ARC Prize verifies Gemini 3.7 Flash at 84.6% on ARC-AGI-2

• ARC Prize verified Gemini 3.7 Flash at 84.6% on ARC-AGI-2 for a reported $0.25 per task, according to arcprize's verified results.
• The model led Artificial Analysis' 80-task AnalystAgent board at 60%, with Claude Opus 5 at 53.8% and GPT-5.5 at 50.0% in WesRoth's benchmark chart.
• The ARC-AGI-2 cost result has become the headline trade-off, which DynamicWebPaige described as a Pareto-frontier point; Google's API prices use a separate per-token unit.
• ARC Prize is still running Gemini 3.7 Flash through ARC-AGI-3 and says it will publish the result once the evaluation is complete in arcprize's ARC-AGI-3 update.

Google's introductory API rate expires on December 31, when the published input and output rates double. Artificial Analysis also reports two deliberately different speed numbers: 1.3 minutes of weighted decode time, excluding first-token time and overhead, and 1.7 minutes of average Time per Task.

ARC Prize reported two verified result-and-cost pairs:

• ARC-AGI-2: 84.6% at $0.25 per task.
• ARC-AGI-1: 95.5% at $0.12 per task.

The verification label has a defined gate. Under ARC Prize's Verified Testing Policy, the program works with selected providers and tests public models from trusted sources. ARC-AGI-2 itself has a 120-task public evaluation set, alongside semi-private and private sets, on the official dataset page.

Artificial Analysis defines AA-AnalystAgent in its benchmark analysis as a test of answering complex questions about spreadsheets and documents. The score chart measures the share of 80 tasks solved on all five of five attempts in WesRoth's benchmark chart.

• Gemini 3.7 Flash, high: 60.0%.
• Claude Opus 5, max: 53.8%.
• GPT-5.5, xhigh: 50.0%.
• Claude Fable 5 with fallback: 48.8%; GPT-5.6 Sol, max: 47.5%, a gap also highlighted by gabriberton's benchmark summary.

ARC Prize's arcprize's comparison post added a direct 3.7-versus-3.6 view. Google's launch post supplies its first-party intergenerational deltas:

• FrontierCode 1.1 Main: 34.4% to 43.6%, up 9.2 points.
• DeepSWE v1.1: 49.0% to 65.3%, up 16.3 points.
• WebDev Arena: 1538 to 1588 Elo, up 50 Elo.
• GDP.pdf: 22.0% to 34.0%, up 12.0 points.
• AutomationBench: 17.0% to 30.4%, up 13.4 points.

ARC's $0.25 is a cost-per-task result, while Google's introductory API rate is $0.75 per million input tokens and $3.75 per million output tokens through December 31. The published standard rate from January 1 is $1.50 and $7.50, respectively. Artificial Analysis calculates $0.40 per Intelligence Index task at high reasoning, 30% below 3.6 Flash, and $0.26 at medium reasoning in its independent report.

The same report lists a one-million-token context window, unchanged from 3.6 Flash, and text, image, video, and speech inputs with text output. It records roughly 340 output tokens per second and a 1.7-minute average Time per Task at high reasoning.

The

labels the lower 1.3-minute figure as weighted average decode time, explicitly excluding time to first token and overhead.

One smaller 3-D coding measurement, as rohanpaul_ai's test summary described it, used Nous Research's Hermes Agent CLI through OpenRouter, identical prompts, and the same Three.js constraints across three voxel-city scenes:

• Muse Spark 1.2: $0.53 total cost.
• Gemini 3.7 Flash: $0.56.
• DeepSeek V4 Pro: $4.57.

Three-agent voxel-city cost comparison

One fofrAI reply to a commenter says simply, “Absolutely.” doodlestein supplied a more concrete scope, saying they would trust Gemini 3.7 Flash with web front-end work and other less important tasks in doodlestein's hands-on note.

An

says a plan's baseline quota would refresh on August 25 and that Antigravity does not support higher rate limits on that Google plan.

ARC Prize says it is currently running ARC-AGI-3 evaluations for Gemini 3.7 Flash and will publish the numbers when they are complete.