American AI isn’t losing the price war

Chinese AI labs are testing higher prices just as American rivals offer more intelligence for less, according to a newsletter analysis. The piece highlights that robotics startup Generalist raised an additional $200 million, Lambda is raising a $3 billion pre-IPO round, and Nvidia is considering investing in Perplexity at a $30 billion valuation.

American AI isn’t losing the price war Chinese labs are testing higher prices just as their American rivals offer more intelligence for less. Tuesday. Today we’re looking at historical trends in AI pricing. How did China become the low-price leader? What really made American AI reviled for its costs earlier this year? And what’s the current state of play? Today’s headline is a hint, but read the whole thing. It was really fun to write. To work — Alex https://x.com/alex 📈 Trending Up: Idiocracy https://www.notus.org/florida/school-vaccine-requirement-rollback …… Apple silicon idiocracy https://abcnews.com/Health/2-people-die-measles-pennsylvania/story?id=135932723&cid=social twitter abcn …… Xiaomi silicon embarrassing laziness https://x.com/scottastevenson/status/2092239039482974226 … agentic payments https://x.com/wongmjane/status/2091967611345289447 … second-order effects https://www.nytimes.com/2026/08/25/business/dealbook/emerald-ai-start-up-data-center-backlash.html … our water future https://t.co/QZikxP1KyH … competition in legal AI https://www.reuters.com/business/google-expands-gemini-ai-platform-law-firms-lawyers-2026-08-25/ … 📉 https://finance.yahoo.com/news/servicenow-pledges-1-5bn-investment-110000403.html Trending Down: Housing affordability https://x.com/DeItaone/status/2092235766193623224 …… Palantir in Europe PSA: Are you a venture firm looking to raise your next fund? A startup announcing a raise? A late-stage unicorn looking to go public? You may need a video, or a whole annual meeting. If so, hit up my former producer, Chris Gates, who has hung his own shingle and is producing films and events for investment firms and founders. He’s a peach. And brilliant to boot Big Money Moves Before we bury our heads in the AI sands, let’s talk dollars: - Robotics startup Generalist raised $200 million more https://www.axios.com/2026/08/24/robotics-ai-generalist-200m , mere months after raising $400 million. Generalist is building general — get it? — intelligence “for the physical world,” in its own words https://generalistai.com/about . Robotics investment is rising, thanks to bullishness on ‘ Physical AI https://news.crunchbase.com/venture/physical-ai-funding-startups-robotics-aerospace-h1-2026/ ‘ and what the technology may unlock. Lambda is raising a $3 billion pre-IPO round https://finance.yahoo.com/technology/ai/articles/ai-cloud-provider-lambda-in-talks-for-3-billion-pre-ipo-round-234205946.html ahead of a possible 2027 IPO, Bloomberg reports. Public neoclouds CoreWeave and Nebis are worth $48.6 billion and $58.8 billion, respectively, so it’s understandable why investors are interested in putting more capital into Lambda. Oura could target a $16 billion valuation when it lists, alongside a $3 billion raise. That’s up from its trailing $11 billion valuation. Oura generated revenue around $500 million and $1,000 million in 2024 and 2025 https://techcrunch.com/2026/08/24/oura-is-reportedly-eyeing-a-september-ipo-that-could-value-it-at-more-than-16b/ , with expectations for as much as $2 billion this year, for reference. Oura pairs hardware and subscriptions, which could give it attractive economics both in CAC and gross margin terms. Nvidia is considering investing in Perplexity at a $30 billion valuation. The startup was once an inescapable member of the AI bullrush. To see the chip giant invest — more on Nvidia’s check-writing here https://www.cautiousoptimism.news/making-sense-of-nvidias-rapid-fire-dealmaking/ — would be a boon for the unicorn. The Information reports that Perplexity’s revenues have soared to $750 million on an annualized basis https://www.theinformation.com/articles/nvidia-discusses-perplexity-investment-30-billion-plus-valuation-considered-tech-licensing-deal .