AI Price War Heats Up as OpenAI and Anthropic Face Mounting Profitability Pressure
0xBroomberg
Enterprise clients are using model-routing tools to dynamically switch workloads to cheap open-source models, cutting some AI costs by up to 95%. OpenAI and Anthropic — already losing billions annually — now face mounting pressure to slash prices, just as both have confidentially filed for IPOs.
How are companies cutting AI costs by 95%?
A new category called model routing lets enterprises dispatch AI tasks like a taxi dispatcher: hard problems go to premium models (ChatGPT, Claude); routine work goes to cheap ones (DeepSeek, Alibaba's GLM, and other open-source options).
Dan Robinson, founder of bug-detection startup Detail, says he has shifted 90% of workloads away from Claude and Google Gemini to custom and Chinese-developed GLM models.
This means → OpenAI and Anthropic's flagship models no longer handle the full workload. Their revenue base is being hollowed out from the bottom by cheap alternatives.
How fast is the shift to cheap models?
On developer platform Vercel, DeepSeek's share of AI usage jumped from 1% in April to 17% in May.
On AI query platform OpenRouter, DeepSeek became the most-used AI provider from mid-May onward. Among high-spend clients, open-source model token usage is growing four times faster than proprietary models. Over 500 organizations have switched from proprietary to open-source.
In plain terms = this is not early adopters experimenting — enterprise clients are migrating at scale.
Do premium models still have an edge?
Anthropic's newly released Fable 5 is priced at over 50 times DeepSeek V4 Pro per token — a stark gap.
Anthropic argues the right metric is not "price per token" but "cost per task completed" — top models sometimes finish complex tasks with fewer tokens, so total cost may not be higher.
Researchers estimate that frontier proprietary models from OpenAI, Anthropic, and Google still lead open-source rivals by roughly four to six months. But Vishal Misra, associate dean for computing and AI at Columbia Engineering, pushes back: "You don't need a model that understands quantum gravity. Open-source is already very capable. AI's premium pricing power will gradually fade."
Frontier model lead — durable moat or ticking clock?
BULL
Complex tasks still need top models
Measured by cost per task, premium models may not be pricier.
~6-month lead persists
Researchers see proprietary models four to six months ahead of open-source.
BEAR
Most tasks don't need the best model
90% of workloads can already be handled by cheap alternatives.
Open-source is catching up fast
Open-source token growth is 4× faster than proprietary — the gap is narrowing quickly.
In plain terms = frontier models are genuinely stronger on the hardest tasks, but most day-to-day enterprise work never reaches that difficulty level — and that is where the premium revenue base is eroding.
How is OpenAI responding?
The Wall Street Journal reports that OpenAI is considering significant price cuts for users and expects Anthropic to follow suit.
OpenAI believes it has a buffer: over the past year it locked in large computing resources at prices well below current market rates. This means → even after cutting prices, its cost pressure is temporarily lower than competitors'.
CEO Sam Altman said at a recent company event that cost has suddenly become "a huge problem."
Why is this price war most dangerous right now?
OpenAI and Anthropic have both confidentially filed IPO paperwork with the SEC — widening losses will directly affect how capital markets value them.
U.S. tech giants are also racing into the cheap-model lane: Microsoft last week released a more efficient suite of small AI models; Nvidia's Nemotron budget model series is rapidly gaining share.
This reflects a squeeze from both directions — when Microsoft and Nvidia themselves build budget models, OpenAI and Anthropic face pricing pressure from suppliers and customers alike. Hedge fund Citadel Securities summed it up this week: "Even the most powerful technology must pass the test of cost curves, capacity constraints, and diminishing returns."
Content is for reference only, not financial advice.