跪拜 Guibai
← All articles
Backend · Programmer · Artificial Intelligence

Kimi K3.1 Pro Strikes Back: 3D and Finance Benchmarks Beat GPT-6 Astra, Prices 80% Lower

By cxuanAI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A Chinese frontier model now matches or beats GPT-6 Astra and Claude Opus 5.5 on 3D and finance benchmarks while undercutting Astra's API pricing by about 80%, forcing Western teams to reconsider model choice and pricing for 3D, financial analysis, and long-context tasks.

Summary

After two months of OpenAI and Anthropic models topping leaderboards, Kimi's K3.1 Pro arrives as the company's new SOTA, supporting a 100M-token context. Benchmarks show it at 65.7% on Terminal-Bench 4.0 (GPT-6 Astra: 58.2%, Claude Opus 5.5: 66.4%), 64.4 on OfficeQA Pro v2 (ahead of both), 58.6 on Frontier Finance (behind only Fable 5.1), and 92.7 on BenchCAD with tools, within a few points of Opus 5.5 and Astra.

Hands-on tests centered on 3D generation. A single prompt for a photorealistic stratovolcano produced a closed-mesh model with PBR materials and 8K textures, which the tester says could be used as a 3A game asset; it outdid GLM-5.3-Flash and DeepSeek-V4.1-Flash on detail. The same model recreated The Wall and Winterfell from Game of Thrones with multiple camera views, and delivered a financial analysis of Porsche that the tester found the best laid out among compared models.

Pricing is the headline: 1.3 yuan per million input tokens on cache hit, 13 yuan on cache miss, 68 yuan per million output tokens. A highspeed variant runs 2-3x faster at 2x cost. Uncached input is 81.4% cheaper than GPT-6 Astra and output 80.6% cheaper; against Opus 5.5, both are roughly half. The tester concludes 3D and modeling work can be safely handed to K3.1 Pro.

Takeaways
— K3.1 Pro supports a 100M-token context.
— Terminal-Bench 4.0: 65.7% vs GPT-6 Astra 58.2% and Claude Opus 5.5 66.4%.
— OfficeQA Pro v2: 64.4 vs Opus 5.5 62.2 and Astra 54.4.
— Frontier Finance: 58.6, trailing Fable 5.1's 60.8 but ahead of Opus 5.5 and Astra.
— BenchCAD with tools: 92.7 vs Opus 5.5 96.2 and Astra 95.9.
— Per-million-token pricing: 1.3 yuan (cache hit), 13 yuan (cache miss) input, 68 yuan output; highspeed 2-3x speed at 2x price.
— Uncached input and output are about 81% and 80% lower than GPT-6 Astra respectively, and roughly half of Opus 5.5.
— Single-prompt volcano generation produced a closed, watertight mesh with PBR materials and 8K textures.
Conclusions

The margins against Opus 5.5 on Terminal-Bench and BenchCAD are under a point, so the 'SOTA' claim is benchmark-specific, not a blanket win.

An 80% price cut relative to Astra suggests Chinese labs are competing on cost as much as capability; that pressure could force Western API pricing down.

Single-prompt 3D scene generation with game-engine integration is a concrete workflow change; it removes multi-step tooling for producing explorable assets, though the tests are self-reported.

Concepts & terms
Terminal-Bench 4.0
A benchmark that evaluates coding agents on resource-constrained terminal tasks in a real environment, with strict dedicated tests for each task.
OfficeQA Pro v2
A benchmark that requires retrieving information from real financial documents, cross-checking, and performing multi-step calculations.
BenchCAD with tools
A benchmark for generating CAD code from multi-view images to reconstruct 3D shapes, scored by voxel overlap between the model and the target shape.
From the discussion
Featured comments
用户3234927109069

The quota is incredibly low.

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