How to Cut AI Costs by 80%: What Port.io's 12,000-Query Study Found
Cloud Platform EngineeringMost AI cost discussions focus on model selection. Port.io’s research found the bigger lever is how you structure the context you send, not which model receives it.
What the Study Found
Port.io ran 12,000 queries across Claude Haiku, Sonnet, and Opus, comparing three context architectures: MCP-only (unstructured), a Context Lake (structured), and a Context Lake with skill files added.
The cost differences were stark. Against unstructured MCP-only setups, the Context Lake approach cut costs by approximately 80% across every model tested. In raw numbers: Haiku dropped from $0.087 to $0.018 per query; Sonnet from $0.333 to $0.059; Opus from $1.761 to $0.354.
The reasoning is structural. Unstructured context generates high token counts because agents spend tokens exploring context they don’t need before finding what they do. A Context Lake — where knowledge is organized and retrievable in structured form — means agents surface relevant context faster, execute in fewer tokens, and reach correct responses more directly.
The Skill File Nuance
The study result that deserves specific attention: adding skill files to an MCP-only setup made costs worse. Haiku costs increased 18%, Sonnet 13%, Opus 24%. The reason: agents treated skill files like checklists and executed every step regardless of whether each step was relevant to the actual task.
In a Context Lake setup, skill files worked as intended — guiding execution without replacing agent judgment. The lesson isn’t that skill files are bad; it’s that skill files only perform correctly when the context they operate in is already well-structured.
According to The New Stack’s coverage, the Port.io team identified messy context as “one of the biggest unsolved problems in enterprise today” — organizations have large knowledge repositories that agents need, but without structure, those repositories become expensive noise rather than useful signal.
This complements the model selection framework covered in the Claude Fable/GPT-5.5 triage analysis: choosing the right model matters, but sending that model 10× as many tokens as needed cancels any per-token cost advantage.
What Engineering Teams Should Do
The practical implication is a sequencing decision: before scaling up AI agent deployments, invest in context architecture first. A team running on unstructured MCP-only today and planning to increase query volume will hit cost walls that better tooling alone won’t solve. Structuring the context layer first means cost efficiency scales linearly with query volume instead of against it.
The Context Lake concept isn’t proprietary to Port.io. It describes a general architectural pattern: indexed, retrievable, structured knowledge stores that agents can query specifically rather than receive as a flood.
The So What
Run a cost audit before your next model or infrastructure decision. If your AI query costs are scaling faster than your usage, the problem is likely context structure, not model pricing. The 80% number comes from a specific comparison, but the principle generalizes: structured context is less expensive to reason over than unstructured context. Fix that first, then optimize for model selection. See Datadog’s Agentic AI Spark research for how observability tooling captures the cost signal you need to diagnose this.
Source: The New Stack
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →