Short version: July 2026's model price war produced a result almost nobody predicted a month ago: the cheapest capable agent backend you can rent today is a closed model. On July 21, Google cut Gemini 3.6 Flash to $1.50/M input and $7.50/M output — and that $7.50 now sits below both Kimi K3's hosted API ($15/M output) and Claude Sonnet 5's promo ($10/M output). The open 2.8-trillion-parameter model everyone was excited to self-host is, for now, the most expensive of the three to rent. Here's the honest math on what the lower bill costs you.
The two July entrants, on one card#
The interesting fight this week isn't K3-versus-the-frontier. It's the two models that both changed the cheap tier in the same week: Google's Flash cut and Moonshot's open drop.
| Gemini 3.6 Flash | Kimi K3 | |
|---|---|---|
| Weights | Closed | Open — full drop July 27 |
| Input $/M | $1.50 ($0.15 cached) | $3.00 ($0.30 cache-hit) |
| Output $/M | $7.50 (was $9.00) | $15.00 |
| SWE-bench Pro | 58.7% | Not independently published |
| Context | 1M tokens | 1M tokens |
| Self-host | No | Yes, after July 27 |
On July 21, Google shipped Gemini 3.6 Flash and cut the output price from 3.5 Flash's $9.00 to $7.50/M, with input at $1.50/M and cached input at $0.15/M. There's a batch/flex tier at $0.75/$3.75. It carries a 1M-token context and scores 58.7% on SWE-bench Pro, up from 55.1% on the prior Flash. Our July 24 Founder's Wire flagged the cut as it landed; this is the decision piece underneath it.
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). 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 run yourself. (If you want to try it now, we wrote how to call the Kimi K3 API in 10 minutes.)
The one number that flips the usual story#
Founders reach for open weights partly on a reflex: open must be cheaper. Right now, it isn't. Rank the three cheap-tier options on hosted output price and you get a clean, counter-intuitive line:
Flash ($7.50) < Sonnet 5 promo ($10) < Kimi K3 ($15). The open model is the most expensive one to rent.
That's the whole non-obvious idea. K3's open weights are a real asset — but on the hosted bill, an open model is priced like any other API, and Moonshot priced it at frontier-adjacent rates. K3 gets cheaper than Flash only when you stop renting and start self-hosting, and your own per-token inference cost drops below $7.50/M. That crossover is real, but it's far out: in 4-bit MXFP4 precision the weights alone need roughly 1.4 TB of fast memory before you load a single token of context. This is a GPU-cluster deployment, not a side-project. Below sustained, high volume, the hosted Flash bill wins on arithmetic every time.
We walked through the same trap in rent vs. self-host for the 2.8T decision: the open weight is an option you exercise at scale, not a discount you get on day one.
What the cheaper bill actually costs you#
Flash wins on price. It doesn't win on everything, and pretending otherwise is how founders end up re-platforming in Q4.
- Agentic coding. Flash lands 58.7% on SWE-bench Pro. That's fine for agents that mostly retrieve, summarize, route, and call tools. If your agents write and repair real code, it trails the models built for that — which is why the coding-heavy comparison is Kimi K3 vs Claude Sonnet 5 (63.2% Pro on a $2/$10 promo), not Flash.
- Lock-in. Flash is closed and Google-hosted. The low price buys zero portability — no weight to move, no fallback host, no data-residency story beyond Google's. K3's entire case is the opposite: a weight you own and can relocate. That's worth a premium to some founders and nothing to others; know which you are.
- Migration cost. Reviewers describe the 3.5-to-3.6 Flash move as non-trivial — prompts and tool-calling behavior shift enough that you should validate on a slice of traffic before cutting over. Budget the engineering hours; they're part of the price.
The decision, in one line#
- Default to Gemini 3.6 Flash if your agents are high-volume and tool-heavy (retrieval, summarization, routing, classification) and you want the lowest bill with zero infrastructure. At $1.50/$7.50 it's the cheapest capable rental in July 2026.
- Choose Kimi K3 when you need an open weight — for data residency, lock-in avoidance, or self-hosted bulk once your volume clears the GPU math. You're buying portability and a 1M context you own, not a lower hosted invoice.
- Weigh Claude Sonnet 5 if the work is coding-first; its $2/$10 promo and 63.2% SWE-bench Pro make it the near-frontier value pick through August 31.
And do the one thing that makes the whole price war a non-event for you: keep the backend model-swappable. The cheapest capable model changed twice in one week. It will change again. The founders who win the price war are the ones who can move volume between Flash, K3, and Sonnet with a config change — a pattern we laid out in how to cost-route open and closed models. Price is a moving target; an abstraction layer is not.
