What you will learn
- At launch, K3 pricing separates cached input, uncached input, and output tokens.
- Output, reasoning, retries, and agent loops can matter more than the headline input rate.
- Cost per accepted task is more useful than cost per million tokens.
01
Read the rate card correctly
Moonshot's July 2026 K3 rate card lists separate prices for cache-hit input, cache-miss input, and output tokens. Always verify the live platform page before budgeting because rates, limits, and promotions may change.
A low cached-input rate is valuable only when the application repeatedly sends a stable prefix that the provider recognizes as cacheable. Dynamic user content, changing tools, or unstable prompt ordering can reduce cache reuse.
- Record cache-hit and cache-miss tokens separately
- Track reasoning and visible output tokens
- Include tool and search charges
- Measure retries, failures, and abandoned runs
02
A simple cost model
For one request, estimate uncached input tokens multiplied by the uncached rate, cached tokens multiplied by the cache-hit rate, and output tokens multiplied by the output rate. Add tool fees and infrastructure costs where applicable.
For an agent, sum every model call in the trajectory. A task that uses planning, browsing, coding, testing, and revision can create many calls. Average cost hides expensive tail cases, so report median, 90th percentile, and maximum cost.
03
Why long context can become expensive
A one-million-token window makes large inputs technically possible, not automatically economical. Sending an entire repository or document archive on every request can increase prefill work, latency, and distraction even when cached.
Use retrieval, file maps, summaries with provenance, and progressive disclosure. Test whether a smaller evidence set preserves or improves acceptance rate. Context reduction should remove irrelevant material, not erase evidence needed for verification.
04
Compare K3 with alternatives fairly
Create workload bands: short extraction, medium drafting, long research, coding, and multi-step agents. For each model, use the same source material, output requirements, and acceptance rubric. Record total tokens and tool calls rather than applying list prices to an imagined average.
Compare against smaller models too. K3 may be cost-effective for difficult tasks but unnecessary for routine routing or formatting. A cascade can begin with a cheaper model and escalate low-confidence cases.
05
The economic metric that matters
Cost per accepted task divides total model, tool, and review cost by the number of results that meet the acceptance criteria. Add the value of human correction time and the expected cost of important failures.
A sustainable production decision balances quality, speed, cost, operational control, and risk. Recalculate monthly because traffic mix, prompts, model behavior, cache hit rates, and provider prices change.
FAQ
Common questions
Is Kimi K3 cheaper than GPT or Claude?
Its list price can be lower than some frontier models, especially for cached input, but the answer depends on output length, retries, tools, and accepted-task rate.
Should I send the full one-million-token context?
Only when tests show that the extra context improves results enough to justify latency and cost. Retrieval is usually preferable to indiscriminate context loading.
Sources
Primary sources and live documentation
These links point to authoritative documentation used to verify and maintain this guide for the July 2026 update.
Turn the method into a reusable instruction.
Explore expert prompts