Files
tutti/packages/agent/runtimeprep/skill_templates/tutti-model-allocation-model-tiers.md
agent-radar 2ed3415fb7
Publish Claude Code Binaries / Stage and publish binaries (push) Canceled after 0s
Publish Tutti App Runtime / Build darwin-arm64 (push) Canceled after 0s
Publish Tutti App Runtime / Build darwin-amd64 (push) Canceled after 0s
Publish Tutti App Runtime / Build linux-amd64 (push) Canceled after 0s
Publish Tutti App Runtime / Build windows-amd64 (push) Canceled after 0s
Sync Tutti Agent Skills / sync (push) Canceled after 0s
Windows Agent Adapters / Test Windows agent adapters (push) Canceled after 0s
Windows Daemon Adapters / Test Windows daemon adapters (push) Canceled after 0s
Publish Tutti App Runtime / Publish runtime catalog (push) Canceled after 0s
源码快照 @ 2026-09-01 (source-only, 剔除非源码素材)
2026-09-01 20:17:27 +08:00

9.6 KiB

Model Tier Reference

This routing prior combines the packaged OpenSquilla provider ladders at revision d8652b72 (2026-07-28), Tutti live-catalog evidence, and current vendor descriptions. It is neither a benchmark result nor an availability catalog.

Always intersect this table with the exact models returned by the current target's agent composer-options command. Never copy an unavailable id from this file into a plan.

Provider ladders

Provider C0 C1 C2 C3 Vision route
OpenAI gpt-5.6-luna low effort gpt-5.6-luna medium effort gpt-5.6-terra medium/high effort gpt-5.6-sol high/max/ultra; Sol Pro when advertised Use current catalog evidence
OpenAI legacy gpt-5.4-nano gpt-5.4-mini gpt-5.5 medium effort gpt-5.5 high effort Use current catalog evidence
Cursor composer-2.5 fast, overqualified composer-2.5 fast composer-2.5 standard or fast composer-2.5 for long-horizon coding only Current catalog evidence
Cursor Grok grok-4.5 low effort, overqualified grok-4.5 low effort grok-4.5 medium effort grok-4.5 high effort Current catalog evidence
SpaceXAI grok-4.5 low effort, overqualified grok-4.5 low effort grok-4.5 medium effort grok-4.5 high effort grok-4.5
OpenRouter deepseek/deepseek-v4-flash deepseek/deepseek-v4-pro z-ai/glm-5.2 anthropic/claude-opus-4.8; moonshotai/kimi-k3 for long-context coding moonshotai/kimi-k3
DashScope qwen3.6-flash qwen3.7-plus qwen3.7-max qwen3.7-max Use current catalog evidence
Qwen Token Plan qwen3.6-flash qwen3.7-plus qwen3.7-max qwen3.8-max-preview qwen3.7-plus
DeepSeek deepseek-v4-flash no thinking deepseek-v4-flash low thinking deepseek-v4-pro medium thinking deepseek-v4-pro high thinking Use current catalog evidence
Gemini gemini-3.1-flash-lite gemini-3.5-flash gemini-3.1-pro-preview gemini-3.1-pro-preview high thinking Use current catalog evidence
Zhipu glm-5-turbo glm-5 glm-5.1 glm-5.2 Use current catalog evidence
Moonshot kimi-k2.6 low thinking kimi-k2.6 medium thinking kimi-k2.7-code; kimi-k3 low effort kimi-k3 high/max effort kimi-k3
Volcengine doubao-seed-2-0-lite-260215 doubao-seed-2-0-lite-260215 low thinking doubao-seed-2-0-pro-260215 doubao-seed-2-0-pro-260215 high thinking Use current catalog evidence
BytePlus seed-2-0-lite-260228 same model, low thinking same model, medium thinking same model, high thinking Use current catalog evidence
TokenRhythm deepseek-v4-flash deepseek-v4-pro kimi-k2.7-code glm-5.2; kimi-k3 for long-context coding kimi-k3

Differentiated route notes

Use these only when the exact route is present in current composer-options. A tier is task-shape-specific: a coding-specialist C3 route is not automatically the best architecture, research, or final-synthesis route.

Family or exact route Prior tier Prefer for Constraint
gpt-5.6-luna C1 Fast, cost-efficient bounded and everyday coding Use Terra or Sol when task/effect requires C2 or C3
gpt-5.6-terra C2 Balanced everyday implementation, debugging, and integration Do not prefer it over Sol solely because it is the planner's default
gpt-5.6-sol, gpt-5.6-sol-pro C3 Hard implementation, architecture, deep review, and final synthesis Prefer Sol Pro only when advertised; never synthesize a -pro route
composer-2.5 and composer-2.5[...] C2-C3 Fast hands-on coding and sustained long-running implementation Count as C3 only for evidence-backed agentic coding; fast keeps base tier
grok-4.5, grok-4.5[...], x-ai/grok-4.5, xai/grok-4.5 C3 Fast frontier coding, engineering, and tool-heavy agentic work Requested-origin entries are not availability evidence
Kimi K3 family C2-C3 Very long context, large repositories, multimodal and long-horizon coding Low effort is C2; high/max is C3
Claude Haiku family C0-C1 Bounded, latency-sensitive work Use current family generation
Claude Sonnet family C2 Substantial implementation and debugging Use current family generation
Claude Opus family C3 Architecture, deep review, and high-stakes synthesis Use current family generation
Gemini Flash-Lite / Flash / Pro families C0/C1/C2-C3 Fast simple / balanced low-latency / strong multimodal work Use description and reasoning controls to distinguish Pro's top tier

Route aliases

Classify aliases as one family, but preserve the exact value returned by composer-options in the plan:

  • Kimi Open Platform/Codex: kimi-k3; Claude-compatible catalogs may return kimi-k3[1m].
  • OpenAI-compatible plans may namespace GPT routes, for example openai/gpt-5.6-sol-pro; classify the family but preserve the namespaced id.
  • Kimi Code: k3 and k3-256k; the 256K route keeps K3's tier but does not satisfy a task that requires more than 256K context.
  • OpenRouter: moonshotai/kimi-k3.
  • Cursor Composer: parameterized values such as composer-2.5[fast=true]; standard and fast keep the same capability tier.
  • Cursor Grok: preserve bare or parameterized grok-4.5 values exactly; do not substitute a Model Plan's namespaced value.

Family inference

When an exact model is absent from the table:

  1. Prefer its current provider description and advertised capabilities.
  2. Match a known family only when the family and generation are unambiguous.
  3. Treat nano, flash-lite, and equivalent small variants as C0 priors.
  4. Treat mini, flash, lite, and equivalent balanced variants as C1 priors unless provider evidence states otherwise.
  5. Treat pro, max, sonnet, and code-specialized strong variants as C2 priors.
  6. Treat opus, frontier, highest-capability, and explicit high-stakes variants as C3 priors.
  7. Treat turbo as a speed signal, not automatically as a lower capability tier.
  8. Do not give the planning Agent, current provider/model, or a provider default any family-inference bonus.
  9. Do not infer image support, context size, tool support, or availability from the name.

Refresh anchors

When updating this prior, recheck the current runtime catalog and the vendor descriptions for GPT-5.6, Composer 2.5, Grok 4.5, and Kimi K3. Vendor descriptions inform task fit; they never override live availability evidence.