Low-cost and high-efficiency workhorse models, October edition
This is the October 2026 edition of our benchmark of low-cost and high-efficiency models: 13 models measured on the latest workloads data available from Metergraph.
This article is for readers choosing a model for well-structured agentic workflows, where cost and latency are as important as quality.
Results
Best default: GPT-6 Luna. It has both the top quality and almost the lowest cost.
Best for latency: GLM 5.3 Flash. It was the fastest model within the top-quality band.
| Model | Cost | Quality | Latency | Quality / cost | Quality / latency |
|---|---|---|---|---|---|
 GPT-6 Luna | 0.43x | 0.993 | 0.60x | 2.31 | 1.66 |
| GPT-5.6 Luna | 0.86x | 0.988 | 0.64x | 1.15 | 1.54 |
 Qwen 3.8 Flash | 1.22x | 0.985 | 3.71x | 0.81 | 0.27 |
 Gemini 3.5 Flash-Lite | 1.32x | 0.984 | 0.35x | 0.75 | 2.80 |
 Grok 4.1 Fast Reasoning | 1.89x | 0.982 | 1.39x | 0.52 | 0.71 |
 GLM 5.3 Flash | 0.54x | 0.976 | 0.31x | 1.79 | 3.18 |
 DeepSeek V4 Flash | 0.86x | 0.968 | 2.50x | 1.13 | 0.39 |
| DeepSeek V4.1 Flash | 2.11x | 0.962 | 0.67x | 0.46 | 1.43 |
 MiniMax M3 | 1.22x | 0.953 | 0.64x | 0.78 | 1.49 |
 Claude Haiku 4.5 | 4.41x | 0.845 | 0.43x | 0.19 | 1.96 |
 Ministral 3 14B | 0.67x | 0.835 | 1.13x | 1.25 | 0.74 |
| Gemma 4 31B | 0.42x | 0.741 | 6.06x | 1.77 | 0.12 |
 Mercury 2.5 | 0.60x | 0.734 | 1.46x | 1.23 | 0.50 |
What changed in this edition
- New in this edition: GPT-6 Luna, Qwen 3.8 Flash, Gemini 3.5 Flash-Lite, GLM 5.3 Flash, DeepSeek V4.1 Flash, Claude Haiku 4.5, Ministral 3 14B and Mercury 2.5.
- Carried over: GPT-5.6 Luna, Grok 4.1 Fast Reasoning, DeepSeek V4 Flash, MiniMax M3 and Gemma 4 31B.
- No longer included: Kimi K2.5, Gemma 4 26B A4B, Nemotron 3 Super, Grok 4.1 Fast and Claude 3 Haiku.
- New workloads: Ingredient catalogue matching.
- 8 cases per workload instead of 6, and every model is scored on the same cases.
- Exact scores are not comparable with the September article, as test data and rubrics were updated.
Workloads
4 workloads from live traffic in 4 production applications. In each, a model receives a prompt with data in it and returns JSON that a program consumes.
- Recommendation normalization. Convert a product recommendation into a strict JSON record: canonical brand name, rank, justification and URL.
- Event extraction. Read a scraped listing page and return its upcoming public events as a JSON array, or an empty array when there are none.
- Company summarization. Write a short summary of a company from enrichment data: market, customers and positioning.
- Ingredient catalogue matching. Given a product ingredient and candidate entries from a nutrition database, return the entry that is the same food, or null.
16 samples per model per workload. Differences of a few hundredths in quality are noise. All workloads are structured output. The results say nothing about prose, conversation or tool use.
Method
Each workload is a fixed set of production traces. Every model replays the same 8 cases twice, which gives 16 scored samples per model per workload. A judge model (Claude Sonnet 5), which is not a candidate, grades each response against a rubric derived from the workload's own instructions. A clean response scores 1.0, a minor issue 0.5, a major issue 0. A response that does not match the workload's required output format scores 0.
Quality is an absolute score from 0 to 1, averaged across workloads. Cost and latency are relative to the field: 0.50x means half of what the typical model here cost, or took, for the same requests. The typical model is the geometric mean of all models on that workload. No baseline model enters the comparison.
All models were called through Vercel AI Gateway, on 29 and 30 September 2026, one workload at a time.
We repeat this benchmark as new models are released. The latest edition is always at metergraph.dev/blog/workhorse-tier.
Earlier editions: September 2026.
Questions or suggestions: vasiliy@psl.com