Short version: Everyone benchmarks Kimi K3 against the frontier — Opus 4.8, GPT-5.6 Sol. But that's not the decision most founders actually make, because most bulk agent work never touches a frontier model. The real July-2026 default is the cheap tier: Kimi K3's open 2.8-trillion-parameter weights against Claude Sonnet 5's promo-priced $2/$10. And the surprise is that the open model isn't the cheaper one — not yet. Here's the math.
The two models, on one card#
On July 16, Moonshot AI shipped Kimi K3 — a ~2.8-trillion-parameter open-weight mixture-of-experts model (896 experts, 16 active per token, a 1M-token context window). It's hosted at $3/M input ($0.30 cache-hit) / $15/M output, and the full open weights are slated for July 27, which makes it the only one of the two you can eventually self-host.
On June 30, Anthropic shipped Claude Sonnet 5 — closed, but running an introductory $2/M input / $10/M output through August 31 (then $3/$15). It scores 63.2% on SWE-bench Pro, about six points behind Opus 4.8's 69.2%, at roughly 40% of Opus's price.
| Kimi K3 | Claude Sonnet 5 | |
|---|---|---|
| Weights | Open — full drop July 27 | Closed |
| Input $/M | $3.00 ($0.30 cache-hit) | $2.00 promo → $3.00 after Aug 31 |
| Output $/M | $15.00 | $10.00 promo → $15.00 after Aug 31 |
| SWE-bench Pro | trails frontier* | 63.2% |
| Context | 1M tokens | 200K standard |
| Self-host | Yes (from July 27) | No |
\On Moonshot's own evaluation suite K3 tops earlier open models, but Moonshot concedes it still sits behind Claude Fable 5 and GPT-5.6 Sol on overall performance. Independent SWE-bench Pro numbers for K3 weren't published at press time — treat its coding claims as vendor-reported until the weights land.*
The surprise: the open model is not the cheaper model#
The instinct is "open weights = cheaper." Right now that's backwards. Agent backends are output-heavy — a task reads a little and writes a lot (diffs, tool calls, retries, explanations), so output price dominates the bill. And on output tokens, Sonnet 5's promo ($10/M) undercuts Kimi K3's hosted API ($15/M) by a third. On input it's $2 vs $3. The hosted open model loses on both.
K3 only wins on price the day you self-host it — and self-hosting a 2.8-trillion-parameter MoE is not free. You're renting (or buying) enough accelerator memory to hold the weights, running your own inference stack, and eating the idle time between requests. That math only clears at real volume — call it hundreds of millions of tokens a month — where amortized GPU cost per token finally drops below $10–15/M.
What it means: below serious scale, the "cheaper" choice today is the closed model on promo. The open weights are a bet on your future volume and your appetite to run infrastructure, not an instant discount.
Where Kimi K3 actually wins#
Three real reasons to pick K3 anyway:
- The 1M-token context window. Sonnet 5's standard context is 200K. If your agent stuffs whole codebases, long transcripts, or big document sets into a single call, K3's 1M window is a capability Sonnet doesn't match at the tier.
- Portability and control. Once the weights drop July 27, K3 is yours — self-host for data residency, air-gap it, fine-tune it, or escape a vendor price change. Sonnet 5 is an API you rent; if Anthropic moves the price (and it will, September 1), you move with it.
- Volume economics, eventually. At high enough throughput, self-hosted K3 undercuts any hosted rate. That's the endgame the open weights buy you.
Where Claude Sonnet 5 wins#
- Cheapest near-frontier coding, today. 63.2% on SWE-bench Pro at $2/$10 is the best price-to-capability on the hosted market this summer. For most bulk agent work — codegen, refactors, tool-calling loops that don't sit at the frontier of difficulty — it's the pragmatic default.
- Zero infrastructure. No GPUs, no inference stack, no on-call for a 2.8T model. You ship an API key.
- A known, dated price. You can plan around the promo and the September 1 step-up to $3/$15. No hardware capacity to forecast.
The decision#
For most solo founders and early builders in July 2026, default your bulk agent backend to Claude Sonnet 5 while the promo runs. It's cheaper per output token than hosted K3, it's near-frontier on the coding benchmark that matters, and it costs you nothing to run.
Reach for Kimi K3 when you need the 1M-token context, when an open self-hostable weight matters for residency or lock-in, or once your volume is high enough that running the weights beats the hosted bill.
And do the one thing that survives either choice: make your backend model-swappable. Put both behind one interface so you can move volume from Sonnet 5 to a self-hosted K3 the week the weights land and your GPU math clears — without rewriting your agent. Two dates are worth a calendar entry: July 27 (K3 weights) and September 1 (Sonnet 5's promo ends and its output price rises to meet K3's).
Next: how to cost-route between these two behind one OpenAI-compatible client, with the code.
