In the era of AI cost engineering adopted by enterprises, Brian's tweet at the end of June https://(x.com)/brian'armstrong/status/2070670644577280109? The value of s=20 is highly valued. After implementing a series of engineering measures, Coinbase has seen a continued increase in token consumption, but AI spending has decreased by nearly half. This means that the main battlefield for enterprise AI cost management is shifting from "restricting employees from using AI less" to "using engineering infrastructure to make AI more useful but cheaper". Let's talk in detail: 1. Key points of Brian's tweet content In Brian's X tweet, he stated that the goal is to achieve "AI spending remains stable while token usage grows exponentially," meaning that AI token usage grows exponentially while AI spending remains stable or even decreases; The method is not to increase usage friction, set more budget reminders, or lower usage caps, but to rely on better default models, automatic routing, and caching. There are five specific methods: 1）Better defaults， Not relying on usage cap. Engineers can still choose any model, but Coinbase has switched the default model to cheaper open weight models such as GLM 5.2 and Kimi 2.7, and issued the default strategy through an internal LLM gateway. 2）Better routing。 Coinbase will preprocess prompts in the custom harness and route requests to appropriate models based on tasks, model prices, and cache hit probabilities. Complex planning can use a frontier model, and executing classes, repetitive classes, and low-risk tasks may not necessarily require the most expensive model; Armstrong's viewpoint is that ultimately, models should not be manually selected by humans, but should be automatically selected by AI or routing systems. 3）Better caching。 This is the most direct cost leverage. Increasing the cache hit rate from 5% to 60% in LibreChat means that a large number of duplicate contexts or requests will no longer trigger complete new model calls. 4）Keep context lean。 He emphasized not to blindly compare, but to open new sessions, narrow file contexts, and disconnect unused tools when changing tasks. The goal is not to use fewer tokens, but to waste fewer tokens. 5）Better visibility。 Coinbase does not prohibit engineers from using tokens excessively, but rather makes the cost of AI usage transparent for everyone and establishes a responsibility mechanism of 'the more you spend, the greater the impact you will have'. The ultimate effect is that Coinbase's enterprise AI spending is nearly halved, and token usage continues to grow. This is of great significance. 2. The impact of adopting AI on enterprises The real signal of this tweet is that enterprise AI adoption is entering the stage of "cost engineering". In the past year, many companies' AI adoption logic was "let everyone use it first", even treating token consumption as a proxy for AI adoption. But now the bills are starting to get bigger, CFO、CTO、 The platform team will begin to ask: why must the most expensive model be used for the same code generation, code review, customer service summary, and research retrieval tasks? Reuters also reported that companies are reassessing the cost of using AI, with more and more CEOs believing that cheaper and smaller models can cover a large number of enterprise needs. For enterprises, the key to future AI adoption is no longer just "buying Claude, buying ChatGPT, buying Copilot", but building an "AI operation and maintenance layer". Model Gateway: Unified access to OpenAI, Anthropic, Google, open source/open weight, and private deployment models. Routing system: automatically distributed based on task difficulty, latency requirements, compliance level, cache probability, and historical success rate. Cache system: Cache and reuse system prompts, code repository context, commonly used documents, and repeated Q&A. Evaluation system: Use evals to determine which model is "good enough" for a certain type of task. Cost visualization: Mapping token spend to teams, projects PR、ticket、 Customer work order. Governance system: Sensitive data, customer data, code IP, and regulatory requirements determine which requests cannot flow to external APIs. This means that the adoption threshold for AI will decrease. Previously, companies were concerned that the more employees used AI, the more out of control their bills would become; The paradigm given by Brian now is not to suppress usage, but to lower the cost per unit of intelligence. Once companies believe that "more AI usage does not necessarily mean higher costs," AI will move faster from pilot to daily workflows. But this will also widen the gap between enterprises: Companies that can use LLM gateway, router, cache, eval, and observability will gain ROI faster than companies that only buy a single supplier account. In other words, the barriers to AI adoption are shifting from "having models" to "having AI platform engineering capabilities". 3. The impact on token economy The token economy here cannot be solely based on the $/1M tokens listed on the model quotation table. What enterprises truly care about is: The cost of each successful task=input token cost+output token cost+cache cost+retry cost+tool call cost+manual verification cost. Brian's method is to reduce costs on each variable: The default model is cheaper, reducing P_input and P_output. Cache hit rate ranges from 5% to 60%, reducing duplicate input costs. Simplify the context and reduce invalid input tokens. Route to the 'sufficient model' to avoid using the highest end model for each task. Visualize costs, bind token spend with business output, and reduce invalid token maxxing. This will drive companies to shift from token maxxing to ROI maxxing. In the past, people boasted about how many tokens we used, but in the future, it will be more important to see how many merged PR were generated for every $1 token spend, how many customer service work orders were closed, how many labor hours were saved, and how much fault recovery time was reduced. Different models have varying capabilities, latency, context, reliability, and security commitments. But enterprise procurement will increasingly use "task level cost-effectiveness" to compare, rather than default to choosing the strongest model. 