Cheapest AI to run a 50-step AI agent (Oct 2026)
The cheapest model overall is Ling 3.0 Flash at 1.7¢ per task. If you want a top-quartile model, the cheapest is DeepSeek V4.1 Flash at 2.4¢.
10 cheapest models
Any model with a price, including small and older ones.
| # | Model | Cost per task | WebDev score | Tokens in / out | Basis | |
|---|---|---|---|---|---|---|
| 1 | I Ling 3.0 FlashInclusionai | 1.7¢$16.91 / 1,000 | - | 2,257,637 / 95,984 | modeled | |
| 2 | N Nex-N2.5-MiniNex Agi | 1.8¢$17.66 / 1,000 | - | 2,235,262 / 102,435 | modeled | |
| 3 |  GPT-5 NanoOpenAI | 2.1¢$20.85 / 1,000 | 1338 | 2,237,500 / 15,000 | modeled best value | |
| 4 |  Muse Spark 1.3 ContributorMeta | 2.1¢$21.24 / 1,000 | - | 2,239,737 / 43,814 | modeled | |
| 5 | Muse Spark 1.2 ContributorMeta | 2.1¢$21.24 / 1,000 | - | 2,239,737 / 43,814 | modeled | |
| 6 | gpt-oss-20bOpenAI | 2.2¢$22.22 / 1,000 | 1317 | 2,237,500 / 15,000 | modeled | |
| 7 |  DeepSeek V4.1 FlashDeepSeek | 2.4¢$24.06 / 1,000 | 1618 | 2,235,262 / 102,435 | modeled best value | |
| 8 |  GLM 5.3 FlashZ.ai | 2.6¢$25.63 / 1,000 | 1606 | 2,239,737 / 43,814 | modeled | |
| 9 | I Mercury 2.5Inception | 2.7¢$26.84 / 1,000 | - | 2,172,612 / 102,015 | modeled | |
| 10 |  Qwen3.7 FlashAlibaba Qwen | 2.8¢$28.02 / 1,000 | - | 2,259,875 / 96,000 | modeled |
Cheapest high-quality models
Only models in the top 25% of LMArena WebDev scores (≥ 1513).
| # | Model | Cost per task | WebDev score | Tokens in / out | Basis | |
|---|---|---|---|---|---|---|
| 1 | DeepSeek V4.1 FlashDeepSeek | 2.4¢$24.06 / 1,000 | 1618 | 2,235,262 / 102,435 | modeled best value | |
| 2 | GLM 5.3 FlashZ.ai | 2.6¢$25.63 / 1,000 | 1606 | 2,239,737 / 43,814 | modeled | |
| 3 | GPT-6 LunaOpenAI | 4.1¢$40.71 / 1,000 | 1578 | 2,237,500 / 22,000 | modeled | |
| 4 |  MiMo-V2.6-FlashXiaomi | 4.4¢$44.33 / 1,000 | 1634 | 2,241,975 / 95,880 | modeled best value | |
| 5 | MiMo-V2.6-ProXiaomi | 6.5¢$64.51 / 1,000 | 1625 | 2,241,975 / 24,380 | modeled | |
| 6 | DeepSeek V4 Flash 0423DeepSeek | 6.5¢$65.25 / 1,000 | 1581 | 2,235,262 / 102,435 | modeled | |
| 7 |  Claude Haiku 5.5Anthropic | 6.6¢$66.35 / 1,000 | 1585 | 3,101,175 / 50,340 | modeled | |
| 8 | GPT-5.6 LunaOpenAI | 8.6¢$85.82 / 1,000 | 1519 | 2,237,500 / 22,000 | modeled | |
| 9 | DeepSeek V4 Pro 0423DeepSeek | 9.7¢$97.25 / 1,000 | 1582 | 2,235,262 / 102,435 | modeled | |
| 10 | Qwen3.8 27BAlibaba Qwen | 34¢$340 / 1,000 | 1593 | 2,259,875 / 96,000 | modeled |
Cheapest model of each lab
| Model | Cost per task | Score |
|---|---|---|
GPT-5 NanoOpenAI | 2.1¢$20.85 / 1,000 | 1338 |
DeepSeek V4.1 FlashDeepSeek | 2.4¢$24.06 / 1,000 | 1618 |
GLM 5.3 FlashZ.ai | 2.6¢$25.63 / 1,000 | 1606 |
Qwen3.7 FlashAlibaba Qwen | 2.8¢$28.02 / 1,000 | - |
MiMo-V2.6-FlashXiaomi | 4.4¢$44.33 / 1,000 | 1634 |
 Ministral 3 3B 2512Mistral | 4.5¢$45.38 / 1,000 | - |
 Gemini 2.5 Flash LiteGoogle | 5.4¢$53.89 / 1,000 | 1375 |
Llama 3.1 8B InstructMeta | 6.6¢$65.76 / 1,000 | 1211 |
Claude Haiku 5.5Anthropic | 6.6¢$66.35 / 1,000 | 1585 |
 MiniMax M2.5MiniMax | 11¢$109 / 1,000 | 1386 |
 Kimi K2.5Moonshot AI | 43¢$429 / 1,000 | 1437 |
 Grok Build 0.1xAI | 67¢$674 / 1,000 | - |
FAQ
What is the cheapest AI to run a 50-step AI agent?
The cheapest model overall is Ling 3.0 Flash at 1.7¢ per task. If you want a top-quartile model, the cheapest is DeepSeek V4.1 Flash at 2.4¢.
Is the cheapest model good enough?
Often not for hard tasks. That is why this page also lists the cheapest models in the top quarter of LMArena scores and the cheapest model of every major lab.