Claude Code 2.1.296 halves Sonnet 5.5 cache reads to $0.10/M
2.1.296 also brings single-call big-file reads, a workflow-subagent model pin, double-size MCP tool descriptions, and a trending HN agent-debug tool.
Copy markdown
Half-price cache reads, everywhere Claude Code counts cost
Sonnet 5.5 cache reads drop to $0.10/M from $0.20/M, now reflected in /cost, the status line, --max-budget-usd, and SDK cost figures. Long agent runs and big cached system prompts cost roughly half to re-read.
Read a whole big file in one call
A new allow_large option on the Read tool pulls a text file past the usual size cap in a single call. Fewer chunked reads when an agent needs the entire file at once.
Pin one model for every workflow subagent
CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL runs all workflow agents on a single model while other subagents keep theirs, and autoCompactWindow lets a subagent compact earlier than the main thread. Cheaper, more predictable fan-out.
MCP tool descriptions get twice the room
The up-front MCP tool-description and server-instruction limits double from 2,048 to 4,096 characters (tool-search descriptions now cut at 16,384). Richer tool docs reach the model without truncation.
Your agent can now draw on your screen
Trending on Hacker News (379 points): big-arrow-on-the-screen lets an agent paint arrows, boxes, and labels over your display. A quick way to see what a computer-use or browser agent is actually doing, and weekend-sized to wire in.