Every lab ships a "coding model" now. Their output prices run 17x apart.
Filter the July 2026 rate cards down to the models sold specifically for coding agents and one thing jumps out: the label tells you nothing about the bill. Qwen3 Coder Next charges $0.80 per million output tokens. GPT-5.3 Codex charges $14 for the same unit. That is a 17.5x gap inside a single category, and because a coding agent writes far more than it reads, the output rate is the number that ends up running your month.
Photo by Mohammad Rahmani on Unsplash
A year ago "coding model" meant one or two options. Now every serious lab ships one as a distinct product: OpenAI has the Codex line, Alibaba has Qwen3 Coder, Moonshot has Kimi K2.7 Code, Mistral has Devstral and Codestral, Kwaipilot has KAT-Coder, and Z.ai sells GLM-5.2 as coding-first. The pitch is always the same, a model tuned to plan, edit, and iterate inside a repo. What nobody puts on the marketing page is that these models do not share a price band. They barely share a zip code.
Here is every dedicated coder we track, sorted by output rate, the number that matters most for agent work. The input column is close to noise by comparison, and you will see why in a minute.
The rate cards, side by side
| Model | Vendor | Input / 1M | Output / 1M | Context |
|---|---|---|---|---|
| KAT-Coder-Air v2.5 | Kwaipilot | $0.15 | $0.60 | 256K |
| Qwen3 Coder Next | Alibaba | $0.11 | $0.80 | 256K |
| Codestral | Mistral | $0.30 | $0.90 | 256K |
| Devstral | Mistral | $0.40 | $2.00 | 256K |
| KAT-Coder-Pro v2.5 | Kwaipilot | $0.74 | $2.96 | 256K |
| Kimi K2.7 Code | Moonshot | $0.95 | $4.00 | 256K |
| GLM-5.2 | Z.ai | $1.40 | $4.40 | 1M |
| GPT-5.3 Codex | OpenAI | $1.75 | $14.00 | 400K |
Rates are the cheapest widely available list price per provider, cache misses, as tracked on our pricing page. Chinese-lab models often route cheaper on OpenRouter than on the vendor's own cloud; Qwen3 Coder Next, for instance, is $0.11/$0.80 on OpenRouter but $0.30/$1.50 on Alibaba DashScope.
Output is the lever, not input
People sanity-check a coding model by the input price because it is the first column they see. That instinct is backwards for agents. A coding agent does read a lot, it pulls files, diffs, and tool results into context, but it also writes constantly: reasoning traces, full file rewrites, patch after patch, and the long self-corrections that eat tokens when a test fails. On the models where the split is measured, output is where the money goes, and output rates vary far more than input rates do across this table.
Take a developer leaning on an agent all month: roughly 40M input and 12M output tokens, output-heavy because the thing keeps regenerating code. Run that identical workload through all eight and the bill does not drift, it detonates.
| Model | 40M in / 12M out | vs cheapest |
|---|---|---|
| KAT-Coder-Air v2.5 | $13.20 | 1.0x |
| Qwen3 Coder Next | $14.00 | 1.1x |
| Codestral | $22.80 | 1.7x |
| Devstral | $40.00 | 3.0x |
| KAT-Coder-Pro v2.5 | $65.12 | 4.9x |
| Kimi K2.7 Code | $86.00 | 6.5x |
| GLM-5.2 | $108.80 | 8.2x |
| GPT-5.3 Codex | $238.00 | 18.0x |
Eighteen times, bottom to top, for the same 52M tokens. Notice how the ranking tracks the output column almost perfectly and ignores input entirely: Qwen3 Coder Next has a higher input rate than KAT-Coder-Air yet lands within a dollar of it, because the 12M output tokens swamp the difference. If your mix ran even heavier on output, say a refactor sprint that is nearly all generation, the top of this table would pull further away still.
The word "coder" hides two opposite bets
Sort the table and a split appears that has nothing to do with quality. The cheap end is Alibaba and Kwaipilot, labs that treat a coding model as a high-volume, efficiency-first product: small active-parameter MoE designs, priced to run flat out inside an agent loop all day. Qwen3 Coder Next activates just 3B parameters out of 80B. KAT-Coder-Air is a distilled sibling built explicitly for high-volume traffic. Cheap output is the entire design goal.
The expensive end is OpenAI. GPT-5.3 Codex is a premium coding-tuned SKU that carries the same $14 output rate as OpenAI's frontier general models, and the "Codex" label is not a discount, it is a specialization tax. Z.ai and Moonshot sit in between, pricing their coders like capable general models rather than like commodity inference. So "dedicated coding model" does not mean "cheaper than a general model." For OpenAI it means the opposite.
You cannot cleanly buy the price with benchmarks
The obvious rebuttal is that you get what you pay for, that the $14 model must be far better on real coding tasks. Maybe. But the published numbers do not let you prove it, because almost none of these models are measured on the same test set with the same scaffold. Here is what each vendor actually claims, which is not the same as what an independent eval would find.
| Model | Claimed coding score | Which test |
|---|---|---|
| Qwen3 Coder Next | 70.6 | SWE-bench Verified (SWE-Agent) |
| Devstral | ~72 | SWE-bench Verified (vendor) |
| KAT-Coder-Pro v2.5 | 65.2 | SWE-bench Pro |
| GLM-5.2 | 62.1 | SWE-bench Pro |
| Kimi K2.7 Code | Not on public SWE-bench | Vendor's own bench only |
| GPT-5.3 Codex | Vendor SOTA claim, no shared number | SWE-bench Pro / Terminal-Bench |
SWE-bench Verified and SWE-bench Pro are different test sets and do not convert to each other; scaffolds differ too. All figures here are vendor-reported unless noted, and Moonshot never submitted Kimi K2.7 Code to a public leaderboard. Read the column as "roughly capable," not as a ranking. Codex may well be the strongest, it just is not on a shared axis with the cheap ones.
Shop the output column, not the label
Start from your output volume, not the leaderboard. If you run a coding agent hard and it regenerates a lot of code, the output rate is your bill, so a $0.80 model and a $14 model are a different business decision, not a rounding difference. Qwen3 Coder Next and KAT-Coder-Air are the default cost-floor picks, both land near $14 a month on the workload above, and both score in the low 70s on their own SWE-bench sheets. Try the cheap coder first and only move up if it visibly fails your tasks.
Move up when the failure mode costs more than the tokens. GLM-5.2 buys you a real 1M context if you feed the agent whole repos; Kimi K2.7 Code and KAT-Coder-Pro sit in the sensible middle; GPT-5.3 Codex is the model you pay $238 for because a wrong patch in production is worth more than the spread. What you should not do is pay Codex rates by default because the word "coder" made all these models look interchangeable. Plug your real token split into the cost calculator before you commit, because the ranking flips with your input-to-output ratio.
Sources
- Qwen3 Coder Next - Alibaba Qwen blog (80B/3B MoE, SWE-bench Verified claim)
- GLM-5.2 pricing - OpenRouter ($1.40 input / $4.40 output, 1M context)
- Kimi K2.7 Code - Requesty model page ($0.95 / $4.00, $0.19 cached input)
- GPT-5.3 Codex - LLM-Stats ($1.75 / $14.00, 400K context)
- KAT-Coder v2.5 - Kwaipilot on Hugging Face (Pro and Air variants, SWE-bench Pro 65.2)
- Devstral and Codestral - Price Per Token, Mistral (Devstral $0.40 / $2.00, Codestral $0.30 / $0.90)