THE VERDICT
● Same price from today: $3 in, $15 out per million.
● Sonnet 5 for ecosystem, tooling depth and Claude Code bundled with a $20 plan.
● Kimi K3 for open weights, self-hosting and a free tier you can genuinely evaluate on.
● The hidden difference: Sonnet 5 got a tokenizer change today adding 10 to 35 percent more tokens on code.
Head to head
| Kimi K3 | Claude Sonnet 5 |
|---|
| API per 1M | $3 / $15 | $3 / $15 from today |
| Tokenizer change today | No | Yes — 10 to 35% more tokens on code |
| Weights | Modified MIT, downloadable | Closed |
| Free evaluation | Adagio — unlimited K2.6 chat | Claude Free, message-capped |
| Bundled coding tool | Kimi Code from $19 | Claude Code from $20 |
| Ecosystem | Narrower | Deep, wide tooling |
SAME RATE DOES NOT MEAN SAME BILL
Sonnet 5 changed tokenizer today. The same file now produces 10 to 35 percent more tokens than last week.
So on identical rate cards, an identical codebase costs more on Sonnet 5 than on K3 — which is a comparison nobody can make from a pricing page.
Which one
| If you are... | Pick |
|---|
| Building on a deep tooling ecosystem | Sonnet 5 |
| Needing to self-host for compliance | Kimi K3. Read the modified MIT terms first |
| Running heavy code workloads on API | Measure both. The tokenizer change makes the rate card misleading |
| Evaluating before committing | Kimi Adagio free. Unlimited chat, no credits drawn |
FAQ
Are they really the same price?
On the rate card, yes — both $3 and $15 per million from today. The effective cost differs because Sonnet 5 changed tokenizer, producing more tokens for the same input on code.
Can I download Kimi K3?
Yes, under a modified MIT licence. Modified matters, so read the terms rather than assuming standard MIT semantics.
Which is better at coding?
Claude Code ranks first among agent harnesses, which is a tooling advantage rather than purely a model one. On raw model comparison the gap is narrower than the ecosystem gap.