Short version: Two "cheap agent backend" options are live the week of July 26, and they're cheap in opposite ways. Meta's Muse Spark 1.1 is the cheapest frontier-class token you can buy — a reported $1.25 / $4.25 per million input/output, about a quarter of Anthropic and OpenAI list prices. Kimi K3's hosted API is 2.4×–3.5× more expensive at $3 / $15 — but its weights drop July 27 under a Modified MIT license, and it's the only one of the two you can take home and run forever. The question isn't which is cheaper. It's whether the risk that would actually sink you is your bill or your dependency.
The two prices, side by side#
Muse Spark 1.1 went to public preview on July 9 as Meta's first-ever paid developer API. The pricing is the story: a reported $1.25 per million input tokens and $4.25 per million output — Mark Zuckerberg pitched it as roughly 25% of what Anthropic and OpenAI charge for comparable models (AI Weekly; Quartz). It ships with a 1M-token context window, computer use across desktop, browser and mobile, MCP support, custom skills, and parallel subagent orchestration (DataCamp). Access is US-first with a waitlist for everyone else.
Kimi K3, Moonshot's 2.8-trillion-parameter open-weight mixture-of-experts (896 experts, 16 active per token, 1M context), lists at $3 per million input — $0.30 on a cache hit — and $15 per million output, flat across the full context window (OpenRouter). It's the largest open-weight model ever released, and by launch-week benchmarks it lands near the frontier — around fourth on the Artificial Analysis Intelligence Index, on par with Opus 4.8 and GPT-5.5 (VentureBeat). The full weights ship July 27 under a Modified MIT license (Interconnects).
On price-per-token today, Muse Spark wins outright. On portability, rate-limit exposure, deprecation risk, and data control, K3 is the only answer. They are not competing for the same slot.
Why cheaper-per-token isn't the whole decision#
If your product's failure mode is a runaway token bill — high-volume, low-stakes agent work where margin is thin — Muse Spark 1.1 is the obvious default. It's the cheapest frontier-class token on the market this week, and there's no infrastructure to run. That's a real and common situation, and for it the answer is Meta.
But there's a second failure mode, and it doesn't show up on a pricing page: dependency. Muse Spark is closed and Meta-hosted. You inherit its rate limits, its preview waitlist, its terms of service, and the standing possibility that the model gets deprecated or repriced on Meta's calendar rather than yours. For a preview product from a vendor shipping its first paid API, that's not paranoia — it's the base rate.
Kimi K3 is the answer to that second risk, and only that one. Once the weights land July 27, you can run the exact model on your own or a neutral provider's hardware, indefinitely: no waitlist, no deprecation, and prompts that never leave your boundary. That's the entire case for paying 2.4×–3.5× more per token on the hosted endpoint in the meantime — you're buying an exit.
The honest catch on "you can self-host it"#
"Open weights" is doing a lot of work in that sentence. A 2.8T-parameter model is not something a team of one spins up on a spare GPU — self-hosting it is an infrastructure project that needs a large multi-GPU cluster, and until your token volume is high enough, a hosted endpoint (K3's own, or a neutral reseller's) will be cheaper than standing up the hardware. We did the rent-vs-host math in what 594GB of open weights actually cost a founder, and the honest conclusion for almost every solo builder is: rent it now, keep the option to self-host, and don't confuse "can" with "should." The value of the open weights for most founders is optionality, not day-one deployment.
The call#
- Optimizing purely for the lowest token bill, no infra, US-based? Muse Spark 1.1. Nothing beats $1.25/$4.25 this week.
- Need portability, data control, or an escape hatch from a single vendor? Kimi K3 — hosted now, self-hosted when volume clears the GPU math.
- Most founders? Both, behind one swappable interface: cheap high-volume work on Muse, portability-sensitive work on K3.
Whichever you pick, the one non-negotiable is making the backend a config value, not a hard-coded dependency — the same discipline that lets you chase the cheapest workhorse as prices keep dropping and, if you're weighing K3 against the other closed cheap tier, the Kimi K3 vs Claude Sonnet 5 math. The models will keep leapfrogging on price. Your architecture is what decides whether that's an opportunity or a migration.
