Catch-up — A Near-Frontier Open-Weight Model: Re-run Your Cost Math
A newly competitive open-weight model is a reason to rerun your own cost-quality comparison, not an automatic reason to switch.
- Moonshot says Kimi K3.1 Pro is a post-trained Kimi K3 version using substantially more post-training compute.
- Moonshot says it will release the full model weights by November 4, 2026.
- Simon Willison reports pricing of $2 per million input tokens and $10 per million output tokens.
- Willison reports an Artificial Analysis score of 55 for Kimi, behind only Claude Opus 5.5 at 58 among the models he lists.
Catch-up written 24 October 2026 for 2026-10-23; covers a source published 2026-10-23.
Aggregate benchmarks and token prices cannot show whether a model handles your workload, latency target, or provider constraints well enough.
Run representative cases through your current model and the candidate. Record quality, latency, and cost for the same workload before choosing whether to switch or route traffic.
Put it to work 30 minutes
A 20-case model cost and quality comparison
Pick one workload, run 20 real cases through your current model and Kimi K3.1 Pro, and record quality, latency, and cost per 1,000 runs.
See a worked example
Use a real, bounded workflow and record unknowns rather than estimating them. Keep the result small enough to review with a teammate.
- Workflow or decision
- A real workflow the team repeats each week.
- Evidence to record
- Quality, time, cost, failures, and reviewer notes.
- Decision rule
- Keep the current approach unless the agreed checks pass.
Your turn
Name the exact case you will test.
List the observations needed to make the decision.
Write what the evidence will make you do.
Your draft stays in this browser. No account needed.
Benchmarks are vendor and third-party aggregate scores, and the weights remain a promise until their stated date. Check provider terms before sending production data.
Reported facts and quotations were checked against the named sources; the recommended exercise is editorial guidance to test in your own context.
Does the candidate beat the current model on the measures that matter for this workload?
Watch: Quality and latency on the 20-case set · Actual cost per 1,000 equivalent runsKeep Going