闭源模型为何在性价比前沿反超开源
论证要点
- Village 的 LLM Atlas 按价格档位追踪最强模型：目前开源权重模型仅在约 3% 的预算档位是最佳选择，且集中在最便宜的一端，由 Google Gemma 4 E4B 和小米 MiMo-V2.6-Flash 占据。
- 今年 6 月开源模型曾覆盖 41% 的预算档位，DeepSeek V4 Flash 单独拿下近三分之一，但 OpenAI 的 GPT-6 Luna 和 GPT-6.1 Terra 以及 Anthropic 的 Claude Haiku 5.5 夺回了大部分份额。
- 开源模型历来靠单次发布冲高：DeepSeek-V3.2-Exp 在 2025 年 9 月发布当周拿下近半预算档位，但每次峰值后下一季度份额都腰斩或更多，因为闭源实验室推出了更便宜的层级。
- 作者给出三种解释：闭源实验室可用自家前沿模型蒸馏出廉价版本；闭源厂商一次发布覆盖全价格带的模型家族（GPT-6 系列占前沿 21 个点中的 11 个）；闭源厂商推理规模更大、单位任务成本可能更低。
- 研究方法：用公开基准估算能力分，用独立成本模型估算每任务成本，在 1% 到 GPT-5 成本三倍的 201 个对数分布预算档位上找最强模型，数据截至 2026 年 10 月 31 日。
- 作者承认成本按托管 API 计算，开源模型自托管可能更便宜；但限制基准覆盖或收窄价格区间后结论不变，放宽到极低价位时开源份额升至 16%，因微型开源模型仍赢在最便宜端。
对实践的启示
做 AI 应用选型时，不能默认开源更便宜：当前闭源厂商的廉价层级在多数价格档位性价比更高。开源模型的性价比窗口往往短暂，依赖单次发布，需持续跟踪价格前沿变化，并考虑自托管成本是否真能抵消闭源 API 的优势。
作者承认的反方与局限
成本反映的是评估者通过托管 API 支付的价格，开源模型在自己的硬件上运行可能更便宜；此外，将预算范围放宽到极低价位时，开源模型份额会升至 16%，因为微型开源模型仍在最便宜端胜出。
原文摘引
Closed models hold almost every budget
Open models win on value in bursts: a strong, cheap release, usually from DeepSeek, takes a wide band of budgets until the next closed model undercuts it.
The latest closed releases have been both cheaper and stronger than anything open, so the band has shrunk to the very bottom of the price range.
▸ 展开原文全文（19 段，抓取自 Hacker News）
Closed AI models now dominate the price-performance frontier
Village’s LLM Atlas maps the price-performance frontier: the strongest AI model at every price point.
Closed models hold almost every budget
- An open-weight model is the best buy at 3% of budgets. Nearly all of those are at the cheap end, held by Google’s Gemma 4 E4B and Xiaomi’s MiMo-V2.6-Flash.
- In June, it was 41%. DeepSeek V4 Flash alone was the best buy across nearly a third of budgets, with Zhipu’s GLM-5.2 just above it.
- Two OpenAI releases took most of that back. GPT-6 Luna in September and GPT-6.1 Terra in October now cover the budgets open models used to win, alongside Anthropic’s Claude Haiku 5.5.
Open models have led on value before
- Today’s share is the lowest we have measured. At the end of every quarter since 2024, open models held at least 6% of budgets.
- The peaks came from single releases. DeepSeek-V3.2-Exp was the best buy at nearly half of all budgets the week it launched in September 2025.
- None of them lasted. After each peak, the open share fell by about half or more by the next quarter as closed labs shipped cheaper tiers.
Open models win on value in bursts: a strong, cheap release, usually from DeepSeek, takes a wide band of budgets until the next closed model undercuts it. The latest closed releases have been both cheaper and stronger than anything open, so the band has shrunk to the very bottom of the price range.
Why closed labs may be winning on value
Three explanations fit the pattern:
- Closed labs can distill their own frontier models. A lab with the strongest model can train cheaper ones on its outputs in-house, and today’s cheap closed tiers followed their flagships within weeks, as Claude Haiku 5.5 did two weeks after Claude Opus 5.5. Open labs have to start from their own, weaker models, since closed labs’ terms generally bar training competitors on their outputs.
- Closed labs release whole families. OpenAI’s GPT-6 and GPT-6.1 models, from Luna and Terra up to Sol and Astra, hold 11 of the frontier’s 21 points between them, so one generation covers nearly every price.
- Closed labs may serve more cheaply. OpenAI, Anthropic and Google run inference at enormous scale, Google on its own chips, which can lower what each task costs; most open-model labs serve far less traffic.
How we built this
This study uses Village’s LLM Atlas, which estimates a capability score for each model and reasoning setting from public benchmark results and, with a separate cost model, its cost per task from evaluators’ runs, described in its methodology. Costs are relative to GPT-5 at high effort. For each date, we take every model released by then and find the most capable one at each of 201 budgets spread evenly on a log scale from 1% to three times GPT-5’s cost. The scores are from October 31, 2026; Atlas updates as new results arrive.
Costs reflect what evaluators paid through hosted APIs, and open models can be cheaper to run on your own hardware. Restricting to models scored on at least 10 benchmark families, or narrowing the budgets to 3% to one times GPT-5’s cost, gives the same pattern; widening them to 0.3% to ten times raises today’s share to 16%, because tiny open models still win at the very cheapest end.