Uber spent its whole 2026 AI budget in four months

AI bills are no longer anchored to a neat token tariff. From March 2023 to April 2025, the same-intelligence inference price dropped from $30 per million tokens to $0.1 to $0.15 per million tokens, yet enterprise AI cloud spending still tripled.

That gap is the story. Woshipm says the cost shape is shifting from a predictable fixed cost into an explosive variable one, because call volume is rising faster than prices are falling. Capacity limits make the point concrete: Anthropic has repeatedly issued capacity-limited notices, OpenAI has repeatedly rate-limited its API, and domestic inference compute approvals are tight. The rest of this piece explains why the bill now depends on usage shape, not just token price.

Token prices are falling, but spend is rising elsewhere

From March 2023 to April 2025, the inference cost of models at the same intelligence level fell from $30 per million tokens to $0.1 to $0.15 per million tokens. Woshipm describes that drop as more than 99.7%.

That collapse does not mean AI bills got simple. Over the same period, enterprise AI cloud spending increased by three times, while the source argues token prices are falling and call volume is rising faster than prices are falling. The bill is shifting from a predictable fixed cost into an explosive variable cost.

The pressure shows up outside the headline rate card. Woshipm says Anthropic has repeatedly issued capacity-limited notices, OpenAI has repeatedly rate-limited its API, and domestic inference compute approvals are tight. Those limits matter because a cheap token is still useless if the service cannot answer at the moment you need it. In that setting, the cheapest part of AI pricing moves toward zero gross margin, leaving model vendors only two paths: scale to the lowest cost or offer unique capabilities that justify a premium. By March 2026, the pricing picture had already turned more uneven: Zhipu, Alibaba Cloud, Tencent Cloud, and Baidu collectively raised token prices by 5% to 460%.

Peak hours, cache hits, and free quota change the math

DeepSeek's pricing shows why the posted token rate is only the starting point. According to appinn, deepseek-v4-pro charges 0.025 yuan per 1 million input tokens with a cache hit, then 0.05 yuan during peak hours. The same pattern applies to cache misses and output tokens, where the normal 3 yuan and 6 yuan rates become 6 yuan and 12 yuan. Deepseek-v4-flash follows the same structure: 0.02 yuan becomes 0.04 yuan for cache hits, 1 yuan becomes 2 yuan for cache misses, and 2 yuan becomes 4 yuan for output.

That gap matters because cache hits and cache misses are not interchangeable. A request that reuses cached input can be priced almost at the floor, while a miss jumps to the full input rate before peak-hour doubling even starts. On August 4, geekpark said DeepSeek V4 Flash API had capacity problems because of "unprecedented traffic," and geekpark also reported that "DeepSeek Slash Line" was trending on Weibo that day. The billing shape and the traffic shape are linked.

Free tiers can hide this until usage rises. Appinn reported that some users can use Gemini 2.5 Flash for free at up to 1,000,000 tokens per minute, but the free quota is not the same for all users. Free-tier Gemini users are also limited to 15 requests per minute and 1,500 requests per day, while Gemini 2.5 Flash-Lite has a 30 requests per minute cap and the same 1,500 requests per day cap. Qing Xiaowa's account, appinn said, had only 250K tokens per minute for Gemini 2.5 Flash and Gemini 2.5 Flash-Lite.

The fine print also changes the bill's meaning. Appinn said the free version of Google AI Studio lets Google use submitted prompts and received responses to improve its models, while users in the European Union are treated differently. Pro models sit outside the free tier entirely. A free quota is useful, but it is not a pricing guarantee.

Agent loops turn cheap calls into runaway spend

Agentic software changes the unit of pain. A single user input can now fan out into dozens or even hundreds of model calls when the product is doing agent collaboration, long-context memory, multi-step reasoning, code generation, or multimodal processing. That means the bill is shaped by loops and retries, not by one neat prompt-response exchange.

The failure mode is easy to see in the ledger. Uber opened AI coding tools to 5,000 engineers and burned through its entire 2026 budget in four months. At Mihoyo, Zheng Yinhe said a multi-agent test without a circuit breaker sent dozens of agents into a loop and created a 200 million yuan token bill in 13 hours. The short version: if you do not cap the loop, the loop caps you.

That is why launch-time budgeting has to come from logs, not from a token sticker price. Monday.com's CEO publicly announced in January that the company replaced an entire 100-person SDR team with AI Agents, and Atlassian reported its first-ever decline in enterprise seat count in the same year because AI Agents took over tasks that used to need dedicated staff. Once the workflow is agentic, token cost becomes only the visible edge of a much larger operating pattern.

