Skip to main content

Choosing an Engine

Benchmarked against: Anthropic โ€” Choosing a model Rule source: Cost Awareness (Company Constitution ยง6) Captain decision: 2026-02-27 โ€” "CP value is NOT cheapest possible but minimum cost that gets the job done RIGHT"

Engine selection is a core operational skill. The wrong choice wastes money (using Opus for trivial tasks) or produces garbage (using free Groq for important analysis). This guide helps match engine to task.


The selection principleโ€‹

Minimum cost that gets the job done RIGHT. โ€” Captain decision, 2026-02-27

This is NOT "use the cheapest engine." It's "use the least expensive engine that produces quality results."

Anti-pattern: Using free Groq for important intel analysis led to hallucinated version numbers, poor quality summaries, and bad meeting notes. The cost of bad output far exceeds the savings of a free engine.


Decision matrixโ€‹

Task importanceEngineCostRationale
Trivial โ€” random searches, cleanup, format conversionGroq (free)$0Not worth spending on
Standard โ€” research, analysis, summaries, translationGemini / DeepSeekCentsQuality matters, cost is minimal
Important โ€” intel analysis, key research, decision supportGemini with citations~$0.014Need authoritative, verified results
Critical โ€” code editing, file operations, system changesClaude (Sonnet/Opus)$$Only engine with full tool access
Architecture โ€” decisions, design, delegation, reviewsOpus (direct)$$$$Worth every token

Role-based engine assignmentโ€‹

Each agent role has a natural engine:

RoleAgentEngineShould doShould NOT do
Chief Engineerๅฐๅ…‹ (Opus)Claude OpusArchitecture, decisions, delegation, reviewsRepetitive coding, data searching
ExecutorๅฐA (Sonnet)Claude SonnetCoding, execution, standard analysisArchitecture decisions
Intel OfficerPatrol workerGroq / GeminiExternal research, web search, translationFile operations, code editing
Couriercron + bashN/A (scripted)Scheduled checks, automationโ€”
Strategistๅฐ่ฅฟ (Chat)Claude SonnetStrategy analysis, reviewsโ€”

Engine selection by task typeโ€‹

Research tasksโ€‹

UB empty on topic?
โ”œโ”€โ”€ Quick background โ†’ Groq Search (free)
โ”œโ”€โ”€ Need citations โ†’ Gemini Search (~$0.014)
โ””โ”€โ”€ Critical decision โ†’ Opus WebSearch (expensive, last resort)

Coding tasksโ€‹

Code change needed?
โ”œโ”€โ”€ Small fix, well-defined โ†’ Sonnet (dispatch_work_order engine=claude)
โ”œโ”€โ”€ Architecture design โ†’ Opus (direct)
โ””โ”€โ”€ No code change, just analysis โ†’ Groq/Gemini (call_model)

Analysis tasksโ€‹

What kind of analysis?
โ”œโ”€โ”€ Simple summary โ†’ Groq (free)
โ”œโ”€โ”€ Need accuracy โ†’ DeepSeek (cheap, strong reasoning)
โ”œโ”€โ”€ Chinese NLP โ†’ Zhipu GLM-5 (cheap, best Chinese)
โ””โ”€โ”€ Need citations โ†’ Gemini (authoritative)

Dispatch engine selectionโ€‹

When creating a Work Order, the engine parameter determines execution:

Engine valueProviderCostCapability
claudeClaude Opus/Sonnet~$1-2/runFull tool access (files, shell, code)
groqGroq Llama 3.3 70BFreeText analysis, summaries
groq-searchGroq CompoundFreeWeb search + analysis
geminiGemini 2.5 FlashCentsGeneral tasks
gemini-searchGemini + Google~$0.014Authoritative search with citations
deepseekDeepSeek R1/V3CentsReasoning, analysis
mistralMistralCentsEuropean alternative
zhipuZhipu GLM-5CentsChinese NLP, tool-calling
ingestMTAAA PipelineFreeBatch file ingestion (no LLM)

Common mistakesโ€‹

MistakeWhy it's wrongCorrect choice
Using Opus for web searchBurns expensive tokens on retrievalDelegate to Groq/Gemini, ingest results
Using Groq for important intelHallucinates version numbers, poor qualityGemini with citations
Using Claude for text analysisOverkill โ€” no file operations neededGroq (free) or DeepSeek
Using any engine without UB checkMay already have the answer in UBAlways search_brain() first
Using engine=claude for research WOs~$1-2 per run for text workengine=groq or engine=gemini-search

The search flowโ€‹

The mandatory information retrieval flow optimizes engine usage:


PageRelationship
Engine OverviewAll engines and capabilities
PricingDetailed cost breakdown
Cost AwarenessGovernance rule
Usage & CostTracking tools