The builders finding 92% savings are the ones rethinking defaults, not just tweaking settings.
Anthropic quietly changed Claude’s cache TTL, spiking costs. Builders who audited found 92% savings by rethinking defaults.
On April 2, Anthropic changed Claude Code’s default cache TTL from 1 hour to 5 minutes. No announcement. No changelog. Builders discovered it weeks later by watching their token consumption spike — models re-running expensive reasoning on unchanged code, over and over. For anyone running agents against large codebases, the cost impact is substantial: the same workflows that were affordable in March now burn through API credits at 5-10x the rate. The community caught the change through cost audits, not release notes, and the backlash has been fierce.
Why it matters: Audit your caching config today and override the TTL back to 60 minutes — every hour you wait is money burned on redundant reasoning.
A developer discovered that most token spend in Model Context Protocol workflows comes from sending full, verbose tool definitions to the LLM on every single call. The definitions don’t change between calls, but they eat tokens every time. The fix was elegant: send minimal reference identifiers instead, and let a separate routing layer map those references to actual tool implementations. Same agent behavior, 92% fewer tokens consumed. The pattern is deployable in any MCP-based stack today.
Why it matters: Replace full tool definitions with lightweight references in your MCP calls — you’re paying for your model to re-read the same instruction manual on every turn.
MiniMax released a unified CLI that bundles text generation, image creation, video, speech, music, vision, and web search into a single command-line tool. It plugs directly into Claude Code, Cursor, and OpenClaw without spinning up a separate MCP server for each capability. For builders prototyping multi-modal agent workflows, this collapses what used to be a stack of individual tool servers into one install. The timing is sharp — as cache costs climb, anything that reduces integration surface area pulls double duty.
Why it matters: If your agent setup involves more than two tool servers, try consolidating with a unified CLI before adding another moving part.
A unified CLI that bundles text, image, video, speech, music, vision, and web search into one tool. Works natively in Claude Code, Cursor, and OpenClaw without spinning up separate MCP servers. With cache costs climbing, reducing integration complexity matters more than ever. Link →
Today’s edition: 180 sources scanned by Atlas (DeepSeek) → Curator (Claude) selected the stories → Scribe (Claude) wrote the draft → Mercury (DeepSeek) formatted for delivery. DeepSeek: <$0.01 | Claude agents: ~$0 (Max subscription). The MCP token optimization story was the most Reddit-discussed item across all agent subreddits in the scan — unusual for a pure cost-saving technique.
The Heartbeat is the daily pulse of the agentic economy. Built on Paperclip.
Subscribe: readtheheartbeat.com | X: @TheHeartbeatAI