Mail

When do you first feel the limits of an AI coding tool? Usually when you see "result truncated" — or when you have no idea how much you've spent this month.

Claude Code v2.1.91 and v2.1.92 target exactly those two frustrations. The first expands the MCP tool result limit to 500,000 characters. The second introduces the /cost command for real-time cost tracking. Released back-to-back in the first week of April 2026, here's why these updates matter.

What is MCP 500K?
What Changes vs. Before
The /cost Command: Finally Seeing Your Spend
Defer Permission: New Control for Headless Sessions
Practical Tips and Educational Perspective

1. What is MCP 500K?

What it actually means for tool results to stop getting cut off.

MCP (Model Context Protocol) is how Claude Code exchanges data with external tools — file systems, databases, APIs, web search. Each tool response has always had a maximum size limit. Before v2.1.91, that limit was a wall developers hit regularly.

Claude Code MCP 500K update

In v2.1.91, the tool result limit was expanded to 500,000 characters. Using it is straightforward:

{
  "_meta": {
    "anthropic/maxResultSizeChars": 500000
  }
}

Add this metadata to your MCP tool response, and you can pass up to 500,000 characters per result. The difference is most visible when analyzing large codebases, processing long documents, or handling complex API responses.

2. What Changes vs. Before

No more "result truncated" in the middle of a complex analysis.

Common problems caused by the previous MCP result size limit:

Large code file analysis cutting off mid-response
Critical data lost from long API responses
Complex database query results processed incompletely
Developers splitting work across multiple runs to work around limits

How large is 500,000 characters? Roughly 300–400 pages of plain text. Anthropic has noted that "most real-world tool responses comfortably fit within 500K characters." The limit is now permissive enough that you should rarely hit it.

Dimension	Before	After v2.1.91
Max MCP result size	Restrictive	500,000 characters
Large file analysis	Requires splitting	Single-pass
Long API responses	Truncation occurs	Fully processed
Configuration	Default value	Explicit via metadata

3. The /cost Command: Finally Seeing Your Spend

In AI development, costs you don't see are the most dangerous kind.

The /cost command added in v2.1.92 shows a breakdown of costs incurred during the current Claude Code session — by model, by cache hit status. Type /cost in the terminal and you immediately see: total tokens used, cost allocation by model, and cache utilization efficiency.

Why does this matter? Two reasons.

First, predictable cost management. Running complex agentic tasks feels uncertain when you don't know what they cost. /cost resolves that anxiety in real time. You can estimate "this task will cost approximately X to finish."

Second, cache optimization feedback. Claude Code can dramatically reduce repeat-request costs through prompt caching. /cost shows how many cache hits occurred in the current session. Use that data to adjust your workflow and improve cost efficiency.

"When you can see your costs, you can build a strategy. When you can't, you just spend."

4. Defer Permission: New Control for Headless Sessions

In automated AI workflows, "hold on" is now possible.

Defer Permission, added in v2.1.89, allows a PreToolUse hook to pause (defer) tool execution under specific conditions. In headless (automated, no-UI) sessions, when the AI is about to run a particular tool, a human can now intervene to review and then resume.

Use cases:

Requiring human review before modifying production files
Cost and permission confirmation before external API calls
Inserting an approval step when accessing sensitive data

How it works: The -p --resume flag resumes a deferred session, and the hook re-evaluates. This gives you a configurable balance between full automation and human oversight.

5. Practical Tips and Educational Perspective

Knowing about an update matters less than knowing how to use it.

Here's how these three updates translate to real usage patterns.

Using MCP 500K:

When analyzing large codebases in a single pass, set maxResultSizeChars in the metadata
Useful when fetching large result sets from external data sources
Combine multi-step work that previously required splitting into a single execution

Using /cost:

Check /cost before and after a task to compare
Monitor cache hit rate to optimize repetitive workflow patterns
Use as a team baseline for establishing AI development cost norms

From an EdTech perspective, transparency features like /cost matter in educational settings. When teachers and students use AI tools for coursework or research, developing an intuition for "this kind of use costs approximately this much" is part of AI literacy. Understanding the cost structure of your tools — not just how to use them — is what separates casual users from informed ones.

Closing Thoughts

Claude Code's April updates went in the direction of removing practical friction rather than adding flashy new capabilities. Raise the limits. Make costs visible. Return control to the human. That's what a maturing tool looks like.

Next time you're working in Claude Code, try typing /cost. You might find you're spending more — or less — than you expected.

Related Posts

Have you ever been surprised by how much an AI coding session cost? Try /cost and share what you find in the comments!

Sources:

Claude Code MCP 500K Update — Breaking AI Dev Cost and Context Limits at Once

Table of Contents