Which AI model gives the most coding ability per dollar in October 2026?

In our calculation DeepSeek V4.1 Flash gives the most agentic-coding ability per dollar: 40 Terminal-Bench 4.0 points (from Artificial Analysis) per dollar of 100 coding-agent requests (our measured cost), followed by Claude Haiku 5.5 (25) and Step 5 Preview (13).

Source: Artificial Analysis (artificialanalysis.ai). Fetched October 13, 2026. Cost per request: Standard Compute gateway data, 79,301 coding-agent requests, Sep 13 to Oct 12, 2026. How we compute this

Run current Claude, GPT, Gemini and open models in your agent from $19/month

See plans

14 current models ranked by Terminal-Bench 4.0 points per dollar of 100 requests (our calculation)

Model	Points per $our calculation	Terminal-Bench 4.0Artificial Analysis	100 requestsour data
1. DeepSeek V4.1 Flashmax effort	40	26.8%	$0.67
2. Claude Haiku 5.5max effort	25	32.8%	$1.31
3. Step 5 Preview	13	33.3%	$2.48
4. Claude Sonnet 5.5max effort	12	63.6%	$5.33
5. GPT-6.1 Solmax effort	8.3	56.1%	$6.74
6. GLM 5.3max effort	7.3	41.9%	$5.72
7. Mistral Large 4Mistral Large 4 Preview	6.3	26.8%	$4.24
8. Claude Opus 5.5max effort	5.6	59.6%	$10.66
9. Gemini 4 Argonhigh effort	5.4	57.1%	$10.55
10. Muse Wondermax effort	4.9	52.0%	$10.55
11. GPT-5.6 Terramax effort	4.2	35.4%	$8.40
12. Claude Fable 5.1max effort	2.2	52.0%	$23.76
13. Grok 4.7xhigh effort	2.0	25.8%	$13.16
14. GPT-6 Astramax effort	1.4	59.1%	$41.25

Source: Artificial Analysis (artificialanalysis.ai). Each model is measured as the variant named under it. Cost per request is ours: every one of 79,301 real coding-agent requests priced at list rates (OpenRouter's listed price for some older and open models). Methodology

Not ranked

Model	Points per $our calculation	Terminal-Bench 4.0Artificial Analysis	100 requestsour data
GPT-6 Lunapasses 12.6% of Terminal-Bench 4.0 tasks	31	12.6%	$0.41
GPT-5.6 Lunasuperseded by GPT-6 Luna	14	11.6%	$0.84
GPT-6 Solsuperseded by GPT-6.1 Sol	5.3	43.9%	$8.25
DeepSeek V4 Propasses 14.1% of Terminal-Bench 4.0 tasks	4.6	14.1%	$3.04
Gemini 3.8 Flashpasses 19.7% of Terminal-Bench 4.0 tasks	4.1	19.7%	$4.83
GPT-5.6 Solsuperseded by GPT-6.1 Sol	2.4	39.9%	$16.50
MiniMax M3passes 2.0% of Terminal-Bench 4.0 tasks	1.6	2.0%	$1.29
Claude Opus 4.8superseded by Claude Opus 5.5	1.3	21.7%	$16.23
Gemini 3.5 Flashsuperseded by Gemini 3.8 Flash	1.3	6.6%	$4.93
Claude Fable 5superseded by Claude Fable 5.1	1.3	42.4%	$32.47
GPT-5.4 Minipasses 2.0% of Terminal-Bench 4.0 tasks	0.9	2.0%	$2.27
Qwen3.7 Pluspasses 1.0% of Terminal-Bench 4.0 tasks	0.7	1.0%	$1.39
GPT-5.5superseded by GPT-6.1 Sol	0.7	14.6%	$20.99
Gemini 3.1 Flash Litepasses 0.5% of Terminal-Bench 4.0 tasks	0.6	0.5%	$0.82
Gemini 3.1 Propasses 4.0% of Terminal-Bench 4.0 tasks	0.4	4.0%	$9.48
Kimi K2.7 Codepasses 1.0% of Terminal-Bench 4.0 tasks	0.4	1.0%	$2.71
Claude Sonnet 4.6superseded by Claude Sonnet 5.5	0.3	3.0%	$9.74
Qwen3.7 Maxpasses 1.5% of Terminal-Bench 4.0 tasks	0.2	1.5%	$6.29
Kimi K2.6passes 0.5% of Terminal-Bench 4.0 tasks	0.2	0.5%	$2.41
GLM 5.2superseded by GLM 5.3	0.2	1.0%	$5.72
Llama 4 Maverickpasses 0.0% of Terminal-Bench 4.0 tasks	0.0	0.0%	$0.95
Devstral 2passes 0.0% of Terminal-Bench 4.0 tasks	0.0	0.0%	$1.01
Mistral Large 3superseded by Mistral Large 4	0.0	0.0%	$1.19
Claude Haiku 4.5superseded by Claude Haiku 5.5	0.0	0.0%	$2.22
Grok 4.3superseded by Grok 4.7	0.0	0.0%	$6.84

Source: Artificial Analysis (artificialanalysis.ai). Cost per request is ours. Methodology

FAQ

How is the value score calculated?

Our calculation: the model's Terminal-Bench 4.0 score from Artificial Analysis, in percentage points, divided by what 100 average coding-agent requests cost at list price on our measured traffic (79,301 requests, Sep 13 to Oct 12, 2026). A model passing 50% of tasks at $5 per 100 requests scores 10.

Why are some cheap models not ranked?

A very low score divided by a very low price can top a per-dollar list without finishing the work, so models that pass fewer than one in five Terminal-Bench 4.0 tasks are listed apart. So are models whose context window cannot take every request in our sample, because pricing only the requests that fit would flatter them.

Is the best-value model the cheapest per finished task?

Not necessarily. Points per dollar is a rough guide. A stronger model can finish a task in fewer steps, so test the top two or three on a task from your own repository.

Current Claude, GPT, Gemini and open models are included with Standard Compute

Standard Compute

Current frontier models Claude Opus 5.5 · Claude Sonnet 5.5