Cache Efficiency Is the New Gas Fee: Why ZCode's 98.6% Hit Rate Matters More Than Model Hype

The most useful API telemetry this week did not come from a model card. It came from a 48-hour traffic snapshot posted by OpenCode co-founder Dax Raad. For DeepSeek traffic, Zhipu's ZCode hit a cache-hit rate of 98.60%. OpenCode V2 finished at 97.86%. Claude Code/CLI trailed at 89.31%. Dax was honest enough to say: "I don't know what ZCode is, but it's doing a really good job." Hype is noise. Standards are signal. DeepSeek prices input tokens with a brutal multiplier. A cache-miss input token costs roughly 50 times a cache-hit input token. That single multiplier transforms a one-point difference in cache behavior into a large difference in effective input cost. ZCode is an Agentic Development Environment launched by Zhipu for GLM-5.2. The critical detail is that it also supports connection to other models. This is not a closed benchmark. It is a client that can call DeepSeek, and someone measured what happens when it does. An agentic development environment reads files, runs commands, calls tools, edits code, and keeps a long conversation context alive. Every one of those actions is a chance to reuse or rebuild the context prefix. Cache hit rate measures how disciplined that reuse is. A high hit rate means the client keeps its system prompt, tool definitions, and active context stable across calls. It is not buying the same intellectual context from scratch every time. Let me formalize the napkin math. Assign 1 to the cost of a cache-hit input token and 50 to the cost of a cache-miss input token. Effective input cost equals hit rate times 1 plus miss rate times 50. Claude Code/CLI: 89.31% hit gives 0.8931 plus 5.345, or 6.238 units. OpenCode V2: 97.86% hit gives 0.9786 plus 1.070, or 2.049 units. ZCode: 98.60% hit gives 0.9860 plus 0.700, or 1.686 units. Same token structure. Same provider. The only meaningful variable is the client. ZCode's effective input cost is only about 27% of Claude Code's. OpenCode V2 is about 33% of Claude Code's. And OpenCode V2 is 21.5% more expensive than ZCode despite being only 0.74 percentage points behind in hit rate. That is not a small efficiency gain. In a world where agentic runs last hours and consume millions of input tokens, the cost per task is the difference between a product that survives a bear market and one that quietly burns its treasury on API bills. I built and audited decentralized finance systems during the 2020 yield boom. One of the most common failures I found was not smart contract reentrancy or phishing. It was gas waste caused by inefficient contract design. Teams obsessed over total value locked, but the cost per transaction was eroding their margins. The same disease is now visible in AI coding agents. Everyone wants to talk about benchmark scores. Few want to talk about the cost per completed task. What separates the top two clients from Claude Code is the missed-token fraction. ZCode's miss rate is 1.40%. OpenCode V2's is 2.14%. Claude Code's is 10.69%. That means Claude Code is sending more than seven extra percentage points of every input token through a 50x charge. No model intelligence compensates for a cost structure that triples the input bill. Now the contrarian view: do not run to ZCode because one 48-hour snapshot beat another. Cache hit rate is not a pure measure of engineering quality. It is also a measure of workload shape. If ZCode users mostly execute repeated tasks with fixed instructions, the hit rate will be high. If Claude Code users are more often exploring novel codebases, designing new architectures, and refactoring large projects, the future context is unique by definition. Unique context cannot be cached. A sophisticated agent that tackles harder problems can produce a lower hit rate and still be the better product. ZCode's performance is impressive. It is not proof of universal superiority. The real insight is structural. Cache discipline is measurable, manageable, and expensive to ignore. Every agentic client should expose a cache telemetry dashboard. Every team running agents should track effective input cost per completed task, not raw hit rate. In my 2017 ICO work, I built due diligence checklists that rejected projects because they could not define token utility in precise terms. I would reject any AI client that cannot define its prompt reuse strategy. Structure wins. Chaos loses. In a bear market, cost discipline is the only strategy that does not depend on a bull run. This data slice from Dax is useful precisely because it arrived by accident. No vendor ran a carefully curated benchmark. A co-founder looked at his own telemetry and saw someone else beating him. That is real evidence. The next step is for every team using DeepSeek to instrument the same metric. Log cache hit rate by client, by task type, by repository. Then compare cost per merged pull request or cost per successful test pass. The battle for agentic AI will not be decided by a leaderboard of model reasoning scores. It will be decided on unit economics. Cache hit rate is the new gas price. Effective cost per task is the new total value locked. I would rather operate at 98.6% hit rate with a slightly weaker model than at 89.3% with a stronger model, because the cheaper agent can run more iterations and learn faster. The market always rewards survival first. ZCode just showed the market what disciplined architecture looks like. Verify everything. Trust the protocol. Compliance is the new crypto currency, and cache efficiency is the first compliance audit.