Kimi K3 Creates a New AI Service: Sell 'Real-Cost' Audits That Price Open-Weight Models by Completed Task, Not by Rate Card.

Kimi K3 landed with a clean rate card and a strong scorecard.

That combination is exactly what makes it dangerous to buy on the headline numbers alone.

What happened

Moonshot AI released Kimi K3, described as "Open Frontier Intelligence," built on a 2.8-trillion-parameter mixture-of-experts architecture (Kimi Delta Attention and Attention Residuals) that activates 16 of 896 experts per token through what it calls Stable LatentMoE. It runs MXFP4 weights with MXFP8 activations and supports a 1-million-token context window with native multimodal input.

"The full model weights will be released by July 27, 2026," the announcement states -- the model is already live on Kimi.com, the Kimi Work desktop app, Kimi Code CLI, and the Kimi API Platform, with mobile apps on iOS, Android, and HarmonyOS.

At "max reasoning effort," Kimi K3 posts GPQA-Diamond "93.5", Terminal Bench 2.1 "88.3", Program Bench "77.8", MMMU-Pro "81.6", and DeepSWE "67.5". Moonshot's own framing is candid about where it sits: "While its overall performance still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol..."

Pricing is published directly: "$0.30" per million tokens for cache-hit input, "$3.00" for cache-miss input, and "$15.00" per million output tokens, with a cache hit rate cited as "above 90% in coding workloads."

The gap this creates

A rate card is not a cost. Hacker News commenters zeroed in on this fast: several pointed out that if a model needs several times more reasoning tokens than a competitor to reach a comparable answer, the apparent price advantage on paper can disappear entirely once you count actual tokens spent per completed task. That is a workload-dependent number no vendor benchmark publishes, because it depends on the client's own prompts, not a leaderboard suite.

This is the exact gap a service business can occupy. Two teams running the same "$0.30 / $3.00 / $15.00" rate card can land on completely different real costs depending on their cache-hit rate and how hard their tasks push the reasoning mode. Nobody sells that measurement as a packaged product yet -- most teams either trust the rate card at face value or never bother to check.

The business idea

Sell three tiers around exactly this measurement gap:

1. A real-cost audit -- a fixed-price, fast-turnaround engagement that runs a client's own tasks (not benchmark suites) through their current model and a candidate like Kimi K3, then reports tokens-per-completed-task and dollars-per-completed-task, not the vendor's advertised rate.

2. A cache-shape review -- since cache-hit and cache-miss pricing differ by 10x on Kimi's own card, map which parts of a client's workload will actually hit cache and which won't, before they commit budget to a switch.

3. A migration engagement -- for clients whose numbers clear the bar, stand up the workflow on the new model through its official access points once the audit says it's worth doing.

Money play

1. Sell a fixed-scope "real cost" benchmark: the client's own production prompts through both models, tokens counted per completed task, not per benchmark question.

2. Break the cost model into cache-hit and cache-miss buckets using the published "$0.30" / "$3.00" split, and measure the client's actual cache-hit rate against the vendor's cited "above 90%" figure for coding workloads.

3. Isolate reasoning-mode token overhead as its own line, since "max reasoning effort" is what the published benchmark scores are measured at, and reasoning tokens bill as output at "$15.00" per million.

4. Price the audit as a short fixed-fee report so it's a low-friction first purchase, with a clear recommendation and break-even volume.

5. Convert audits that favor switching into a migration engagement using Kimi K3's existing access points -- Kimi.com, Kimi Work, Kimi Code CLI, or the API platform -- ahead of the full weight release on July 27, 2026.

Bottom line

Kimi K3 ships with a rate card that reads like a bargain and a benchmark sheet that reads like a near-miss on the frontier. Neither number tells a buyer what their own workload will actually cost. That measurement -- real tokens per real task -- is the service, and almost nobody is packaging it yet.

Sources:

https://www.kimi.com/blog/kimi-k3

https://news.ycombinator.com/item?id=48935342