The line that isn’t there
Search the site’s tables for either model and you get nothing back. We track 2843 LLM price rows and 117 GPU quotes, refreshed daily, and neither Seed 2.0 Code nor KAT-Coder has a metered input/output line in that set. That gap is the finding, not a caveat to apologize for.
Both ship through vendor contracts or invite-only access rather than a public metered API. That’s common for the current wave of purpose-built coding models: the lab wants usage data and lock-in before it commits to a transparent per-token rate. So the comparison the title promises, a straight $/M-token race, doesn’t exist yet in a form anyone can verify. What does exist, and what you can price today, is the architecture pattern both models copy.
The pattern both models copy
Every purpose-built coding release this cycle takes a large mixture-of-experts base and fine-tunes it for tool use, repo navigation, and long agent loops, keeping active parameters small relative to total parameters so inference stays cheap even as the model gets smarter. Two open-weight releases on our tables follow that exact shape and give you real numbers to reason from: Kimi K2.7 Code, a coding-specific finetune at 1059B total params with 32B active, and Qwen3 Coder, 480B total with 35B active. If Seed 2.0 Code and KAT-Coder resemble anything priceable, it’s one of these two bands.
Run the memory math. At INT4 (roughly 0.55 GB per billion params) plus 25% KV headroom, Kimi K2.7 Code’s 1059B total lands near 730GB, which pushes you past a single 8-card node and into multi-node territory on H200 or B200. Qwen3 Coder’s 480B total lands near 330GB, which fits inside a single well-stocked 8x H100 box with room to spare. That’s a real gap in infrastructure commitment between the two bracket cases. Seed 2.0 Code and KAT-Coder sit somewhere inside it, whichever tier their undisclosed active-param count actually lands in.
Why the headline rate wouldn’t tell you much anyway
Even where per-token rates are public, they mislead. Qwen3 235B A22B shows a provider spread over 20x between the cheapest and priciest host carrying it, as of this writing, purely from hosting margin and infrastructure choice with the identical weights underneath. A closed coding model’s single contract quote carries that same kind of hidden variance, just collapsed into one number you can’t audit. You have no way to know if the quote reflects a lean deployment or a padded one, because there’s no competing host publishing a second number to check it against.
Agentic coding work compounds the problem. A task that a chat benchmark scores as one prompt and one completion runs through dozens of tool calls, file reads, and retries in an actual coding agent loop, and the token bill scales with that loop count, not with the headline per-request price. A vendor quoting you a flat contract rate for Seed 2.0 Code or KAT-Coder is pricing against their own internal average loop length. If your repo tasks run longer loops than that average, and in agentic coding they usually do, you’re paying above the number they quoted you.
What this can’t tell you
This can’t tell you what Seed 2.0 Code or KAT-Coder’s actual active-parameter count is, because neither vendor publishes it, and it can’t tell you whether either model’s real inference cost sits closer to the Kimi K2.7 Code bracket or the Qwen3 Coder bracket. That’s a genuine unknown, not a rounding error, and no amount of reasoning from analogs closes it. All the analog math gives you is a floor and a ceiling to negotiate against.
The rule
Before you sign a contract rate for either model, price the closer analog’s node cost on /gpus and its metered rate on /models, then run your own repo tasks, full agent loops, not single-turn benchmarks, through the calculator against that analog. If the vendor’s contract quote beats the analog’s total cost by a margin wide enough to cover the risk of single-vendor lock-in, take the contract. If it doesn’t clear that margin, you’re paying a premium for a model you can’t verify against a second host, and the open analog is the safer default until one of these vendors publishes a real per-token rate you can check.