DeepSeek 75 Percent Price Cut Permanent AI Inference Pricing War 2026 became the defining backdrop to DeepSeek’s permanent 75 percent price cut on its V4-Pro model, announced May 26, 2026, is the AI industry event this week that received the least mainstream coverage and has the most serious long-term implications for the frontier AI companies currently racing to go public. AI Weekly’s May 26 bulletin confirmed the cut: DeepSeek makes 75 percent V4-Pro price cut permanent, escalating the inference pricing war. DeepSeek’s V4-Pro, which reaches approximately 90.5 percent on GPQA Diamond scientific reasoning, within a few points of Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 on that specific benchmark, now costs $0.27 per million input tokens. Claude Opus 4.8, which launched on May 28 at unchanged pricing, costs $5 per million input tokens. GPT-5.5 costs approximately $3 per million input tokens. The ratio between the premium Western frontier models and DeepSeek’s most capable offering is not a rounding error or a marginal efficiency difference. It is an 18.5-times gap on price for near-comparable capability on the benchmarks that enterprise buyers evaluate most heavily. DeepSeek is not the only Chinese model making this aggressive move: Kimi K2.6 costs $0.95 per million input tokens with a 90.5 GPQA score; Zhipu GLM-5.1 costs $0.44 per million with strong performance on coding tasks. The permanent nature of DeepSeek’s cut signals that the Chinese AI labs have concluded that cost-competitive performance against Western frontier models is not a temporary promotional strategy but a permanent market position.
What Permanent Means — The Pricing War Has a New Baseline
The distinction between a temporary and a permanent price cut is the most commercially significant aspect of the DeepSeek announcement. Temporary promotional pricing is a customer acquisition tool that can be withdrawn after market share targets are achieved. Permanent pricing is a cost structure statement: DeepSeek’s leadership has concluded that $0.27 per million input tokens is a sustainable margin-positive price for V4-Pro, or that they are willing to operate at current margins indefinitely to capture and hold developer market share. AI Weekly’s framing — that this escalates the inference pricing war — is the right characterisation: once a significant player makes a price cut permanent, it creates asymmetric competitive pressure on all other providers. If DeepSeek can sustainably serve near-frontier capability at $0.27, then every developer who chooses to pay $5 (Claude) or $3 (GPT) for tasks that V4-Pro can handle adequately is making an active choice to pay an 18.5x or 11x premium for marginal capability differences. Some will make that choice — quality-sensitive enterprise tasks, tasks where safety and alignment matter, tasks where Anthropic’s enterprise integration depth creates switching costs. Many will not.
The structural challenge for Anthropic and OpenAI is not that DeepSeek will capture their most demanding enterprise customers — it will not. Those customers value reliability, safety, governance, enterprise integration, and the specific capability advantages (Claude’s code flaw detection, GPT-5.5’s reasoning at the very frontier) that justify premium pricing. The structural challenge is the advisor model that Databricks CEO Ali Ghodsi described in May 2026: enterprises route routine and cost-sensitive tasks to cheap models and call frontier models only for tasks those cheaper models cannot handle. As DeepSeek’s capability improves and its price permanently stays at $0.27, the set of tasks that enterprises have economic justification to route to Claude or GPT-5.5 narrows. If the set narrows enough, the $43.6 billion Anthropic ARR and the $25 billion OpenAI ARR are not defended positions — they are high-water marks that require sustained frontier differentiation to maintain.
“DeepSeek makes 75% V4-Pro price cut permanent, escalating the inference pricing war. The cut directly challenges the pricing assumptions behind the enterprise AI revenue models of both OpenAI and Anthropic.” — AI Weekly, May 26, 2026
AI Inference Pricing Comparison — May 31, 2026
| Model | Provider | Price (Input per M tokens) | Price (Output per M tokens) | GPQA Diamond (est.) | Gap vs Claude Opus 4.8 |
| Claude Opus 4.8 | Anthropic (US) | $5.00 | $25.00 | Leads agentic benchmarks | Baseline (most expensive) |
| GPT-5.5 | OpenAI (US) | ~$3.00 | ~$15.00 | ~92% | 3.3x cheaper input |
| Gemini 3.5 Flash | Google (US) | ~$0.30-0.75 | ~$1.20-3.00 | Strong — leads science | 6-16x cheaper input |
| Kimi K2.6 | Moonshot AI (China) | $0.95 | ~$3.00 | 90.5% | 5.3x cheaper input |
| DeepSeek V4-Pro (permanent cut) | DeepSeek (China) | $0.27 | ~$1.10 | ~90% (est.) | 18.5x cheaper input |
| Zhipu GLM-5.1 | Zhipu AI (China) | $0.44 | ~$1.76 | ~87% (est.) | 11.4x cheaper input |
| Gemini 3.5 Flash-Lite | Google (US) | ~$0.075 | ~$0.30 | Competitive | 66x cheaper input |
The IPO Implication — Can Claude’s $43B ARR Survive Permanent $0.27 Competition?
