Artificial-intelligence users and businesses are increasingly paying for every exchange with a language model, but many are using expensive systems for tasks that cheaper models could handle. A guide published by Decrypt examines how to reduce that waste by matching jobs to models, limiting unnecessary context and putting controls around automated agents.
The issue matters because AI services generally charge by the token—a small piece of text used to measure both prompts and responses. A token averages roughly three-quarters of a word. For subscribers, token use can determine how quickly a monthly usage limit is reached. For developers using application programming interfaces, or APIs, token consumption directly affects the bill.
Why model choice can change the bill
The price gap between AI models can be substantial. Anthropic’s most capable model costs about 100 times more per token than its smallest model. OpenAI’s pricing also spans a similar range, from $0.10 to $10 per million input tokens, according to the source material. That makes a premium model an expensive choice for routine work such as drafting emails, summarizing text or rewording copy.
The cost problem has also appeared inside large technology companies. Meta employees used 60 trillion AI tokens in 30 days earlier this year while competing on an internal leaderboard known as “Claudeonomics.” Tech columnist Gergely Orosz estimated the usage at about $900 million at Anthropic’s list prices, although Meta is likely to pay less under its commercial arrangements.
Similar token-use contests have reportedly existed at Microsoft, Shopify and Amazon. At Amazon, employees acknowledged using an internal AI tool for irrelevant tasks to increase their totals, prompting a senior vice president to tell staff not to use AI merely for the sake of using it. Uber, meanwhile, used up its 2026 budget for AI tools in roughly four months.
The newer interpretation of “tokenmaxxing” is therefore not to consume as many tokens as possible, but to extract more value from each one. Bain & Company consultant Jue Wang told the Associated Press that many companies and users default to Anthropic’s high-end Opus model even for basic jobs.
The simplest ways to cut usage
The first step is to test a less expensive model against the work it is expected to perform. Users can run five recent tasks through a smaller model and a flagship model side by side. If the cheaper system produces an acceptable result, it can become the default, while the more advanced model is reserved for difficult reasoning or decisions where an error would be costly. The source says this approach can reduce API costs by up to 99% in suitable cases.
Clear prompts can also prevent repeated requests. Asking a model to revise an answer several times resends the conversation and uses more tokens. A single instruction that specifies the required fields, format and fallback wording is more efficient. Users should also open a new chat when changing subjects, because the model rereads earlier conversation history. Anthropic refers to the deterioration caused by excessive accumulated context as “context rot.”
Output limits matter as well. Anthropic and OpenAI charge more for output than input on their leading models, so instructions such as “three bullets, under 100 words” or “only the changed lines” can reduce unnecessary responses. Free model tiers and local tools such as LM Studio or Ollama can be useful for experimentation, although the source cautions that free offers may disappear and that users should check data policies before entering sensitive information.
Extended thinking should be avoided for straightforward questions where it is unnecessary. Anthropic bills the full thinking process as output, including portions users do not see. Claude Pro and Max subscribers who enable usage credits can also set a monthly spending limit in the service’s settings.
Agents create new sources of waste
The cost challenge becomes more complicated when users deploy agents. An agent can call tools, search the web, run code, read the result and repeat the cycle until it decides the task is complete. In an orchestrated system, one model may divide work among several subagents, each with its own loop and memory. Anthropic says agent teams in Claude Code can use about seven times as many tokens as a standard session when teammates operate in plan mode.
For terminal-based workflows, the guide recommends starting with a cheaper model and escalating only after a specific failure. Unused conversation history can be cleared, while context should be compacted only when it is genuinely needed. Large spreadsheets should be processed with formulas or scripts rather than pasted into a language model, which can miscount despite its price.
Tool integrations can add hidden context. MCP servers, which connect models to services such as GitHub or Google Drive, add tool names and instructions to every session. Removing unused servers can reduce that overhead. Anthropic reported one example in which loading only the required tool definitions reduced an agent’s use from 150,000 tokens to 2,000.
For larger systems, caching stable instructions can lower the cost of repeated prompts. OpenAI and Anthropic also offer batch processing at a 50% discount, with OpenAI promising results within 24 hours. Developers can further limit risk by setting maximum steps, per-run dollar caps and clear definitions of completion. The guide cites Shopify’s use of circuit breakers to stop access when spending spikes.
Moneyiar market picture
Moneyiar market data at 9:48-9:49 p.m. on October 17, 2026, showed mixed performance among the listed digital assets:
- Compound (COMP): $21.2120723448, down 3.92% over 24 hours and 14.92% over seven days.
- Decimal (DEL): $0.4349783584, down 3.70% over 24 hours but up 11.47% over seven days.
- Prom (PROM): $4.9297408289, up 0.69% over 24 hours and down 6.98% over seven days.
- ConstitutionDAO (PEOPLE): $0.0075072471, down 4.76% over 24 hours and 8.37% over seven days.
Why cheaper AI has not reduced total spending
Lower prices do not automatically produce lower aggregate spending. The source cites Silicon Data’s Token Expenditure Index, which found that the price of a token has fallen by more than 90% since 2023 even as spending on large language models has doubled since late 2025. A Bain & Company brief said token costs halved between December 2024 and 2025 while consumption increased 450%.
That pattern is known as the Jevons paradox: when a resource becomes cheaper, people may use enough more of it to increase total spending. For AI users, the practical lesson is to combine lower prices with better routing, shorter context, caching, batching and spending limits. The result can be lower costs without automatically sacrificing quality.
