AI Coding Models: The Full Reference Table (July 2026)

I keep researching this for my own work, so I turned it into a reference. One big table, everything in it: prices, benchmarks, licenses, where to access each model. Scan it, pick what you need, go build.

Quick mental model before the table: Agent = Model + Harness. The model is the intelligence, the harness (Claude Code, Cursor, Cline, Aider...) turns it into something that navigates your repo and fixes its own mistakes. Pick both.

The table

How to read it:

• Prices are first-party rates as of late June 2026. OR = OpenRouter weighted average, noted where it differs a lot from direct.
• OpenRouter adds ~5.5% credit fee ($0.80 minimum) and 5% BYOK fee above 1M requests/month. The per-token rates themselves match provider list prices.
• Benchmarks mix vendor-reported and independent numbers. Compare within the same column only, and test on your own code before committing.
• Fable 5 caveat: a US export-control directive suspended access on June 12, expected back for US users around July 1. If you are outside the US (like me), verify before building on it.

The agents (harnesses)

Providers cheat sheet

Three numbers to remember

The 10x cliff. Five models score between 80.2% and 80.6% on SWE-bench Verified (DeepSeek V4 Pro, Gemini 3.1 Pro, MiniMax M3, Qwen3.7 Max, Kimi K2.6), with output prices from $2.40 to $12 per million. The next 8 points up (GPT-5.5, Opus 4.8) cost $25 to $30. The last 8% of quality costs 10x. Know if your work lives in that gap.

The 21.7 point gap. Claude Fable 5 scores 80.3% on SWE-Bench Pro vs 58.6% for GPT-5.5. Frontier benchmarks usually move in single digits. This one did not.

1/36th. DeepSeek V4 Flash delivers about 82% of Opus's SWE-Bench Pro score at roughly 1/36th the input price. It does not replace Opus. It means Opus should stop doing DeepSeek's job.

My routing shortcut

• Cheap and low-risk work: DeepSeek V4 Flash
• Default coding workhorse: Kimi K2.6 or Sonnet-class
• Hard architecture, repeated failures: frontier (Opus 4.8, GPT-5.5, Fable 5 if you can get it)
• Private or offline: Qwen local via Ollama
• Always: measure cost per successful task, not cost per token. A cheap model that causes rework is expensive.

This space moves monthly. I will update this table when the next shakeup lands. Building with agents and want to compare notes? I am @itseduvieira pretty much everywhere.