Cost per task: agentic coding session
One long tool-using loop: 60,000 tokens in across turns, 4,000 tokens out. Cost = input tokens × input price + output tokens × output price. “With caching” assumes three-quarters of the input is served from the prompt cache at the provider's cache-read price.
As of 21 Oct 2026, BenchLeader ranks Gemma 4 E4B first on the Cost per task: agentic coding session board at $0.0016, ahead of Granite 4.2 3B ($0.0023) and LFM2 24B A2B ($0.0023).
284 models · data as of 21 Oct 2026
- 1Gemma 4 E4B (thinking)$0.0016
- 2Granite 4.2 3B$0.0023
- 3LFM2 24B A2B$0.0023
- 4Qwen3.5-4B (thinking)$0.0024
- 5Llama 3.2 1B Instruct$0.0024
- 6Gemma 4 E2B (thinking)$0.0027
- 7Llama 3.1 8B$0.0027
- 8Granite 3.3 8B (none)$0.0028
- 9Qwen3.5 Flash$0.0029
- 10NVIDIA Nemotron Nano 9B v2 (thinking)$0.0030
- 11Nova Micro$0.0030
- 12Sarvam 105B (high)$0.0031
- 13Mistral Small 3$0.0033
- 14Gemma 3 4B$0.0034
- 15Gemma 3 12B$0.0036
284 of 284
| # | ||||
|---|---|---|---|---|
| 1 |  GPT-5.5 Proxhigh | $2.52 | – | 65.3 |
| 2 | GPT-5.4 Pro | $2.52 | – | 63.5 |
| 3 | GPT-4 | $2.04 | – | 43.4 |
| 4 | GPT 4 0613 | $2.04 | – | 39.8 |
| 5 | o3-prohigh | $1.52 | – | 53.9 |
| 6 | GPT-5 Pro | $1.38 | – | 59.8 |
| 7 |  Claude Opus 4.1thinking | $1.20 | $0.593 | 56.5 |
| 8 | Claude Opus 4thinking | $1.20 | $0.593 | 55.1 |
| 9 | Claude 3 Opus | $1.20 | – | 39.2 |
| 10 | o1 | $1.14 | $0.802 | 53.1 |
| 11 | GPT-6 Astramax | $0.800 | $0.395 | 70.6 |
| 12 | Claude Fable 5.1max | $0.800 | $0.361 | 70.6 |
| 13 | Claude Fable 5max | $0.800 | $0.395 | 69.4 |
| 14 | GPT-4 Turbo | $0.720 | – | 41.0 |
| 15 | GPT-5.5xhigh | $0.420 | $0.217 | 66.1 |
| 16 | GPT-5.5 Instant | $0.420 | – | 58.1 |
| 17 | Claude Opus 5max | $0.400 | $0.198 | 68.9 |
| 18 | Claude Opus 4.7 | $0.400 | $0.198 | 63.4 |
| 19 | Claude Opus 4.8max | $0.400 | $0.198 | 63.4 |
| 20 | Claude Opus 4.6high | $0.400 | $0.198 | 62.9 |
| 21 | Claude Opus 4.5thinking | $0.400 | $0.198 | 60.0 |
| 22 | Claude Opus 5.5max | $0.320 | $0.149 | 72.7 |
| 23 | GPT-5.6 Solmax | $0.320 | $0.158 | 68.0 |
| 24 |  Muse Wondermax | $0.320 | $0.149 | 67.5 |
| 25 |  Mistral Large 24.02 | $0.288 | – | 40.3 |
| 26 |  Kimi K3maxopen ↗ | $0.240 | $0.118 | 64.9 |
| 27 | Claude Sonnet 4.5high | $0.240 | $0.118 | 58.3 |
| 28 |  Grok 4 | $0.240 | – | 58.0 |
| 29 | Claude Sonnet 4.6 | $0.240 | $0.118 | 57.1 |
| 30 | Claude Sonnet 4thinking | $0.240 | $0.118 | 53.3 |
| 31 | Grok 3 | $0.240 | – | 51.5 |
| 32 | Claude 3.5 Sonnet | $0.240 | – | 46.3 |
| 33 | GPT-5.4xhigh | $0.210 | $0.109 | 64.8 |
| 34 |  Nova Premier | $0.200 | $0.116 | 45.3 |
| 35 | GPT-4o | $0.190 | $0.134 | 44.0 |
| 36 |  Command Aopen ↗ | $0.190 | – | 41.9 |
| 37 |  Qwen3 7max | $0.180 | $0.082 | 60.7 |
| 38 | GPT-5.6 Terramax | $0.168 | $0.087 | 64.4 |
| 39 |  Gemini 3.1 Pro | $0.168 | $0.087 | 60.7 |
| 40 | GPT-5.3-Codexxhigh | $0.161 | $0.090 | 63.8 |
| 41 | GPT-5.2xhigh | $0.161 | $0.090 | 60.7 |
| 42 | GPT-5.2-Codexxhigh | $0.161 | $0.090 | 59.4 |
| 43 | Gemini 4 Argonhigh | $0.160 | $0.074 | 71.0 |
| 44 | GPT-6.1 Solmax | $0.160 | $0.074 | 70.0 |
| 45 | Claude Sonnet 5.5max | $0.160 | $0.074 | 69.5 |
| 46 | GPT-6 Solmax | $0.160 | $0.079 | 66.6 |
| 47 | Claude Sonnet 5high | $0.160 | $0.079 | 60.3 |
| 48 | o3 | $0.152 | $0.085 | 59.5 |
| 49 | GPT-4.1 | $0.152 | $0.085 | 48.4 |
| 50 |  Jamba 1.5 Largeopen | $0.152 | – | 42.3 |
| 51 | Jamba Large 1.6 | $0.152 | – | 26.0 |
| 52 | Grok 4.6medium | $0.144 | $0.076 | 64.2 |
| 53 | Grok 4.7high | $0.144 | $0.076 | 64.0 |
| 54 | Qwen3.8 Maxmax | $0.144 | $0.064 | 63.3 |
| 55 | Qwen3.8 2.4T A95Bopen ↗ | $0.144 | $0.065 | 63.2 |
| 56 | Grok 4.5high | $0.144 | $0.068 | 59.3 |
| 57 | Mistral Large 2.1open | $0.144 | – | 42.5 |
| 58 | Pixtral Largeopen | $0.144 | – | 41.9 |
| 59 | Mistral Large 2open | $0.144 | – | 40.8 |
| 60 | Mixtral 8x22B Instructopen ↗ | $0.144 | – | 37.5 |
Cite as: BenchLeader, “Cost per task: agentic coding session”, https://www.benchleader.com/leaderboards/cost-agent-session, data as of 21 Oct 2026.