Answers
What's the cheapest API for Cursor?
The cheapest coding-capable API you can plug into Cursor today is MiniMax-M2.7 at $0.18 input / $0.72 output per 1M tokens ($0.288/1M blended at a 4:1 split) through an OpenAI-compatible endpoint. If you want Claude-class quality, Kimi-K2.6 — 80.2 on SWE-Bench Verified against Claude Opus 4.6's 80.8 on the same published harness — costs $0.35 / $1.46 per 1M. Cursor accepts either by overriding the OpenAI base URL; no plugin needed.
| Model | Input / 1M | Output / 1M | Blended (4:1) | vs cheapest |
|---|---|---|---|---|
| MiniMax-M2.7via Mingles Router | $0.18 | $0.72 | $0.288 | — |
| Kimi-K2.6via Mingles Router | $0.35 | $1.46 | $0.572 | 2.0× |
| Claude Sonnet 5 | $2.00 | $10.00 | $3.60 | 12.5× |
| GPT-5.4 | $2.50 | $15.00 | $5.00 | 17.4× |
Vendor list prices; sources: Anthropic — Claude pricing, OpenRouter — models & pricing, Moonshot AI — Kimi-K2.6 model card. Prices verified August 14, 2026.
Set it up in Cursor (60 seconds)
- Get a free key at router.mingles.ai/app — no card.
- Cursor → Settings → Models → paste the key as your OpenAI API key.
- Override the OpenAI base URL with
https://router.mingles.ai/v1. - Add a custom model:
moonshotai/Kimi-K2.6(orMiniMaxAI/MiniMax-M2.7for the cheapest runs).
Full walkthrough with screenshots on the agents hub.
The honest caveats
- Cursor Pro bundles editor features we don't replace — a cheap key replaces the model bill, not the subscription's tab-completion models.
- Frontier models keep a lead of a couple of points on hard reasoning benchmarks (AIME, GPQA). For agent loops and code edits the published gap is within one point.
- Two main models, not three hundred. If you need vision or a specific frontier model, use both endpoints side by side.
FAQ
▸Does Cursor work with a custom OpenAI-compatible API?
Yes. In Cursor: Settings → Models → add your API key and override the OpenAI base URL with https://router.mingles.ai/v1, then add the model id (e.g. moonshotai/Kimi-K2.6). Cursor treats it like any OpenAI endpoint — chat, inline edits and agent mode all work.
▸Is a cheap model good enough for coding in Cursor?
Kimi-K2.6 scores 80.2 on SWE-Bench Verified on its published model card — Claude Opus 4.6 scores 80.8 under the same harness. For routine completions MiniMax-M2.7 is about half the price again. The honest gap is hard reasoning, where frontier models keep a couple of points.
▸Is this cheaper than Cursor Pro?
Different thing: Cursor Pro ($20/mo) bundles the editor features with model credits. Bringing your own key means you pay for exactly the tokens you use — the Mini plan is $9/mo for an allowance worth ~220M DeepSeek-V4-Flash tokens — but heavy Claude/GPT usage inside Cursor Pro can also hit its own caps. Many users run both: Pro for the editor, a cheap key for the heavy agent loops.
▸Is there a free tier to test with?
Yes — a free API key with a weekly allowance (~3.9M tokens on DeepSeek-V4-Flash), no card required. Enough to run Cursor for an evening and decide with your own code.