StepFun's Cache-First Pricing: A Land-Grab Price, Not a Steady State

The market doesn't reward cheap reads. It rewards predictable total cost. StepFun's Step 5 Preview, launched with a cache-read price of $0.05 per million tokens, is being framed as a revolution in agent economics. But based on my audit experience across token fund allocations and infrastructure due diligence, this pricing table is incomplete in a way that should concern every developer building persistent agents. Let's cut through the launch noise. The structure is clean: $1.00 input, $2.70 output, $0.05 cache read. That output-to-input ratio of 2.7x sits comfortably in the industry's normal band. The cache read discount at 1/20 of input price matches the pattern Anthropic and Meta have already commoditized. So what's actually new? The absolute price point. $0.05 versus $0.20 at frontier competitors is a 4x gap. That's a real number. But the discount shape is not the differentiator—the absolute level is. Here's the blind spot in the launch narrative. StepFun published input, output, and cache read prices. They did not publish a cache write price. Anthropic, when releasing Opus 5.5, disclosed full write terms: $5.00 per million for standard writes and $8.00 per million for one-hour storage—both above their $4.00 input price. StepFun's omission is not an oversight. It's a strategy. If their write price exceeds input, publishing it would puncture the cache-first marketing story. The two-stage land-grab is textbook: acquire developers with cheap reads, defer write-side pricing and monetization until adoption creates lock-in. We didn't see this asymmetry coming when we first scanned the pricing page. We assumed the absence of write terms was a documentation gap. It's not. It's a deliberate information asymmetry that makes total cost of ownership unmodelable. You cannot compute your real per-agent cost without knowing the write price, the TTL, and the eviction policy. StepFun provided none of those variables. Now consider the traffic-side evidence they did publish. The top applications consuming Step 5 Preview are all persistent, multi-step agents: Hermes Agent at 2.12 trillion tokens, Kilo Code at 1.45 trillion, OpenClaw at 1.07 trillion, Cline at 808 billion, and StepCode at 695 billion. This is a load profile of write-once, re-read-every-step. It's consistent with the cache-first hypothesis. But it also reveals the economic fragility. A codebase or document-level prefix can generate tens of gigabytes of KV cache. If TTL is long, storage costs compound. The $0.05 read price assumes storage is dramatically cheaper than recomputation. That assumption holds only under specific context-length distributions and hit rates—neither of which StepFun disclosed. The technical architecture itself is not groundbreaking. The model is a 600B total parameter, 27B active parameter sparse Mixture-of-Experts. That's a ~22x sparsity ratio, more aggressive than DeepSeek-V3's 671B/37B at ~18x. This signals a pure inference-cost priority over capability. Prefix caching is a mature mechanism—implemented via RadixAttention in SGLang or automatic prefix caching in vLLM. StepFun described the mechanism but provided no implementation details, no benchmark against SOTA on reasoning, code, or long-context tasks. The innovation is engineering-level and combinatorial-level, not architectural. The "Preview" tag and OpenRouter distribution confirm this is a POC transitioning to production—commercially available but without SLA commitments. The contrarian angle here is sharper than the launch materials suggest. The market's narrative reads the $0.05 cache read as a permanent price floor. I read it as a promotional price designed to capture a single workload class. The three missing variables—write price, TTL, eviction policy—determine whether this pricing is actually sustainable. Without them, StepFun can adjust any term at any time. The "Preview" label is not a technical disclaimer; it's a pricing flexibility clause. Developers who architect their cost model around $0.05 reads are building on sand. License transparency compounds the risk. StepFun has not published its model license. That's either pending legal review or intentional commercial flexibility. Either way, enterprise adoption faces unquantifiable legal uncertainty. In a bull market where FOMO drives API integration decisions, this is the kind of detail that gets skipped. It shouldn't be. My takeaway is forward-looking. The cache-first economics will work for a specific load profile: long-context, multi-step agents with high hit rates and short TTLs. That's a real segment. But the pricing is a land-grab price, not a steady-state price. The real question is not whether StepFun can attract developers at $0.05—they already have. The question is what happens when they publish the write price. If it follows the Anthropic pattern, the TCO for persistent agents shifts materially. The developers who win are the ones who model their costs with the write price already included, not the ones who wait for the other shoe to drop.