Cost-Per-Pass, Not Cost-Per-Call#

Practitioners price LLMs by the per-token rate on the provider’s pricing page. For autonomous agents, that number is misleading. Two layers of indirection sit between the per-token rate and the cost you actually pay to get work done: variable prompt sizes turn per-token into per-call, and variable pass rates turn per-call into per-pass. Each layer can invert the ranking.

For autonomous fleets where failed attempts trigger reviewer cycles, retries, and reputational drag, cost-per-pass is the only metric that ranks models correctly. This article shows how to compute it, when it dominates, and where the cheapest-per-token model becomes the most expensive in production.

LLM Adapter Audit Checklist#

When you wrap an OpenAI-compatible LLM provider (Moonshot, DeepSeek, xAI, Together, Fireworks, OpenRouter, vLLM, anything else that exposes POST /v1/chat/completions) in a Go HTTP client, the same ten bug classes show up. They all silently degrade or break the agent — none of them crash loudly. Each was observed in production across at least one of xAI, DeepSeek, or Moonshot during a two-week audit period.

This checklist is the audit. Run it against any new adapter before shipping. Each entry is Symptom → Cause → Fix with a code shape you can grep your repo for.

Moonshot Kimi K2.6 Operational Quirks#

Kimi K2.6 is one of the cheapest competent reasoning models — $0.95/M input cache-miss, $0.16/M cache-hit, $4.00/M output, 256K context. It is also one of the most opinionated. Half of what works on OpenAI breaks here, and the failures are silent: empty content, mid-reasoning truncation, 400 errors that don’t mention the actual problem, and a cache key parameter that makes cost go up instead of down.

OFAT Matrix LLM Tuning#

When a new provider or model lands and you have to decide what temperature, max_tokens, tool_choice, prompt-shape, and turn budget to ship in production, the default is to pick by hunch. Read the model card, copy a partner adapter’s defaults, ship. A week later you find out reasoning_effort=high doubled cost for no quality gain, max_tokens=2048 silently truncated half your tier-3 runs, and the “prompt-rich” pattern you copied from grok-4.3 actively hurts kimi.

Reasoning-Model Tuning Asymmetry#

Practitioners assume “better prompt = better output”. For one model class, that assumption is correct. For the other, the same prompt makes things measurably worse. This article documents the asymmetry, names the dividing line, and gives you a 4-cell test to confirm it on your own canary before you commit to a prompt.

The asymmetry is empirical, not theoretical. It shows up cleanly across four independent OFAT (one-factor-at-a-time) matrices run between 2026-05-18 and 2026-05-20: sonnet POC, grok matrix v1+v2, deepseek matrix v1, kimi matrix v1.

The Self-Ask Trap#

Practitioners ask the LLM about itself as a research shortcut: “What are your common quirks? What temperature should I use? Do you need reasoning_content echoed in multi-turn?” The output looks plausible, often cites specific behaviors, sometimes includes API parameter names. It is often wrong.

The 2026-05-20 kimi-k2.6 tuning research surfaced a clean example. Self-ask said one thing. Documentation, partner adapter source, GitHub issues, and direct API probes said the opposite. The model is provably wrong about itself, and the failure mode is structural — not specific to kimi.