- And, finally, Unitree’s IPO went vertical, rising some 542% in its market debut https://www.cnbc.com/2026/08/19/china-backflipping-robot-maker-unitree-jumps-shanghai-ipo.html . However, after reaching a peak of 1,100 RMB per share, the robotics company’s shares now trade at 603.80 RMB, a dramatic comedown from its early trading sessions. Companies go public earlier in China than here in the States, which I appreciate greatly. The AI price war changed sides As OpenAI and Anthropic prepare to go public, opinion about their prospects is in flux. While OpenAI’s third-quarter growth reacceleration https://www.cautiousoptimism.news/openai-is-not-dead-yet-wants-to-go-for-a-walk/ and Anthropic’s chart-topping first half of 2026 are undoubtedly impressive, many expect low-cost, open-weight AI models from China to topple the leading American AI labs. Why? OpenAI and Anthropic charge too much for their models, the argument goes, and as Chinese models manage to stay close to the frontier, their price advantage will earn them the future. I doubt it. Why? Because the era of cheap Chinese AI models is fading just as American labs raise the capability bar and offer more intelligence for less. Meanwhile, Chinese AI labs are raising prices and retaining more of the economics their models create for third-party inference providers. China’s AI price war How did China end up with the reputation as the low-cost AI leader? The nation’s AI labs bought the moniker honestly through years of aggressive price cuts. Let’s rewind the clock to May 2024, when GPT-4o was released https://openai.com/index/hello-gpt-4o/ . That month brought cuts from Chinese AI labs worth as much as 99.3% https://www.reuters.com/technology/chinese-tech-giants-slash-prices-language-models-used-power-ai-chatbots-2024-05-21/ ByteDance and 97% https://www.bloomberg.com/news/articles/2024-05-21/alibaba-sparks-china-ai-price-war-with-spate-of-steep-discounts https://www.reuters.com/technology/chinese-tech-giants-slash-prices-language-models-used-power-ai-chatbots-2024-05-21/ Alibaba . In response, Baidu began offering certain AI services to businesses for free. Low-cost AI models from Chinese AI labs kept coming. In the final days of 2024, DeepSeek announced its V3 model https://api-docs.deepseek.com/news/news1226/ , priced at $0.14 per million input tokens and $0.28 per million output tokens thanks to promotional pricing a 50% discount . In late 2024, Alibaba cut the price of its visual Qwen-VL by 85% https://www.cnbc.com/2024/12/31/alibaba-baba-cloud-unit-slashes-prices-on-ai-models-by-up-to-85percent.html . A few months later, in February 2025, DeepSeek announced off-peak discounts https://x.com/deepseek ai/status/1894710448676884671?lang=en for V3 50% and its then-famous R1 reasoning model 75% . In early 2025, Baidu’s Ernie 4.5 Turbo launched with an 80% price reduction https://jingdaily.com/posts/china-baidu-upgrades-ernie-ai-models-slashes-prices compared to its predecessor. Similarly, Baidu’s X1 Turbo model came in at half the price of its own ancestor. In November 2025, Alibaba cut the price of its Qwen3-Max model by 50% https://finance.yahoo.com/news/alibaba-slashes-qwen3-max-model-093000463.html . The price war continued into the new year. We could detail every price cut made by Chinese AI labs heading into the last few months, but that would be silly. We get the point: Chinese AI labs purchased market share from 2024 to 2026 with quickly improving AI models priced below their American competitors. Meanwhile, in America While Chinese labs were engaged in a price war, American AI labs kept pushing the frontier. This year, for example, brought a series of impressive Opus models from Anthropic, and the well-received GPT-5.6 model family from OpenAI https://www.cautiousoptimism.news/the-ai-agents-are-coming-for-microsoft-office/ . We also saw Mythos and Fable arrive from A\, setting new high watermarks for model performance. However, while American AI labs did a praiseworthy job of keeping their home nation atop the AI intelligence charts in recent years, they also managed material price reductions over the same time period. Today’s flagship American models are far cheaper than their predecessors. In 2023, OpenAI’s best model, GPT-4, cost a staggering https://openai.com/index/gpt-4-research/ $30 and $60 per million input/output tokens. The company’s latest leading model GPT-5.6-Sol arrived in July at $5 and $30 per million input/output tokens. Similarly, Anthropic’s early 2024 Claude 3 Opus was priced at $15 and $75 per million input/output tokens https://www.anthropic.com/news/claude-3-family . The most recent Opus launch, its 5.0 release, came priced at $5 and $25 per million input and output tokens https://www.anthropic.com/news/claude-opus-5 . If AI prices have come down from American AI labs as their performance has improved, why have we spent this year fretting about AI costs? The Cost Crisis was driven more by the erosion of all-you-can-eat AI subscriptions as providers adopted more usage-based pricing, and an explosion in token demand driven by increasing agentic AI usage than raw token