July 2026 Edition! Summary of the Latest LLM Coding Capability Comparison
A summary of the latest LLM coding capabilities as of July 2026, including model scores like Claude Fable 5, ideal VRAM sizes for local deployment, recommended hardware, and costs.
As of July 2026, the automatic coding capabilities of AI have undergone dramatic evolution. In particular, the score in "SWE-bench Pro," which measures the ability to solve software engineering tasks, has become a very important metric for evaluating the coding power of each LLM (Large Language Model). In this article, we compare the scores of major closed and open models, and further explain in detail the hardware configuration, costs, and break-even point against cloud APIs for running open models locally.
Latest LLM Coding Capability Comparison (SWE-bench Pro)
The scores and features of the major models at present are as follows. *Scores are evaluated on SWE-bench Pro| Rank | Model Name | Developer | Type (Required VRAM) |
Score | Features |
|---|---|---|---|---|---|
| 1 | Claude Fable 5 | Anthropic (USA) |
Closed | 80.0% | The absolute champion at present with overwhelming debugging capabilities. Cleared by the US government and now available in other countries. |
| 2 | Claude Mythos | Closed | 77.8% | An ultra-powerful reasoning model provided in advance. Highly specialized in security, available only to authorized corporations like banks and securities firms. | |
| 3 | Claude Opus 4.8 | Closed | 69.2% | The highest performance among available models when Fable 5 cannot be used. | |
| 4 | GPT-5.6 Sol | OpenAI (USA) |
Closed | 64.6% | The pinnacle within OpenAI. A reasoning model excellent at agentic behavior. |
| 5 | GLM-5.2 | Z.ai (China) |
Open (2TB) |
62.1% | Ranked 1st globally in open weights, not limited to Chinese models, closing in on closed models. |
| 6 | Claude Sonnet 5 | Anthropic (USA) |
Closed | 59.6% | An ultra-cost-effective model delivering equivalent coding power at less than 40% of the price of Opus 4.8. |
| 7 | Qwen3.7 Max | Alibaba (China) |
Open (2TB) |
60.6% | A highly versatile open model with various weight versions, chosen by many ventures as the base for their own AI. |
| 8 | MiniMax M3 | MiniMax (China) |
Open (1TB) |
59.0% | Just released in June 2026, strong in handling long context. A multimodal model capable of image and video input. |
| 9 | DeepSeek-V4 | DeepSeek (China) |
Open (1TB) |
55.4% | A model with excellent cost performance. |
Focus Point: While the top ranks are dominated by Anthropic's Claude family, it is noteworthy that Chinese open models (GLM-5.2 and Qwen3.7 Max) are closing in on powerful closed models like GPT-5.6 Sol.
Hardware Requirements for Running Large-Scale Local LLMs
Many open models have a massive number of parameters in the 1TB to 2TB class. We explain the "ideal VRAM size" and hardware configuration for running these at a practical speed (with quantization applied) in a local environment.Estimated Required VRAM Size
2TB Class (GLM-5.2, Qwen3.7 Max, etc.): Even with 4-bit quantization (INT4/AWQ, etc.), a minimum of 192GB or more of VRAM is required for loading the model and the context window (KV cache).1TB Class (MiniMax M3, DeepSeek-V4, etc.): Similarly, 128GB or more of VRAM is ideal with 4-bit quantization.
Summary of the Cheapest Recommended Hardware Configurations and Costs for Local LLMs
In the current GPU market, the cheapest way to build a large-capacity VRAM setup is arguably AMD right now. In the past, Mac Studio and others would have been candidates, but due to memory shortage issues, the current maximum memory limit is 96GB (or rather, this is the only option), and prices have risen by 30%, starting at a minimum of around 1.2 million JPY.| Model | CPU(TOPS) | VRAM (Memory) | Estimated Cost | Features & Evaluation |
| NVIDIA DGX Spark | 1000 TOPS (INT 4) | 128 GB (Unified Memory) | 800,000 JPY (1 unit) | Runs on Linux without a PC. Can be linked to a second unit to increase to 256GB. Low power consumption and excellent quietness. |
| NVIDIA A100 | 1248 TOPS (INT 4) | 80 GB (VRAM) | 4.3 Million JPY (1 unit) + PC cost | Looks like a graphics card, but since it can only perform AI processing, it has no image output terminals. Consumes an abnormal amount of power. Used only. |
| NVIDIA H100 | 3958 TOPS (INT 8) | 80 GB (VRAM) | 5.7 Million JPY (1 unit) + PC cost | Successor to the A100. Bandwidth has greatly increased, doubling the calculation speed. Subject to export restrictions to China. |
| RTX 3090 × 8 units + PC Build | 2504 TOPS (INT 8) | 192 GB (VRAM) | From 1.8 Million JPY (Including PC) | Requires a motherboard capable of seating 8 graphics cards. |
| AMD AI Max+ 395 | 50 TOPS | 128 GB (Unified Memory) | From 590,000 JPY | The strongest cost performance. Low power consumption and excellent quietness. |
| Apple Mac Studio M3 Ultra | 36 TOPS | 96 GB (Unified Memory) | From 1.16 Million JPY | Low power consumption and excellent quietness. |
Token Unit Price and Break-even Point When Using Closed Model APIs
Should you invest in a PC with such specs, or should you continue using cloud APIs (like Claude Opus or GPT-5.6) on a pay-as-you-go basis? We briefly estimate the break-even point. Currently, the typical API unit prices for high-performance closed models are around the following (assuming Sonnet 5 class as an example):
- Input: $3.00 / 1M tokens
- Output: $15.00 / 1M tokens
In coding tasks, a massive amount of context (such as existing codebases) is loaded, resulting in very high input tokens.
For example, if one exchange consumes Input: 50,000 tokens / Output: 2,000 tokens, the cost per exchange is about $0.18 (approx. 27 JPY).
Break-even calculation: Monthly cost of 27,700 JPY ÷ 27 JPY/exchange = approx. 1,025 inference requests per month. In other words, for heavy users or development teams executing coding tasks with massive codebases at least 34 times a day, purchasing hardware to run open models locally is calculated to be more cost-effective than continuously using cloud APIs (assuming the response accuracy of open models can adequately substitute for closed models).
In reality, everyone uses subscriptions, so the monthly token cost is much lower. If memory is abnormally expensive, honestly, it's doubtful whether even heavy users can recoup their investment up to the break-even point.