Skip to main content
Every Check Run Agent runs on a model you choose in its front matter. This page covers the available models, how reasoning and effort behave, per-model pricing, and how to keep costs down.

​
Choosing a Model

Set model: in your front matter to choose which model powers the agent. The default model is claude-opus-4-6.
If you omit model:, the agent uses the default (claude-opus-4-6). If you set model: to a value Macroscope doesn’t recognize (or a deprecated one), the check run completes as skipped — neither pass nor fail — with a “model is not available on Macroscope” message on its details page. It never silently falls back to a default.
qwen-3-7-plus is being deprecated on August 27, 2026. Any Check Run Agent configured to use it will fail after that date. We recommend switching to qwen-3-8-max — it’s the newer model in the same family.
  • Default = the value used when the field is omitted from front matter.
  • Effort applies to Anthropic models only — GPT, Grok, and open-source models ignore it (shown as —). On those models every effort level collapses to the same variant, so it has no effect.
  • Newer Anthropic models (Opus 4.6 / 4.7 / 4.8 / 5, Sonnet 4.6 / 5, Fable 5) set thinking automatically, so their reasoning field is ignored (shown as auto).
  • grok-4-6 always reasons, and its four thinking levels are all distinct: low, medium, high, and xhigh each select their own depth. Only off collapses, mapping to low.
  • Open-source models each map reasoning onto their own thinking levels: deepseek-v4-flash always reasons and grades effort (off/low map to low, medium/high to high, xhigh to max); kimi-k3 always reasons and grades effort (off/low map to low, medium/high to high, xhigh to max); kimi-k2-7-code always reasons; glm-5-2 distinguishes off, high, and xhigh (with low/medium collapsing to high); qwen-3-8-max keeps all five levels distinct — off, low, medium, high, and xhigh each select their own depth, and turning it off really does disable thinking; minimax-m3, qwen-3-7-plus, and deepseek-v4-pro are binary — off or on.
  • xhigh is supported only by glm-5-2, kimi-k3, deepseek-v4-flash, grok-4-6, and qwen-3-8-max.
Open-source models are opt-in per agent (the default model stays claude-opus-4-6) and typically cost a fraction of the frontier options — a good fit for specialized, high-volume, or narrowly-scoped checks, while Opus remains the default for reviews that need the strongest model. model: only changes which LLM powers the agent — every other front-matter control (tools, input, include, exclude, and the rest) works the same on every model.
Open-source models are served through Fireworks AI’s hosted inference — Macroscope does not run local or self-hosted models, and agents cannot point at your own endpoints. As we expand the set of inference providers we support, capacity fluctuations from these providers (during heavy load) mean latency may occasionally be higher due to request queuing.

​
Reasoning and Effort

  • reasoning accepts off | low | medium | high | xhigh
  • effort accepts low | medium | high
Which field applies depends on the model:
  • Anthropic — effort always applies. claude-opus-4-5 and claude-sonnet-4-5 additionally honor reasoning, which maps to extended thinking; newer Anthropic models set thinking automatically and ignore reasoning.
  • OpenAI, xAI, and open source — reasoning selects the thinking level; effort is ignored.
xhigh is supported only by glm-5-2, kimi-k3, deepseek-v4-flash, grok-4-6, and qwen-3-8-max. On any other model the run falls back to low and surfaces a warning on the check run details page. For example:

​
Pricing

Check Run Agents bill in credits based on the underlying model’s token cost. Rates below are USD per 1M tokens. claude-opus-4-6 is the default when no model is set.
All prices are USD per 1M tokens. Agent usage is billed in credits: raw model cost + 5% markup, then ÷ $0.01 per credit (1 credit = $0.01). claude-sonnet-5 shows introductory pricing through Aug 31, 2026 (standard afterward: $3.00 / $15.00 / $0.30 per 1M).

​
Managing Costs

To keep costs down: use include/exclude or .macroscope/ignore.md to scope to relevant files, prefer full_diff over code_object, use pr_metadata for checks that only judge the PR’s title, labels, description, or commit messages, and use lower effort for simple checks. Cost information is available in two places:
  • Settings → Billing — Admins can see a full cost breakdown across all Check Run Agents.
  • GitHub Check Run details page — Shows the billed total for each individual run.