OpenAI Starts a Token Price War: Is the Token Economy at a Turning Point?

1. Key Numbers Indicating a Signal

Anthropic reports a developer spending about $13 per day on tokens with Claude Code, which scales to $75,600 annually for a ten‑person team—roughly the salary of a mid‑level engineer.

Salesforce expects to spend $300 million on Anthropic this year, while Uber burned its entire annual token budget in the first four months, coining the term tokenmaxxing for wasteful usage.

Entelligence.AI tracked 2,444 companies and found that for every $1 spent on tokens, only $0.18 translates into actual user value; $0.44 covers AI‑generated bugs, $0.27 goes to rework, and $0.11 to review friction.

Goldman Sachs reports that AI token spend already accounts for up to 10% of some enterprises' total labor costs and is still rising.

The industry is asking the same question: Is token spending worth it?

2. Three‑Year, Three‑Step Evolution

The commercialization of AI has progressed through three stages, each reinforcing the principle that what you sell determines your pricing, and pricing determines product longevity .

First jump (2023): Monthly subscription model – e.g., ChatGPT Plus at $19.99/month, adopted by Chinese tech giants. Subscriptions provide predictable revenue but hide true usage costs, treating heavy and light users alike.

Second jump: Subsidy wars – Google offered 15 months of free Gemini Advanced, ByteDance advertised “99.3% cheaper than industry” for Doubao, and Baidu made core models free. These subsidies mask the real cost, as Microsoft’s GitHub Copilot loses over $20 per user per month, $80 for heavy users.

Third jump (June 1 2026): Microsoft announced that Copilot will switch to pure token billing, converting the $19/month plan into a token allowance that expires when used. A single agent‑coding session can consume $30‑$40, exhausting a month’s allowance in one go, exposing the true cost.

Each jump brings AI pricing closer to its actual cost, making profitability harder.

3. The Upcoming Fourth Jump

OpenAI’s price cut is less about competitor strength and more about its own situation: both OpenAI and Anthropic are filing for IPOs, making a simultaneous price war financially untenable.

Raising prices would improve profit margins for investors, but cutting prices is seen as necessary because enterprise customers are being forced into “tokenmaxxing” and may abandon usage otherwise.

Amazon has halted its internal AI usage leaderboard, urging employees “not to use AI for the sake of using AI,” and Microsoft plans to drop some Claude Code subscriptions.

Sam Altman’s admission that “cost is a huge problem” translates to a warning: if costs aren’t curbed, customers will stop using the services.

The price war is not a choice but a survival tactic: sell tokens cheaper, hoping higher volume compensates, which only works if demand is elastic. Entelligence’s data (only $0.18 of value per $1 spent) suggests demand elasticity may be limited.

4. Three Unavoidable Questions

1) Model selection criteria now include price volatility. After OpenAI’s cut, Anthropic is likely to follow, and other platforms may join. A model’s pricing could halve within six months, forcing product strategies that depend on static pricing to be re‑evaluated.

2) What your product actually sells – if the core value is API calls, the price war directly threatens margins. Mid‑layer products that merely wrap APIs must derive value from data, workflow integration, user habits, or industry knowledge, not from the model itself.

3) Who controls pricing power – a $50/month plan versus a $0.10 per request model hinges on upstream token costs. If OpenAI reduces token prices by 90%, existing pricing strategies may become untenable.

Many AI products have let API pricing dictate their own pricing without establishing independent value anchors. The price war acts as a forced exam, testing what remains valuable once model margins disappear.

Conclusion: Tokens Can Get Cheaper, but Good Products Won’t

The token price war marks a shift from selling scarcity (model capability) to selling value beyond the token. As tokens become commoditized, the differentiators are data, integration, and domain expertise.

Goldman Sachs notes AI spend is rising as a share of labor costs; Amazon warns against using AI for its own sake; Uber’s “tokenmaxxing” term highlights the emerging ROI focus. Users are moving past the novelty phase and demanding clear returns – if products can’t demonstrate that, customers will seek alternatives.