The same pattern shows up at the individual level. A researcher in the AI field reportedly consumed nearly 4 billion tokens in a single day while coding. At GLM-5.2's price, that would cost about $1,084 in one day. That is still only one day. If a product can turn one task into repeated calls, the right controls are circuit breakers, caps, and budget estimates drawn from real logs before launch, not faith in the posted rate.

Subscriptions, quotas, and outcome pricing can beat metered APIs

Zendesk and HubSpot show why the bill can move away from a raw token meter. Zendesk's AI Agent charges about $1.5 per automatic resolution under committed volume and about $2 per automatic resolution on demand, so the unit is an outcome, not a token. HubSpot changed Breeze Customer Agent in April from $1 per handled conversation to $0.5 per successfully resolved conversation. That shift rewards finish rates and makes a cheap-looking per-request system expensive if it misses too often.

GitHub Copilot moved fully to usage-based billing starting on June 1, but usage-based does not mean simple metering. Woshipm says Futurum's survey in the first half of the year found that 43% of enterprise software decision-makers preferred consumption-based pricing and 27% preferred outcome-based pricing. The same source says the median growth of hybrid pricing, which combines a subscription base price and a usage component, is 21% higher than pure subscription or pure usage pricing, based on Stripe data. That helps explain why vendors keep mixing fixed access with variable charges.

The practical advantage for heavy users often comes from bundles and quotas, even when the list price looks higher. The researcher used a Codex subscription priced at $20 per month for Plus and starting at $100 per month for Pro. Another heavy developer said he spent 300 yuan in one week on Codex without company reimbursement. Woshipm also says Kimi Code has four tiers and its quota refreshes weekly, while GLM's coding subscription limits usage by the number of calls every 5 hours and deducts quota at three times the normal rate during peak periods. A friend working in operations said Alibaba Qoder's enterprise version ran out in three days under his usage.

That is the core tradeoff. Metered APIs look clean on a price sheet, but subscriptions, quotas, and per-resolution pricing can cap volatility or reward completion. For teams that generate lots of calls, a bundle can be cheaper than a supposedly low per-token rate, especially when the work is bursty or the vendor counts success instead of raw volume.

When token price matters less than the bill you actually incur

• You expect heavy, bursty, or agentic usage where one prompt can trigger many downstream calls. Do not budget only off posted token rates. The sources say a single user input can trigger dozens or even hundreds of model calls in agent collaboration, long-context memory, multi-step reasoning, code generation, and multimodal processing, and they give examples of explosive variable cost, including a 200 million yuan token bill in 13 hours when a multi-agent loop lacked a circuit breaker. In this situation, usage shape and loop control matter more than headline token price.
• Your workload depends on uninterrupted throughput during busy periods. Prefer vendors with clear capacity guarantees or generous quotas, and avoid assuming a cheap API will stay cheap under load. The ledger says Anthropic issued capacity-limited notices, OpenAI rate-limited its API, and DeepSeek planned to double prices during 9:00-12:00 and 14:00-18:00 Beijing time. For peak-sensitive workflows, surcharge exposure can outweigh a low base rate.
• Your usage is highly repetitive and benefits from cache hits or fixed workflow packages. Look for pricing models that reward successful resolution, cache hits, or bundled quotas rather than raw tokens. The sources show deepseek-v4-flash charging 0.02 yuan for 1 million input tokens with cache hit versus 1 yuan with cache miss, and HubSpot changing Breeze Customer Agent from $1 per handled conversation to $0.5 per successfully resolved conversation. These structures can be much cheaper than paying for every token.
• You are choosing between a flat subscription and a metered plan for developer tooling or agent workflows. Use consumption-based or hybrid pricing when usage is variable, but validate the quota rules carefully. The sources say 43% of enterprise software decision-makers preferred consumption-based pricing and 27% preferred outcome-based pricing, while hybrid pricing grew 21% faster than pure subscription or pure usage pricing. At the same time, Kimi Code has four tiers with weekly quota refreshes, and GLM's coding subscription limits calls every 5 hours and deducts quota at three times the normal rate during peak periods.
• You are a vendor or buyer deciding whether cheap tokens alone are enough to win or de-risk the stack. Treat token price as commodity-level input, not the core differentiator. The ledger shows model prices collapsing from $30 per million tokens in March 2023 to $0.1 to $0.15 per million tokens by April 2025, while enterprise AI cloud spending tripled. That points to competition moving toward capacity, quotas, workflow controls, and pricing design rather than simple per-token discounts.

Provider strategy is diverging, so procurement has to look past list price

DeepSeek's pricing story is no longer only about a low posted rate. On August 6, according to geekpark, it said in its open platform backend that it planned to raise DeepSeek API service prices overall in the near future and expected a large increase. That warning lands after OpenRouter's latest weekly statistics showed DeepSeek V4 Flash at the top globally in model call volume on August 4, and after OpenCode recorded 8 trillion tokens in a single day on August 1, with 5 trillion from free quota and 3 trillion from paid calls.

