Models / Use cases
Best coding models, ranked with exact prices
Coding is the workload where model choice matters most: tool-calling reliability, long-context repo comprehension, and cost per completed task diverge sharply between models. These are the picks we route the most code through, with exact prices.
Ranked
The picks
Claude Sonnet 5.5
Anthropic
The default for coding agents, the strongest tool-use reliability per dollar in the library.
Context
1M tokens
Max output
128K tokens
Tools
Yes
Vision
Yes
View Claude Sonnet 5.5 pricing and playground →
Kimi K3
Open-weights agentic pick, most of the tool-calling reliability at a fraction of the price.
GLM-5.2
The open-weights frontier challenger, strong coding and reasoning at a fraction of frontier pricing.
Gemini 3.8 Flash
Google builds this one for long-horizon software engineering and autonomous agents, at flash-tier prices.
DeepSeek V4.1 Flash
Bulk codegen, test scaffolding, and autocomplete at commodity prices, pair it with a frontier reviewer.
Replaces DeepSeek V4 Flash, retired by DeepSeek on 2026-09-10.
Retired
Previously recommended here
Bulk codegen, test scaffolding, and autocomplete at commodity prices, pair it with a frontier reviewer.
Use DeepSeek V4.1 Flash instead, same endpoint, same key, one line changes.
FAQ
Frequently asked questions
Which model should a coding agent default to?
Start with Claude Sonnet 5.5 for the main loop and route high-volume sub-tasks (lint fixes, summaries, commit messages) to a cheap tier. With one key across all of them, the split is a config change rather than a migration.
Can I use these models in Claude Code, Cursor, or Cline?
Yes. Every tool that accepts an OpenAI-compatible base URL works, because that is the only surface we expose. Point the tool at https://api.companyfabric.com/v1 with an sk-cf- key and change the model id. See the integration guides for exact setup per tool.
Do these models support tool calling and function calling?
Every pick on this page declares tool support in our catalog, and the facts panel on each model page says so per model. If a request needs tools, companyfabric/auto will only route it to a model that has them.
How much context do I need for a whole repository?
More than most repos need in one call. GLM-5.3-Flash and DeepSeek V4.1 Flash carry a million tokens, which fits a large codebase, but retrieval over the files that matter beats stuffing the window and costs far less per task.
What is the cheapest way to run a coding agent continuously?
Split the loop. A frontier model for planning and review, a flash tier for the mechanical edits. Because the price difference between tiers is 10x or more, where you draw that line matters more than which frontier model you picked.
One key. Every model. Exact prices.
Route every pick on this page through one key. Free starter credits included, no subscription required.