4. Impact on the input-output ratio of AI This tweet has a significant impact on AI ROI as it transforms the AI cost curve from linear to optimizable. The traditional intuition is: AI usage ↑ → token usage ↑ → AI billing ↑ The path presented by Brian is: The usage of AI has increased, but through routing, caching, context management, and low-cost models, the effective token cost can be reduced or remain stable The improvement in ROI comes from both sides: 1) On the cost side, the same task may switch from the most expensive frontier model to the cheaper open weight model; Repeating context to cache; Reduced low value output; Invalid long context reduction. 2) The revenue side is on the rise, without relying on cap restrictions on employee usage, reducing the psychological resistance of 'I'm afraid of exceeding the budget, so I don't use AI'. The increasing frequency of AI usage makes it easier to release actual productivity benefits. But ROI will not automatically increase. Low priced tokens may also lead to new waste, especially in agentic coding, long context research, and multi-agent workflows, where token consumption can be very fast. Reuters mentioned that although token prices are decreasing, the cost of completing a task may actually increase due to companies shifting towards usage based pricing, increasing task steps, and longer inputs, resulting in more unpredictable bills. So what companies need to look at is not "the cheaper the token, the better", but three indicators: 1) successfully completing tasks, 2) saving time, and 3) producing results. For example, in software engineering scenarios, you can see: If a team spends $100000 on tokens but delivers an additional $2 million worth of software output, this is a good ROI. On the other hand, spending $10000 in tokens but only generating a large amount of low-quality code and increasing review bursts is poor ROI. 5. What does it mean for a model company? Brian's tweet implies' medium to long-term pressure 'for AI model companies. For closed source Frontier Labs, the risk is that companies no longer treat the 'strongest model' as the default, but rather as a high-end tool. After a large amount of routine workloads are routed to the cheaper open weight/open source/self host model, the token volume may continue to grow, but the share of the most expensive API will decrease. If the model layer cannot prove a significantly higher task success rate, it will be considered as "over configuration" by the router. For open-source model and inference infrastructure companies, this is a positive development. http://Z.AI 、 Moonshot, DeepSeek, Mistral, Llama ecosystem, inference cloud, model router, LLM gateway, observability, and eval platform will all benefit. Because what companies really want to buy is not just models, but the system capability to 'complete tasks with the cheapest and best models'. It may also be beneficial for cloud providers. Even if the unit price of API tokens decreases, the usage of enterprise tokens may increase significantly, and the inference computing power GPU/ASIC、 Cache, vector database, logging, monitoring, and private deployment requirements will all increase. This It's a bit like the early days of cloud computing: unit computing prices have decreased, but total computing volume has exploded. 6. Will it drive a wave of token price wars in the future? From a personal perspective, there may be price wars, but not all tokens will engage in price wars, but rather 'layered price wars'. The three types of tokens that are most prone to price wars are: 1) Universal inference/execution token. For example, code completion, simple scripts, summaries, categorization, information extraction, format conversion, and regular customer service Q&A. These tasks have low dependence on the 'strongest model', and as long as small models or open weight models reach sufficient quality, prices will quickly fall. 2）cached input token。 Each company is already making cached input a low-cost entry point. The GLM-5.2 cached input is $0.26/1M, the Kimi K2.7 Code cache hit input is $0.19/M, and the OpenAI chat late cached input is $0.50/1M. The more popular caching becomes, the more enterprises will engineer repeatable contexts, and suppliers will be forced to continue lowering cache prices. 3) Mid end model token. Reuters reports that companies are embracing cheap models and routing tools, leaving complex tasks to premium models; The proportion of open-source tokens on OpenRouter has increased from 34% in January to 65% in June, indicating that traffic is shifting towards low-priced models. But Frontier tokens may not immediately enter a fierce price war. The strongest model still has several moats: Complex reasoning, system design hard coding、 We are still willing to pay a premium for high-risk tasks such as scientific research, financial analysis, and legal/medical care. The real computational cost of outputting tokens, especially reasoning heavy outputs, is higher. Enterprises will also consider security, auditing, data isolation, compensation clauses SLA、 Delay and context length payment. OpenAI, Anthropic, Google, and others can maintain comprehensive ARPU through enterprise contracts, subscription packages, batch discounts, reserved capacity, and agent platforms, rather than just lowering list prices. So what is more likely to occur is: Low end/mid-range tokens: significantly reduced prices, trending towards computing power costs+thin profits; Cached token: continue to significantly reduce cost; Frontier reasoning token: maintains premium, but will be forced to launch cheaper sub flagship/mini/fast versions; Enterprise contract: There are a large number of private discounts beyond the public price. In my personal opinion, Brian's tweet is not simply saying 'Coinbase has found a cheap model', but rather announcing the next stage of enterprise AI: From: Buy the strongest model+encourage everyone to use it more To: Multi model architecture+automatic routing+caching+cost visualization+ROI constraints It will accelerate the adoption of AI by enterprises, as it proves that AI usage growth and cost control are not contradictory. It will change the token economy as enterprises shift from $/1M tokens to $/successful tasks. It is also likely to trigger a token price war, but it mainly occurs on mid to low end, cacheable, routable, and replaceable inference tokens; Frontier reasoning tokens will still retain a premium, but the default usage share will be compressed. The future is not about not using expensive models, but about no longer default to expensive models. The model layer will be commodified by routers, and the real barriers for enterprises will shift towards data, processes eval、 Cache, governance, and AI platform engineering. This content is sponsored by @ BITstocks_CN. Buy BIT-16000+US stocks and ETFs on the US stock market, hold real positions, and enjoy dividends.