Premium products rarely lose to better products. They lose to cheaper ones.
That has been the quiet rule of every technology cycle since the personal computer, and it is the rule American AI labs have been hoping does not apply to them. OpenAI’s GPT-5.5 sits at $5 per million input tokens and $30 per million output, according to public API rate cards compiled by BenchLM.ai. Anthropic’s Claude Opus 4.7 lands at $5 in and $25 out, the same source confirms.
Developers have largely accepted those rates because the cheapest serious open-source alternative still came with quality gaps that made the savings feel risky.
Then, on Saturday, May 23, a Hangzhou-based startup quietly changed the floor. Chinese AI company DeepSeek told developers it will make permanent a 75% price cut on its flagship V4-Pro model, locking in API prices at roughly a quarter of their original level, as highlighted in an official X post.
DeepSeek permanently cuts price of V4-Pro
The 75% cut was originally framed as a promotional sale set to expire on May 31, as in an April 29 X post. DeepSeek instead made this price the permanent floor.
The new pricing puts V4-Pro between $0.0035 and $0.83 per million tokens, depending on usage type, down from a range of $0.0145 to $3.48, Reuters noted. That is the rate any developer building a real product against DeepSeek will see on their invoice from this point forward, not just during a promotional window.
More AI:
- Micron sits at the center of a red-hot chip rally
- IBM CEO sends blunt message on AI and quantum computing
- Anthropic CEO makes shocking admission about AI
For context on how aggressive American labs have been holding the line, Morgan Stanley recently revamped Nvidia‘s price target ahead of big Q3 on the assumption that frontier compute demand stays priced at a premium.
I have been tracking AI API rate cards since R1 first jolted the market in early 2025, and May 23’s move is structurally different from earlier promotional pricing. Promotional discounts come with an expiration date that competitors can simply outwait. Permanent discounts do not.
Why DeepSeek’s price cut cut lands hardest on OpenAI, Anthropic
The pain is not the gap on a single API call. The pain is what the gap does to every build-versus-buy decision American AI labs have been winning by default.
A team budgeting for 100 million output tokens per month sees Claude Opus 4.7 at roughly $2,500 and GPT-5.5 at roughly $3,000, based on pricing analysis published by BuildFastWithAI.
Related: Anthropic just scored major win in escalating Al talent war
The same workload on DeepSeek V4-Pro at the new permanent rate comes in near $348, according to the same analysis. The output cost gap alone is nearly seven-to-one against Anthropic and nearly nine-to-one against OpenAI.
That changes the conversation in every engineering meeting where someone has been quietly running both APIs on the same prompts. The math now favors the Chinese model on cost so heavily that even real quality gaps on the hardest benchmarks become defensible trade-offs for most production workloads.
The Huawei chip story hiding behind DeepSeek’s move
The price cut is not an act of charity, and DeepSeek did not explicitly explain what enabled it.
DeepSeek “did not disclose whether the permanent price cut was due to increased supply of Huawei’s Ascend 950 chips, which it used to maximize V4’s performance,” Reuters reported. The company has previously said Pro pricing would fall sharply once Huawei Ascend 950 supernodes ship in large quantities in the second half of 2026, Reuters indicated.
When I ran the numbers against published API rates from the major frontier labs, three data points stood out.
- DeepSeek V4-Pro at the new permanent rate is roughly one-sixth the cost of Claude Opus 4.7 per blended million tokens, according to VentureBeat.
- GPT-5.5 doubled in price versus the previous GPT-5.4 generation on output tokens, BenchLM.ai noted.
- DeepSeek is reportedly chasing a $45 billion funding round while it cuts prices, AndroidHeadlines reported.
That is not the cost structure of a company that needs API revenue to survive. It is the cost structure of a company building a long-term moat at the expense of immediate margins.
What this means for your AI stock exposure
The American AI labs are private companies, so the price war hits the public market sideways. It hits the names that supply the labs.
Nvidia (NVDA), already the most visible AI proxy on the Nasdaq, has spent a year benefiting from the assumption that the largest AI workloads would route through its most expensive chips. A serious open-source alternative running on Huawei Ascend silicon complicates that thesis.
The chip story matters more, not less, when DeepSeek can stay cheap because it is not paying Nvidia margins.
Hyperscalers such as Microsoft (MSFT), Alphabet (GOOGL), and Amazon (AMZN) face the same question from a different angle. They have been reselling premium AI access at premium margins. A permanent low-cost frontier alternative drags every conversation about cloud AI billing toward DeepSeek’s number.
The AI software margin signal investors should not miss
Permanent pricing changes everything about how the next 12 months play out for developer adoption.
A discount that expires is a marketing event. A discount that does not expire is a market floor. Every customer-acquisition deck written by an OpenAI or Anthropic sales team now has to assume that prospective enterprise customers know they can route a meaningful share of their workload to V4-Pro and absorb the trade-offs in exchange for a 70%-plus cost reduction.
For investors, the practical read is straightforward. The AI capex story still favors the names that supply the picks and shovels. The AI software margin story just got a structural ceiling, and the recent news from Hangzhou is the most important data point on that ceiling all month.
Related: Tesla’s Chinese rivals open bold new front beyond price
