DeepSeek raised its peak API prices by nearly 50% on May 10, 2026. Hours later, Zhiyu GLM-5.3 appeared with a price tag that was almost identical — ¥8 in, ¥28 out versus V4-Pro's ¥9/¥27. The difference is a single yuan per direction. That's less than the cost of switching a single API key.
This isn't a price war. It's a surgical repositioning of two competing inference stacks, and the signal for anyone building autonomous agents onchain is deafening: the cost of intelligence is no longer a differentiator. The architecture underneath the pricing is.
Context: The Two-Stage Competitive Landscape
DeepSeek V4 has been the default choice for Chinese Web3 developers since its launch in late 2025. Its open-weight philosophy, aggressive pricing, and strong coding capability made it the go-to model for smart contract auditing, MEV strategy backtesting, and on-chain agent workflows. Zhiyu's GLM series, meanwhile, dominated the B2B enterprise market but struggled to break into the developer-first crypto space.
That changed when DeepSeek announced a price increase on its V4-Pro tier. The timing was peculiar: a market where AI API tokens are becoming commoditized, raising prices seems counterintuitive. But the fine print reveals a layered strategy. DeepSeek introduced off-peak pricing (50% discount during low-traffic hours) and an aggressive cache-hit rate of ¥0.15 per million tokens — roughly 1/60th of the peak input price. Zhiyu responded with GLM-5.3, which matched the peak pricing while claiming superior performance on nine agent benchmarks, seven of which it won.
From a blockchain infrastructure perspective, the parallel is obvious: just as Layer2s fragment liquidity, these two models are now fragmenting the developer mindshare. The question is which one will survive the coming consolidation.
Core: Code-Level Dissection of the Pricing War
Let me walk through the numbers as if I were auditing a smart contract. The unit economics are brutal.
Peak input pricing: DeepSeek V4-Pro at ¥9/M tokens, GLM-5.3 at ¥8/M. Output: ¥27 vs ¥28. The delta is ¥1 per direction. For a typical coding agent task — say, auditing a Uniswap v3 fork with 5M input tokens and 0.5M output — the cost difference is ¥4.5 in favor of GLM-5.3. That's negligible compared to the developer's opportunity cost of switching toolchains.
Cache pricing: Here the gap is a canyon. DeepSeek charges ¥0.15 per million tokens on cache hits during off-peak, and ¥0.3 during peak. Zhiyu charges ¥2. That's a 13x multiple at peak. DeepSeek's cache discount relative to normal input is 1/60th; Zhiyu's is 1/4th. This is not a pricing difference — it's a structural advantage in inference infrastructure.
Based on my experience auditing protocol fee structures, a cache-to-normal ratio below 1/10 suggests the operator has achieved near-marginal cost for cache reads. DeepSeek's 1/60 ratio implies they have optimized their KV-cache reuse to the point where the marginal cost of serving a cached request is approaching zero. This is the equivalent of having a Layer2 with infinite block space during low demand. Zhiyu, by contrast, is still pricing cache as a premium feature — possibly because their attention mechanism doesn't allow efficient prefix reuse, or their cluster is not yet optimized for speculative execution.
Off-peak pricing: DeepSeek's 50% discount during off-peak hours is a smart demand-shaping mechanism. It's akin to Ethereum's base fee mechanism adjusting to block demand. For developers running non-time-sensitive batch jobs — like backtesting historical MEV data — this is a killer feature. Zhiyu doesn't offer any off-peak discount, which limits its appeal to cost-sensitive bulk users.
Now, let's examine the benchmark claims. Zhiyu's published comparison shows GLM-5.3 winning 7 of 9 agent-specific benchmarks. The wins are narrow: 66.9 vs 62.7 on DeepSWE, 28.5 vs 25.7 on HLE with Tools, 84.5 vs 83.3 on Agent's Last Exam. These are within statistical noise for most real-world use cases. DeepSeek actually leads on NL2Repo and Toolathlon. The selective framing — only agent benchmarks, no general reasoning, math, or multilingual tasks — is a red flag. It's like a DeFi protocol cherry-picking TVL metrics while ignoring its impermanent loss curves.
During the 2020 DeFi Summer, I spent six weeks deriving impermanent loss curves using stochastic calculus. I learned that narrow metrics can deceive. The same applies here: if Zhiyu were truly superior across the board, they would have published a broader set of results. The fact that they didn't suggests a weakness in general intelligence that DeepSeek could exploit.
Contrarian: The Infrastructure is the Real Winner
While the narrative pits DeepSeek against Zhiyu, the underlying signal is that inference infrastructure is becoming the moat, not the model. DeepSeek's cache pricing and off-peak discounts reveal a company that has invested heavily in engineering its GPU cluster, scheduling algorithms, and cache architecture. Zhiyu's model may be marginally better on agent benchmarks, but its infrastructure is playing catch-up.
This mirrors the early Layer2 wars: Optimism and Arbitrum had similar theoretical throughput, but the one with better sequencer architecture and lower costs won the TVL war. DeepSeek is the Arbitrum of AI models — it may not have the flashiest benchmarks, but its execution layer is more efficient.
Furthermore, the price increase may actually be a positive signal for DeepSeek's sustainability. In the 2021 NFT mania, I ignored Bored Apes to analyze EIP-1559's fee market dynamics. I found that burn mechanisms introduce non-linear deflationary pressures during low traffic. DeepSeek's pricing strategy is analogous: by raising peak prices while lowering off-peak and cache, they are creating a deflationary incentive for developers to optimize their API usage patterns. This increases stickiness. Once a developer builds their agent pipeline around DeepSeek's cache and off-peak schedule, switching becomes painful.
Zhiyu's response — matching the price but not the infrastructure — is a short-term tactical win but a long-term strategic risk. They are betting that model quality will override cost. But in the Web3 world, where margins are razor-thin and autonomous agents run 24/7, cost efficiency often trumps marginal capability gains.
One blind spot: the benchmark contamination risk. DeepSeek has been open-weight for a long time, meaning its architecture is widely studied. Zhiyu keeps its model weights private. If Zhiyu trained GLM-5.3 on leaked DeepSeek data, the benchmark comparisons become meaningless. I've seen this happen in smart contract audits — copy-paste code with slight modifications can pass unit tests but fail integration. The same applies to neural networks.
Takeaway: The Coding Agent Market is About to Fragment
Entropy wins. Always check the fees. The current equilibrium — DeepSeek and Zhiyu within ¥1 of each other — will not hold. Within three months, I expect one of two outcomes: either DeepSeek releases a V4.1 with improved agent benchmarks to reclaim the narrative, or Zhiyu slashes its cache pricing to match DeepSeek, triggering a race to the bottom in infrastructure investment.
For developers building on-chain agents, the takeaway is clear: do not lock into a single provider. Build a model-agnostic routing layer that can switch between DeepSeek and Zhiyu based on cost and latency. The next six months will see several new entrants — Alibaba's Qwen, ByteDance's Doubao, and possibly even a decentralized AI inference network — all fighting for the same developer wallet. The one that offers the lowest total cost of ownership, not just the lowest API price, will win.
2017 vibes. Proceed with skepticism. The ICO mania taught us that infrastructure is easy to build, hard to scale, and even harder to maintain. DeepSeek's cache pricing is the equivalent of a well-audited smart contract — it's boring, but it works. Zhiyu's benchmark wins are the flashy frontend that may hide a leaky backend. I know which one I'd trust with my agent's runtime.
