Cheapest AI API for coding in 2026, by quality level

What one coding agent task costs (our estimate)

Coding agents resend your files and the conversation on every step, so most of what you pay for is input, and most of that input repeats. Providers charge far less for repeated input that hits their prompt cache. That makes the cached-input price the number to watch for agent work, more than the headline input price.

To compare like for like, we priced one long agent task: 100K fresh input tokens, 900K cached input tokens and 50K output tokens. These inputs are our assumption, not a measurement, and real tasks vary a lot. Cache-write fees are left out: Anthropic charges more than the normal input price to write its cache (Kimi does for its 1-hour cache), so those models cost a little more than shown.

On those inputs, MiMo-V2.6 Flash is cheapest at about $0.031 per task, and Claude Fable 5.2, the top model on Code Arena, costs about $2.98. That is roughly 98x more for the best score on the board.

Qwen3.8 Max and Qwen3.8 27B have no public cache-hit price, so their estimate assumes no cache discount.

Which cheap coding API should you pick?

• Lowest bill that still codes well: MiMo-V2.6 Flash is #20 on Code Arena at $0.17 per 1M tokens blended. Good for bulk edits, tests and quick fixes.
• Best quality under $1: DeepSeek V4.1 Flash is #15 at $0.53 blended, with a cached-input price of $0.006. DeepSeek says off-peak use is half price, see DeepSeek off-peak hours.
• Cheapest top 10 model: GPT-6.1 Terra is #12 at $2 blended. A sensible default if you want a frontier lab model without frontier prices.
• Cheapest top 3 model: Claude Sonnet 5.5 is #3 at $4 blended, against $16 for Claude Fable 5.2, the top of the board.

Our take: run routine work on a cheap model and switch to a top 5 model only for the hard bugs and big refactors. If your coding tool lets you change models, that split alone can cut the bill by several times.

Is a subscription cheaper than the API for coding?

Often, if you code every day. Claude Code, Codex and Cursor plans bundle a lot of usage for a flat monthly price, while the API bills every token. Run your own numbers in the plan vs API calculator, or read Claude Pro vs API cost. If you already have a plan and hit its weekly cap, the cheapest way to keep going is in Claude weekly limit: cheapest API fallback and Codex weekly limit: cheapest API fallback.

Head to head: DeepSeek V4.1 Flash vs Claude Haiku 5.5, GPT-6.1 Terra vs DeepSeek V4.1 Flash, GPT-6.1 Sol vs Claude Sonnet 5.5. Want it on your own machine instead? See the best local LLM for coding.

Prices that change soon

• Gemini 4 Carbon: Launch price through Nov 30, 2026. From Dec 1: $4 in, $20 out. Implicit caching only. Not on the Free tier. Public API from Nov 10, 2026. (per Google's pricing page)
• Gemini 4 Argon: Launch price through Nov 30, 2026. From Dec 1: $4 in, $20 out. (per Google's pricing page)
• Gemini 3.8 Flash: Through Dec 31, 2026. From Jan 1, 2027: $1.50 in, $7.50 out. (per Google's pricing page)

How we rank

Quality is each model's Code Arena score from LMArena, built from blind side by side votes on real web app and agentic coding tasks. Where a model is listed at several effort levels we use its best one, which can cost more tokens than the default. Close ranks are often inside the margin of error. Prices are direct list prices; resellers and routers may charge differently. Our best model for coding page shows the same scores as a 0 to 100 relative score.