OpenAI Weighs Token Price Cuts as Anthropic Grows

OpenAI is discussing significant reductions to the per-token fees it charges developers and enterprise customers, according to people familiar with the talks. The deliberations come as OpenAI and Anthropic move toward confidential filings for initial public offerings and face rising pricing pressure from competitors and open-source providers. Sam Altman told an audience recently, “I think we’ll have a lot of ways we can help people get more value for less spend.”

Financial and usage data highlight the pressure on OpenAI. The company reported a negative 122% adjusted operating margin in the first quarter of 2026, meaning it lost $1.22 for every dollar of revenue. Market share on generative-AI web traffic for ChatGPT declined from 77.6% in May 2025 to 53.7% by April 2026. By contrast, Anthropic’s annualized revenue run rate rose from about $9 billion at the end of 2025 to roughly $47 billion by May 2026, driven largely by demand for its coding model. Anthropic recorded its first profitable quarter in Q2 2026.

Commercial users are shifting from capped consumer plans to metered API usage, which increases compute consumption and billing. Industry participants use the term “tokenmaxxing” to describe heavy token consumption by organizations, often without clear immediate returns. Companies report unexpectedly large AI bills: one large employer’s technology chief said the group’s 2026 AI budget was exhausted by April, and a payments division leader at a major bank warned that employee AI use could generate costs comparable to payroll. Bank analysts published a note titled “AI Bills Are Out of Control.”

Open-source inference providers are offering models developed in China, including DeepSeek V4, GLM, MiMo, Kimi and Minimax, at substantially lower prices than many closed-model APIs. Some providers are serving DeepSeek V4 at roughly one-thirteenth the price of comparable closed alternatives. The cost gap reflects licensing differences: several Chinese labs publish models under open licenses, so inference providers do not pay for model access. A venture investor summarized the dynamic: “The model is the single biggest cost an inference provider has, and they get it for free.”

Earlier pricing strategies for consumer tiers relied on a low flat monthly fee to attract users while expecting heavy enterprise usage to move to metered APIs. That structure exposes incumbents when enterprise adoption grows rapidly and cheaper inference options exist. Some users on social platforms allege that certain cloud AI vendors charge rates that exceed the underlying compute costs by wide margins.

OpenAI has raised the priority of its coding tools and expanded work on its Codex offering to better compete with Anthropic’s coding products. Both companies are pursuing enterprise customers as they prepare for public listings. Some industry observers say downward pressure on per-token pricing will continue while open-source models remain openly available; they add that if Chinese labs restrict access to their models, pricing trends could change.