The DeepSeek permanent price cut lands in the same week that Anthropic closed its $30 billion funding round at a $950 billion valuation on a thesis of reaching $50 billion or more in annual revenue within 18 months. The coexistence of those two facts — a $950 billion valuation thesis and a permanent 18.5x price disadvantage on a near-comparable model — is the tension at the heart of the frontier AI investment story in 2026. CNBC’s May 20 investigation, What Cheap AI Could Mean for OpenAI and Anthropic IPOs, identified this pricing gap as the primary structural threat to both companies’ public market valuations. The investigation found that the cheapest capable Chinese models cost approximately nine times less than Claude for equivalent tasks — and that this differential is already driving what Ghodsi calls the advisor model in enterprise deployments.
Anthropic’s defenders make two arguments. First, enterprise switching costs are real: a team that has integrated Claude Code into their production CI/CD pipeline, relied on Dynamic Workflows for codebase-scale migrations, and invested in enterprise compliance governance for Claude does not migrate to DeepSeek V4-Pro based on per-token pricing alone. Second, the specific capability advantages that Claude Opus 4.8 has over DeepSeek V4-Pro on agentic tasks — the 4x better code flaw detection, the Dynamic Workflows parallel subagent architecture, the Mythos-level prosocial alignment scores — matter disproportionately for the use cases that generate the highest-value enterprise revenue. Both arguments are true. Neither resolves the structural question of what happens over 18 to 36 months as Chinese model capability continues to improve while their pricing remains at $0.27. The gap between $0.27 and $5 is not closing from the top down — Anthropic has not cut its Opus pricing. It is being pressured from the bottom up as DeepSeek makes its price permanent and its capability continues improving.
“We used to debate whether AI would take jobs. Now we’re watching AI labs race to the bottom on price to win those jobs before they can charge what their R&D costs. The permanent DeepSeek cut is the clearest signal yet that inference pricing is commoditising faster than frontier labs planned.” — AI Weekly analysis of the DeepSeek permanent price cut, May 26, 2026
What Developers and Enterprises Should Do Now
The practical implications of the DeepSeek permanent cut are immediately actionable for developers and enterprise AI architects. For new workloads, the advisor model is now the rational default: evaluate whether the task requires frontier capability (Claude Opus 4.8, GPT-5.5) or whether near-frontier capability at 18.5x lower cost (DeepSeek V4-Pro) is adequate. For existing deployments, audit current token usage to identify cost-sensitive routine tasks — inbox summarisation, document classification, first-pass code review — that can be routed to DeepSeek or Gemini 3.5 Flash-Lite without meaningful quality loss. For enterprise compliance and governance buyers, note that DeepSeek is a Chinese-operated model with data processing that may not satisfy EU GDPR, US federal data handling, or industry-specific compliance requirements — the compliance argument for Western frontier models is the most durable competitive advantage that pricing does not erode.
For investors evaluating the Anthropic and OpenAI IPOs, the DeepSeek permanent price cut is the most important single data point to incorporate into valuation models. If the advisor model captures 40 to 60 percent of current premium AI token usage and routes it to models priced at $0.27 rather than $5, the revenue impact on Anthropic’s $43.6 billion ARR is not 40 to 60 percent — it is the revenue from those specific use cases, which may represent a much smaller share of total revenue than their share of total tokens suggests. Premium frontier enterprise customers who use Claude for mission-critical tasks with high switching costs are not the ones routing to DeepSeek. The marginal developer who is experimenting with AI in new workflows is exactly the customer most price-sensitive to an 18.5x gap. The IPO prospectuses due in Q3 2026 must address this dynamic directly. Any S-1 that does not include detailed analysis of the Chinese model competitive threat will face pointed questions from SEC reviewers and institutional investors who have been reading the same CNBC investigation that the CEOs’ job-prediction walk-backs were designed to soften.
Key Takeaways
• DeepSeek made its 75% V4-Pro price reduction permanent on May 26, 2026, setting the new sustainable price at $0.27 per million input tokens — 18.5 times cheaper than Claude Opus 4.8 ($5) and 11 times cheaper than GPT-5.5 ($3) for input tokens.
• DeepSeek V4-Pro achieves approximately 90% GPQA Diamond scientific reasoning — within a few benchmark points of Claude Opus 4.8 and GPT-5.5 — making the pricing gap a genuine enterprise decision variable rather than a capability-justified differential for routine tasks.
• The permanent cut escalates the AI inference pricing war: other Chinese models (Kimi K2.6 at $0.95, Zhipu GLM-5.1 at $0.44) are also in the sub-$1 per million token range for comparable capability, collectively constituting 60% of OpenRouter usage as of May 2026.
• The structural threat to Anthropic’s $43.6 billion ARR and OpenAI’s $25 billion ARR is the advisor model: enterprises route routine, cost-sensitive tasks to cheap models and call frontier models only for tasks those cheaper models cannot handle — narrowing the set of tasks that justify premium pricing.
