Context is the new compute bill: your tools cost more than your model

Every tool definition is a tax

OpenRouter's analysis of what MCP servers, CLIs, and skills cost in context starts with a fact most agent builders never see on an invoice: every tool definition provided to an AI agent is sent to the model on every request and incurs input token costs. MCP servers can add tens of thousands of tokens to an agent's context window. On Claude Sonnet 5.5, the GitHub MCP server's 46 default tools add 19,436 tokens to every request — before the model does anything. A single GitHub release object returned as a tool result cost 20,400 tokens on the same model. The tools are the context now. m:…5516

The same analysis shows how cheap the fix is. In a GitHub repository test, CLI tools used 1,421 tokens compared to 59,417 for an MCP server, because CLI tools rely on the model already knowing the tool. Tool Search, which defers loading tool definitions until needed, reduces the up-front cost of 46 tools to under 600 tokens. Scoping the GitHub MCP server to two tools cut the first request from 19,487 tokens to 1,406. The price also varies by model: the same 46 tools cost 19,436 tokens on Claude Haiku 5.5 but 8,903 on GPT-6 Luna. Prompt caching lowers the price of tokens but does not remove them from the context window, so caching is a discount, not a cure. m:…5516

Less context, better answers

Iris.ai makes the same argument from the enterprise side, with bigger numbers. The company defines Minimum Complete Context as the smallest set of information that holds everything a task needs, and says its approach reduces context size by 100 to 1,000 times compared to connect-everything setups, with overall costs falling by 10 to 100 times and request processing time dropping from 10 minutes to 30 seconds. Connecting every enterprise data source into one large model context, the company argues, increases token costs, latency, and error rates. Its framework builds context in four steps: organizing data at the source, normalizing content, retrieving task-specific context, and live evaluation. m:…5109

The most interesting claim is about model size. Iris.ai says smaller 30-billion-parameter models can replace trillion-scale models when using Minimum Complete Context, because the smaller model is no longer drowning in irrelevant context. Its live evaluation framework checks every model call for instruction following, context usage, and business rule compliance, at a cost the company puts at 1/100th of an extra LLM call. These are the company's own numbers and should be read as such — but the direction of the argument matches OpenRouter's independent measurements. Context is the scarce resource, not parameters. m:…5109

Memory should carry receipts

Zep, a context-graph vendor, attacks the precision problem from another angle: entity types. Flat ontologies, it argues, make agent context either too vague or too brittle, forcing a trade-off between overly broad types and overly narrow ones. Its new type hierarchies let context graphs assign both a broad type and the most specific subtype to an entity, with up to four subtype levels below a top-level type. Classification happens in two stages — top-level extraction followed by a subtype walk — and the feature is currently in closed beta. The idea is the same as Iris.ai's: give the model less, but give it exactly the right less. m:…7597

In a preprint, researchers add the missing piece: memory that can be checked. WikiTrace links distilled lessons to their original execution trajectories, allowing agents to perform evidence checks when encountering ambiguous or conflicting information. The system reduces memory-induced degradation — the failure mode where agents fail because they rely on incorrect or misapplied past lessons — and outperforms existing memory approaches across five agentic benchmarks. The through-line across all four stories is simple: an agent's context is a budget, and the teams that treat it like one are the ones whose agents stay affordable and accurate. m:…9439