Gemini points in a different direction. According to appinn, some users can use Gemini 2.5 Flash for free at up to 1,000,000 tokens per minute, but the free quota is not the same for all users. Free-tier users are limited to 15 requests per minute and 1,500 requests per day, while Gemini 2.5 Flash-Lite has a 30-request ceiling and the same 1,500-request daily cap. The free version of Google AI Studio also reserves the right to use submitted prompts and received responses to improve its models, though free-tier data is not used for training for users in the European Union.

That split is why procurement has to look past input and output rates. Woshipm lists DeepSeek-V4-Pro at $0.435 per million input tokens and $0.87 per million output tokens, while Gemini 2.5 Flash is listed at $0.15 and $0.60. Those numbers matter, but so do quota ceilings, data-use rules, and whether a vendor can change the bill after adoption. Geekpark's author argues that the token is moving toward commodity status, and that relying on DeepSeek's floor-price API is risky for startups and independent developers. In practice, the purchasing question is who controls access, who can train on your prompts, and how exposed you are when a provider shifts prices or quota policy.

How AI pricing shifts from token rates to usage shape

When you compare Gemini models, start with the workload ceiling, not the sticker price. According to appinn, the account Qing Xiaowa checked had only 250K tokens per minute for Gemini 2.5 Flash and Gemini 2.5 Flash-Lite, while all Pro models sit outside the free tier. That changes the bill before a single token is counted.

Then price the path your app actually takes. Flash-Lite is listed at $0.075 input and $0.30 output per 1M tokens; Flash is $0.15 input and $0.60 output; Pro is $1.25 input and $10.00 output. In the free version of Google AI Studio, Google can use prompts and responses to improve its models, though free-tier data from the European Union is excluded. If your use case can run on another model, appinn reports Qing Xiaowa argues DeepSeek can replace Gemini here.

For readers outside China

• Availability: The sources discuss products and quotas from Chinese and U.S. vendors, but they do not provide a single unified availability map for outside China. DeepSeek, Zhipu, Alibaba Cloud, Tencent Cloud, Baidu, Google AI Studio, OpenRouter, Anthropic, OpenAI, HubSpot, Zendesk, GitHub Copilot, and Gemini are all mentioned, but country-by-country availability is not disclosed in sources.
• Pricing: Sourced examples include deepseek-v4-flash at 0.02 yuan input cache hit, 1 yuan input cache miss, and 2 yuan output, with peak-hour prices doubling to 0.04 yuan, 2 yuan, and 4 yuan; deepseek-v4-pro at 0.025 yuan input cache hit, 3 yuan input cache miss, and 6 yuan output, with peak-hour prices doubling; Gemini 2.5 Flash-Lite at $0.075 input and $0.30 output per 1M tokens; Gemini 2.5 Flash at $0.15 input and $0.60 output per 1M tokens; Gemini 2.5 Pro at $1.25 input and $10.00 output per 1M tokens; GLM-5.2 at $1.4 input and $4.4 output per million tokens; DeepSeek-V4-Pro at $0.435 input and $0.87 output; GPT-5.6 Sol at $5 input and $30 output; Claude Fable 5 at $10 input and $50 output; Zendesk AI Agent at about $1.5 per automatic resolution under committed volume and about $2 on demand; HubSpot Breeze Customer Agent at $1 per handled conversation before April and $0.5 per successfully resolved conversation after April. Exact local taxes and billing terms are not disclosed in sources.
• Closest Western equivalents: OpenAI API; Anthropic API; Google AI Studio / Gemini; GitHub Copilot; HubSpot AI agents; Zendesk AI Agent; OpenRouter
• Data residency: The sources do not provide full data-residency terms. They do say that in the free version of Google AI Studio, Google reserves the right to use submitted prompts and responses to improve its models, while for users in the European Union, free-tier data is not used for training. For the other products, data-residency and training-use details are not disclosed in sources.

Sources

• geekpark DeepSeek 的「斩杀线」，正在重塑 AI 行业的世界观 https://geekpark.net/news/368556
• woshipm Token 便宜了 99.7%，为什么你的 AI 账单还在涨 https://woshipm.com/ai/6442490.html
• appinn DeepSeek V4 调价预告：高峰时段价格翻倍，正式版7月中旬发布 https://appinn.com/deepseek-v4-pricing-preview-2607
• woshipm 高强度写代码的场景里，最便宜的 Token 反而是 Codex 套餐 https://woshipm.com/ai/6442487.html
• appinn 谷歌开始提高免费 Gemini API 配额，部分模型已达每分钟 100 万 Token，但不是全部 https://appinn.com/google-increases-free-gemini-api-quota-1m-tokens