• The DeepSeek cut arrives as both Anthropic and OpenAI are preparing Q3-Q4 2026 IPO filings with valuations of $950 billion and $852 billion respectively — both requiring sustained frontier revenue growth that depends on pricing power that the DeepSeek permanent cut directly challenges.
• Enterprise compliance and governance requirements — EU GDPR, US federal data handling, industry-specific regulations — are the most durable competitive moat for Western frontier models against Chinese alternatives that may not satisfy those requirements regardless of capability.
Conclusion
DeepSeek’s permanent 75 percent price cut is the most strategically significant AI pricing event of 2026 — more consequential for the frontier AI business model than any model launch or funding announcement. It signals that the Chinese AI labs have concluded that near-frontier capability at sub-$0.30 per million tokens is a sustainable, permanent market position, not a promotional tactic. For Anthropic and OpenAI, whose IPO valuations depend on sustained enterprise revenue growth at $5 and $3 per million tokens respectively, the permanent cut is a structural competitive challenge that their prospectuses must address honestly. The switching-cost and compliance arguments for premium frontier models are real and durable. They are also bounded: switching costs erode over time as integration patterns mature and alternatives develop, and compliance requirements evolve as Chinese AI providers invest in international data governance. The 18.5x price gap is not going away. The question that the IPO process will force both companies to answer in writing is: how much of the current revenue trajectory is protected by durable enterprise switching costs and compliance requirements, and how much is accessible to a competitor that is 18.5x cheaper and closing the capability gap every quarter?
Frequently Asked Questions
What is the DeepSeek price cut?
DeepSeek made its 75% V4-Pro model price reduction permanent on May 26, 2026, setting the new price at $0.27 per million input tokens and approximately $1.10 per million output tokens. This compares to Claude Opus 4.8’s $5 per million input tokens and GPT-5.5’s approximately $3 per million — an 18.5x and 11x price disadvantage respectively for the Western frontier models.
How does DeepSeek V4-Pro compare to Claude and GPT-5.5 in quality?
DeepSeek V4-Pro achieves approximately 90% on GPQA Diamond scientific reasoning benchmarks, within a few points of Claude Opus 4.8 and GPT-5.5. Claude leads on agentic coding, code flaw detection (4x better than Opus 4.7), and Dynamic Workflows parallel execution. GPT-5.5 leads on SWE-bench Verified coding benchmark. For routine tasks not requiring those specific capabilities, DeepSeek V4-Pro at $0.27 versus $5 is a genuine enterprise decision variable.
Does the DeepSeek price cut threaten the Anthropic and OpenAI IPOs?
It is a material competitive risk that must be disclosed in both companies’ S-1 filings. CNBC’s May 20 investigation identified the cheap Chinese model pricing gap as the primary structural threat to both companies’ IPO valuations. The advisor model pattern (using cheap models for routine tasks, frontier models for specialist tasks) is already reducing the share of total tokens going to premium providers. The IPO thesis depends on whether switching costs and compliance requirements protect enough revenue to sustain the $43.6B and $25B ARR trajectories.
Should enterprises switch from Claude to DeepSeek?
It depends on the use case. For tasks requiring Claude’s specific advantages (Dynamic Workflows, code flaw detection, Mythos-level alignment, enterprise governance), the premium is justified. For routine tasks where DeepSeek V4-Pro’s quality is adequate, the 18.5x cost difference is a genuine argument for the advisor model. Important caveat: DeepSeek is Chinese-operated and may not satisfy EU GDPR, US federal data handling, or industry-specific compliance requirements — which is a binding constraint for many enterprise deployments regardless of pricing.
Will Anthropic or OpenAI cut their prices in response?
Neither has announced price reductions as of May 31, 2026. Anthropic launched Opus 4.8 at unchanged pricing from Opus 4.7. Claude’s Fast Mode (2.5x faster, 3x cheaper than previous configurations) represents a cost efficiency improvement within the existing pricing structure. A direct price cut to Opus-class models would require either accepting lower margins or finding equivalent efficiency gains — which the GB200 Blackwell Ultra compute ramp at Colossus 2 may enable in H2 2026.
References
AI Weekly. (2026, May 26). DeepSeek makes 75% V4-Pro price cut permanent, escalating the inference pricing war. https://aiweekly.co/
CNBC. (2026, May 20). Cheap AI could derail OpenAI and Anthropic IPOs. https://www.cnbc.com/2026/05/20/cheap-ai-could-derail-openai-and-anthropic-ipos.html
Build Fast with AI. (2026, May 30). AI news today — May 30, 2026: 11 biggest stories. https://www.buildfastwithai.com/blogs/ai-news-today-may-30-2026
Artificial Analysis. (2026, May). AI model cost and performance benchmarking — May 2026. https://artificialanalysis.ai
OpenRouter. (2026, May). Model usage statistics — Chinese model share data. https://openrouter.ai
Build Fast with AI. (2026, May 27). AI news today — May 27, 2026. https://www.buildfastwithai.com/blogs/ai-news-today-may-27-2026
AI Weekly archive. (2026, May 31). 2360 stories tracked — DeepSeek pricing coverage. https://aiweekly.co/
