"Which model should I run my coding agent on?" almost always turns into a price question once the first invoice lands. Coding agents are token-hungry — they read whole files, reason across a repo, and emit long diffs — so the model you pick shows up on your bill in a big way.
Here's the part nobody tells you: for a coding workload, output price dominates. An agent that burns ~90M input and ~25M output tokens a month pays for those 25M output tokens at rates that swing from $0.28 to $30 per million. That single number decides most of your bill.
The same coding agent, priced across 15 models
Below is one fixed workload — ~90M input + ~25M output tokens/month (a busy single-developer coding agent) — priced against each model's current, official API rates. Nothing here is invented; every figure is pulled live from AI Model Watch, which tracks these prices daily from provider pricing pages.
| Model | Input $/M | Output $/M | Est. monthly cost |
|---|---|---|---|
| Qwen3.5-Flash | $0.10 | $0.40 | $19 |
| DeepSeek-V4-Flash | $0.14 | $0.28 | $20 |
| Codestral (v25.08) | $0.30 | $0.90 | $50 |
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 | $60 |
| DeepSeek-V4-Pro | $0.435 | $0.87 | $61 |
| Mistral Large 3 | $0.50 | $1.50 | $83 |
| Kimi K2.7 Code | $0.95 | $4.00 | $186 |
| Qwen3-Max | $1.20 | $6.00 | $258 |
| Grok 4.5 | $2.00 | $6.00 | $330 |
| Gemini 3.5 Flash | $1.50 | $9.00 | $360 |
| GPT-5.4 | $2.50 | $15.00 | $600 |
| GPT-5.6 Terra | $2.50 | $15.00 | $600 |
| Claude Sonnet 5 | $3.00 | $15.00 | $645 |
| Claude Opus 4.8 | $5.00 | $25.00 | $1,075 |
| GPT-5.6 Sol | $5.00 | $30.00 | $1,200 |
That's a 63× spread — $19/mo to $1,200/mo — for the same number of tokens. The choice of model, not the amount of work, is what moves the bill an order of magnitude.
Three things the table makes obvious
1. Output tokens are where coding agents bleed. Compare DeepSeek-V4-Flash ($0.28 out) to Claude Sonnet 5 ($15 out): a 54× output-price gap that a chat benchmark, which weights input heavily, would hide. Agents write a lot, so weight the output rate accordingly.
2. "Cheap" and "specialist" aren't the same axis. The two cheapest here are general-purpose small models (Qwen3.5-Flash, DeepSeek-V4-Flash), not the code-branded ones. Codestral and Kimi K2 Code are tuned for coding, but you pay for the tuning. Whether that tuning earns its 4–9× premium depends on your task — benchmark it on your repo, not on a leaderboard.
3. The frontier tier is a different budget entirely. Opus 4.8 and GPT-5.6 Sol land above $1,000/mo on this workload. They may well close the loop in fewer iterations — a frontier model that one-shots a task can be cheaper in practice than a cheap model that needs five tries. But that's an efficiency argument you have to verify, not assume.
The honest caveats
- These are token-price estimates, not your invoice. Real cost depends on how many tokens your agent actually burns, how often it retries, and whether you use prompt caching (which bills reused input at ~10× less — worth turning on for a static system prompt + tool defs).
- Preview pricing can move. DeepSeek V4-Flash/Pro are preview-tier; treat those two rows as provisional.
- Cheaper-per-token ≠ cheaper-per-task. A weaker model that loops more can cost more end to end. Price is the floor of the decision, not the whole of it.
Prices change — and coding models change fast
The numbers above are current as of publication, but this corner of the market moves weekly: new coding models ship, prices get cut, and preview tiers graduate or get retired. If you're running an agent in production, a 2× output-price change is a real budget event.
AI Model Watch tracks every LLM's price, context window and deprecation status daily from official sources, and sends a free email alert the moment a model you rely on changes price or gets an end-of-life date. If you'd rather not re-check a pricing page every week: aimodelwatch.dev.
Full ranked coding-cost table and methodology: aimodelwatch.dev/guides/cheapest-llm-for-coding
Top comments (0)