costs. List token pricing is only one of several vectors that impact overall AI cost. Token usage is another. A model with a low per-token cost and huge token burn may cost more to use than a model that has a higher list price but more efficient token consumption. No matter the cause, however, by mid-2026 it became clear that companies adopting AI could not stomach the cost of letting their staff go hog-wild with the technology. That was a problem: business leaders wanted AI to help reduce their overall costs, while AI labs wanted usage to grow. What happened next? In response to market concerns, American AI labs got serious about offering better value per token. Thus, statements from Meta Muse Spark 1.1 and SpaceXAI Grok 4.5 https://www.cautiousoptimism.news/the-ai-agents-are-coming-for-microsoft-office/ about their latest AI models and their prices: Low, and cheaper than what OpenAI and Anthropic were serving up. But more efficient models with attractive token list prices weren’t the only goodies American AI labs had cooking. They’ve leaned into discounts and price cuts, too. Temporary promotions mean you can use Grok 4.6 with the Hermes agent for 50% off today https://x.com/SpaceXAI/status/2091957125941543034 . Google’s Gemini 3.7 Flash model is currently 75% off on OpenRouter https://openrouter.ai/models?discount=true , while OpenAI’s GPT-5.6 Sol model is currently 50% off at the same vendor. More importantly, OpenAI has led the low-cost charge by aggressively cutting the prices of its current models. A few weeks after announcing the GPT-5.6 model family, OpenAI permanently cut the price of Luna the smallest variant by 80% https://www.cautiousoptimism.news/openai-reclaims-its-place-at-the-frontier/ , and Terra the midsize version by 20%. A few days ago, OpenAI also cut the price of Sol https://x.com/MadhavMalani11/status/2091950416082968762 the largest GPT-5.6 model by 20% for the next three months though I suspect the discount may be extended . Even Anthropic is getting in on the fun: In mid-August, the company announced https://www.anthropic.com/news/claude-sonnet-5 that its Claude Sonnet 5 the midsize Claude varietal would retain its introductory pricing for good, keeping it at $2 and $10 per million input and output tokens. As a result of OpenAI’s Luna price cuts in particular, an American AI lab took first place on the critical intelligence/price charts https://artificialanalysis.ai/ intelligence at least per Artificial Analysis . Talk about a turnaround. Chinese AI labs charge more Recently, as Chinese AI labs have matured, gone public, and raised investor expectations by attracting ever-higher valuations, commercial pressures have made it harder to offer so much for so little. In response, Chinese AI leaders are testing price increases across their product portfolios, from raw token rates to subscription tiers. We got a preview of the trend in February 2026, when Z.ai of GLM fame raised prices for several of its subscription plans https://x.com/Zai org/status/2021656635668901985 . March brought price hikes from Alibaba and Tencent https://www.caixinglobal.com/2026-03-19/alibaba-tencent-hike-cloud-prices-as-ai-boom-drives-up-hardware-costs-102424744.html . In July, Z.ai made further changes to its pricing plans https://docs.z.ai/devpack/notice/usage-revision , discontinuing previous, more generous versions of its service. MiniMax tried to price its M3 model at twice what the preceding M2 cost https://platform.minimax.io/subscribe/token-plan?tab=api-enterprise , but wound up extending promotional pricing in perpetuity. Most recently, we’ve seen the impressive Moonshot Kimi K3 model launch with list prices of $3 and $15 per million input and output tokens https://www.cautiousoptimism.news/china-finally-built-an-american-ai-model/ . That pricing lands K3 between Anthropic’s Sonnet and Opus 5 models, and between OpenAI’s Terra and Sol models. As we wrote at the time, Kimi K3 was China’s first American AI model in terms of its cost. There’s more. In recent months, we’ve seen Chinese AI labs change https://www.cautiousoptimism.news/moonshot-opens-the-open-model-tollbooth/ their open-weight model licenses, adding a commercial agreement requirement for third-party inference providers. The new rules of the road will either reduce neocloud margins or force price increases for Chinese AI models abroad. All told, the era of rapid-fire price cuts from China has been replaced with a period of moderately rising prices. Therefore, we’re hardly barreling toward a world in which China sells cheap, open intelligence and America slings expensive, closed intelligence.