Cheapest model for coding agents

Per-token price lists hide what matters: agents are input-heavy and cache-heavy. Here's every model ranked by an actual monthly bill.

Data verified 2026-10-13 from official pricing pages.

Ranked by monthly cost

Workload: 3M input + 150K output tokens/day, 60% cache hit rate, 22 working days, standard list prices.

#	Model	Vendor	In / Cached / Out per 1M	Est./month
1	GPT-6 Luna	OpenAI	$0.1 / $0.01 / $0.5	$4.69
2	Gemini 3.5 Flash-Lite	Google	$0.3 / $0.03 / $2.5	$17
3	Gemini 3.8 Flash	Google	$0.75 / $0.075 / $3.75	$35
4	Grok Build 0.1	xAI	$1 / $0.2 / $2	$41
5	Claude Haiku 5.5	Anthropic	$0.5 / — / $2.5	$41
6	Grok 4.3	xAI	$1.25 / $0.2 / $2.5	$49
7	GPT-6.1 Sol	OpenAI	$2 / $0.1 / $10	$90
8	Gemini 4 Argon	Google	$2 / $0.1 / $10	$90
9	Grok 4.7	xAI	$2 / $0.5 / $6	$92
10	GPT-5.3 Codex	OpenAI	$1.75 / $0.175 / $14	$99
11	Gemini 3.1 Pro Preview	Google	$2 / $0.2 / $12	$100
12	Claude Sonnet 4.6 (legacy)	Anthropic	$3 / $0.3 / $15	$141
13	Claude Sonnet 5.5	Anthropic	$2 / — / $10	$165
14	Claude Opus 5.5	Anthropic	$4 / $0.2 / $20	$180
15	GPT-5.6 Sol	OpenAI	$4 / $0.4 / $20	$187
16	Claude Opus 5 (legacy)	Anthropic	$5 / $0.5 / $25	$234
17	GPT-6 Astra	OpenAI	$10 / $1 / $50	$469
18	Claude Fable 5.1	Anthropic	$10 / — / $50	$825

Change the workload in the calculator. Note scheduled changes: Gemini 4 Argon doubles Dec 1, 2026; Gemini 3.8 Flash doubles Jan 1, 2027.

FAQ

What is the cheapest model for coding agents?

At list prices with our default agent workload, GPT-6 Luna is cheapest at about $4.69/month, followed by Gemini 3.5 Flash-Lite ($17) and Gemini 3.8 Flash ($35).

Why does cache pricing matter so much?

Agents resend the same context every turn, so most input tokens can be cache hits. Cached input is often 10x cheaper than fresh input.

Should I use the cheapest model for everything?

Usually not. Route easy work (search, summaries, boilerplate) to cheap models and keep frontier models for planning and hard debugging.