DeepSeek announced on May 22 that the 75% discount on its V4 Pro pricing, originally scheduled to expire at the end of the month, is now permanent. The new sticker rate sits below OpenAI’s GPT-5, Anthropic’s Claude Opus 4.7, and Google’s Gemini 3.5 Flash. The gap is widest on the reasoning-heavy workloads that enterprise customers run when accuracy matters more than throughput.
The pricing move is the latest escalation in a price war that has been compressing flagship-tier rates for the past nine months. The structural pressure comes from two directions at once. From the open-weight side, DeepSeek, Qwen, and other Chinese labs have built models that match Western frontier performance on most enterprise benchmarks while running on hardware that costs a fraction of what a U.S. lab pays for the equivalent capacity. From the demand side, the “advisor model” pattern that CNBC documented in mid-May (where enterprises route routine requests to cheap models and reserve frontier APIs for the final reasoning step) is shifting volume away from premium tiers regardless of price.
The competitive math on Western flagships now looks tight. Artificial Analysis ran identical evaluations across the three flagships and DeepSeek in mid-May: Claude at $4,811, GPT at $3,357, DeepSeek at $1,071. A 75% permanent cut on DeepSeek’s headline pricing further widens that gap to a roughly 7-to-1 ratio between Claude and DeepSeek at the flagship reasoning tier. Even allowing for the quality differential that regulated buyers cite as justification for paying Western rates, the gap is now large enough that an enterprise routing the majority of non-sensitive inference through DeepSeek and reserving Claude for the final 5% of requests can plausibly cut its annual AI spend by an order of magnitude.
Two skeptical notes deserve weight. First, DeepSeek has not disclosed its inference economics, and the price cut may reflect strategic subsidy rather than durable margin. Chinese government interest in AI capability proliferation creates a plausible non-commercial rationale for keeping prices below cost-recovery for an extended period. Second, regulated industries, which CNBC noted earlier this month, will continue to pay frontier prices regardless of the open-weight gap. Healthcare, finance, defense, and government procurement have compliance requirements that DeepSeek cannot satisfy on a timeline that matters for 2026 procurement cycles.
That said, the regulated-industry segment is shrinking as a share of total AI spend. The broader enterprise market is exactly the segment where the advisor-model pattern is taking hold fastest, and that segment is what Western flagships need to win to justify the revenue trajectories they are pricing into IPO valuations. Tomasz Tunguz argued last week that flagship prices are climbing because labs need margin ahead of going public. DeepSeek making its 75% cut permanent is a direct counter-pressure to that pricing strategy, and it forces OpenAI and Anthropic to choose between protecting margin and protecting market share.
The October IPO timing for Anthropic concentrates this question. Anthropic’s path to $10.9 billion in Q2 revenue assumes pricing power on Claude Opus 4.7. If DeepSeek’s permanent 75% discount accelerates customer migration to the cheap tier through the back half of 2026, the revenue projections in Anthropic’s roadshow will be running against a different cost-of-substitution curve than they were in April.
For enterprise procurement teams negotiating multi-year AI contracts now, the practical move is to insist on a price-protection clause that benchmarks Western flagship rates against an open-weight comparator updated quarterly. Without that protection, you are signing a multi-year price floor at exactly the moment the floor is moving.
Reported by The Next Web on 2026-05-22.
