Chinese AI startup DeepSeek ignited a pricing war in the large language model (LLM) market with two consecutive price cuts within 48 hours, pushing its V4-Pro model's cache-hit input cost to just RMB 0.025 per million tokens (US$0.0035)—roughly one-fortieth of its original price. The aggressive move, coupled with the April 24 launch of the open-source DeepSeek-V4, sent API call volumes surging nearly fourfold on third-party routing platforms and threatens to upend pricing expectations for agent-optimized models industrywide.

The timing underscores a strategic pivot: as AI transitions from conversational interfaces to autonomous agents executing multi-step workflows, token consumption—and therefore cost—has become the bottleneck for enterprise adoption. DeepSeek's offensive directly targets this pain point, forcing rivals including OpenAI's GPT-5.5, Anthropic's Claude Opus 4.7, and domestic competitors Kimi K2.6 and GLM-5.1 to reassess their pricing anchors.

Back-to-Back Discounts Collapse Cost Floor

On April 25, DeepSeek announced a limited-time 75% discount on its V4-Pro model API, effective through May 5, 2026. Twenty-four hours later, it extended a permanent 90% reduction on cache-hit input pricing across its entire API suite, with Pro models receiving stacked discounts during the promotional window.

The combined effect is dramatic. Before discounting, DeepSeek-V4-Pro's cache-optimized input cost was RMB 1.00 per million tokens; after both cuts, it dropped to RMB 0.025. The Flash variant, designed for lower-latency tasks, now charges RMB 0.02 per million tokens for cached inputs—equivalent to US$0.0029.

By comparison, processing one million input tokens plus one million output tokens costs US$35 on GPT-5.5 and US$30 on Claude Opus 4.7 at list prices. DeepSeek-V4-Pro's equivalent workload, without cache optimization, runs US$5.27—one-seventh of GPT-5.5 and one-sixth of Claude. With caching enabled, the total drops to US$3.66, or roughly one-tenth of leading Western models.

"DeepSeek is essentially subsidizing agent adoption," said Hu Yanping, distinguished professor at Shanghai University of Finance and Economics. "Agent workloads consume tokens at exponentially higher rates than chatbot queries. By undercutting rivals on this dimension, DeepSeek is directly courting enterprise developers and automation-focused users."

Call Volumes Spike 300% Post-Launch

Data from OpenRouter, a global AI model API aggregator, confirms immediate market response. On April 25—the first full day after V4's debut—DeepSeek-V4-Pro logged 13.6 billion tokens in API calls, a 297% jump from the prior day. V4-Flash recorded 50.2 billion tokens, up 85.9%.

By April 26, following the second price cut announcement, Flash's volume climbed to 81.4 billion tokens (a further 62.2% gain), though Pro's usage dipped to 9.6 billion. Neither model yet appears on OpenRouter's weekly or daily top-call rankings as of April 27, suggesting early traction remains concentrated among early adopters testing cost arbitrage opportunities.

Still, the trajectory signals a potential inflection. Unlike prior DeepSeek releases focused on benchmarking against GPT-4 or Claude in conversational tasks, V4 explicitly targets agent capabilities—including long-context reasoning (1 million tokens), tool invocation, and automated code execution. Third-party evaluations by Artificial Analysis scored V4-Pro at 52 on its Intelligence Index, second only to Kimi K2.6 (a Chinese rival) and surpassing V3.2's 42. V4-Flash, at 47, aligns with Claude Sonnet 4.6 performance levels.

Pressure Mounts on Domestic Rivals

The pricing offensive arrives as Chinese LLM providers—including Moonshot AI (Kimi), Zhipu AI (GLM), Alibaba (Qwen), and MiniMax—have collectively raised API fees over the past quarter amid rising inference compute costs and investor pressure to demonstrate unit economics.

DeepSeek's subsidized rates now force a strategic dilemma: match prices and risk margin compression, or maintain premiums and risk developer attrition. Hu noted that high-performance domestic models could face "sustained pricing pressure," particularly if DeepSeek scales inference infrastructure through partnerships with domestic cloud providers or chipmakers navigating U.S. export controls.

For Western incumbents, the impact may prove more contained. OpenAI and Anthropic derive pricing power from brand trust, regulatory compliance for enterprise clients, and ecosystem lock-in via fine-tuned models. DeepSeek's advantage centers on cost-sensitive segments—startups, academic researchers, and emerging markets where budget constraints outweigh compliance requirements.

However, the cache-hit pricing innovation itself merits attention. By charging one-tenth standard rates for repeated content (common in agent workflows with persistent system prompts), DeepSeek exploits a technical efficiency that others have implemented but not aggressively monetized. If developers optimize workloads accordingly, the effective cost gap versus GPT-5.5 or Claude could widen beyond headline comparisons.

Agent Economics Reshape Model Competition

The broader context reveals a market transitioning from foundational model performance to application-layer cost efficiency. Where 2024–2025 competition centered on MMLU scores and human preference benchmarks, 2026 increasingly prioritizes dollars-per-completed-task metrics for real-world automation.

Agent workloads differ from chatbot queries in token consumption profiles: a single autonomous task—say, debugging code, researching regulatory filings, or planning travel—might invoke 50,000 to 200,000 tokens across multiple LLM calls, tool accesses, and iterative refinements. At legacy pricing (US$15–30 per million tokens for input), monthly costs for active agent deployments could exceed thousands of dollars per user.

DeepSeek's intervention collapses that barrier. A developer running 10 million agent tokens monthly would pay US$146 at standard V4-Pro rates, or US$37 with caching—versus US$1,500+ on premium Western models. This arithmetic explains why OpenRouter traffic spiked despite V4's modest benchmarks relative to Claude Opus 4.7 or GPT-5.5.

Sustainability Questions Linger

Whether DeepSeek can sustain such pricing remains uncertain. The startup has not disclosed inference infrastructure details, though industry observers speculate heavy reliance on Huawei Ascend 910B accelerators (China's flagship AI chip under U.S. sanctions) and cost-sharing arrangements with state-backed data centers.

Hu suggested that if domestic inference capacity expands significantly—potentially through government-subsidized compute programs—then sub-cent token pricing could stabilize as a competitive baseline for Chinese models. "But this assumes continued buildout of inference clusters and willingness to operate on thin margins," he cautioned.

For now, DeepSeek's 75% discount expires May 5, though the permanent cache-hit reduction persists. Market participants will watch whether rivals respond with matching cuts (signaling commoditization) or double down on differentiation through accuracy, safety features, or vertical-specific fine-tuning.

The immediate takeaway: agent economics have entered the spotlight, and DeepSeek just reset the cost floor—at least within China's rapidly evolving AI ecosystem.

Related Coverage：

DeepSeek Unveils V4 Preview With Million-Token Context Window