Language

Code generation & completion

Models and APIs built for writing, completing and fixing code.

What it is

Code models are LLMs trained and tuned specifically for programming: autocomplete inside an editor (including fill-in-the-middle), generating functions and tests, editing whole repositories as an agent, and explaining or reviewing code. They sit behind IDE assistants and coding agents.

How pricing works

Billed per token like any LLM, with input and output priced separately; we rank on output price. Coding agents send large repository context repeatedly, so input price, prompt caching and context window often matter more than the output price. Some providers sell a faster 'high-speed' variant at about double the price.

We compare everything in one unit: USD per 1M output tokens.

Typical uses

Autocomplete and inline edits in IDEs
Agentic coding: reading a repo, editing files, running tests
Writing tests, docs and migrations
Code review and bug explanation
Code translation between languages

What to compare

Pass rate on SWE-bench / LiveCodeBench-style benchmarks
Context window (how much of the repo fits)
Fill-in-the-middle support and latency for autocomplete
Tool-use reliability for agents
Output speed

Watch out for

Cost per task, not per token, is what counts: a cheaper model that needs more retries can cost more.
Autocomplete needs low latency and FIM; a chat-style model is a poor fit even if it is smart.
Agents re-send long context each step, so input and cache pricing dominate.

Capabilities explained

Input price · $ / 1M tokens: Cost of the code and context you send. Large for coding agents because the repo is re-sent often.
Context window · tokens: How much code the model can read at once. 256K tokens fits roughly a mid-sized codebase module set.
Fill-in-the-middle: Supports completing code between a prefix and suffix, required for editor autocomplete.
Tool use / agent mode: Can call tools (shell, editor, search) so it can act as a coding agent, not just write snippets.
Faster tier available: The provider sells a higher-speed version of the model, usually at a premium price.
Open weights: The model weights are downloadable, so you can self-host it or switch hosting providers without changing models.

Glossary

Fill-in-the-middle (FIM): Completing code given the text before and after the cursor, rather than only continuing from the end.
SWE-bench: A benchmark where models must fix real GitHub issues; a common yardstick for coding agents.
Agentic coding: The model plans, edits files, runs commands and tests in a loop with tools.

All 5 APIs, cheapest first

Free to browse

Tick 2 to 4 APIs to compare them side by side.

API	$ / 1M out tokens	Context window	Fill-in-the-middle	Tool use / agent mode	Quality	Price data
Codestral Mistral AI	$0.90 in $0.30	256K tokens	Yes	Yes	–	Unverified sample
grok-build-0.1 xAI (Grok)	$2.00 in $1.00	256K tokens	—	Yes	–	Verified Oct 6, 2026
Kimi K2.7 Code Moonshot AI (Kimi)	$4.00 in $0.95	262.144K tokens	—	Yes	–	Verified Oct 6, 2026
Qwen3-Coder Alibaba Cloud (Qwen)	$5.00 in $1.00	256K tokens	Yes	Yes	–	Unverified sample
Kimi K2.7 Code (high-speed) Moonshot AI (Kimi)	$8.00 in $1.90	262.144K tokens	—	Yes	–	Verified Oct 6, 2026