On GDPval-AA Elo expert tasks, Claude posts 1633, ahead of Gemini 3.1 Pro’s 1317.
OpenAI’s GPT-5.3-Codex leads on Terminal-Bench 2.0 for terminal programming, scoring 77.3% against Gemini 3.1 Pro’s 68.5%. The report also says Claude models show hallucination rates of about 3%, lower than the roughly 6% average cited for Gemini and GPT. Taken together, the benchmark picture points to Google pushing hard in reasoning and agent workloads, Anthropic holding ground in precision and safety, and OpenAI keeping an edge in code generation and ecosystem reach.