Decrypt's TokenMaxxing guide opens with a number that sounds like a typo: Meta employees reportedly used 60 trillion AI tokens in 30 days, racing on an internal leaderboard nicknamed "Claudeonomics." Gergely Orosz put that at roughly $900 million at Anthropic's list prices, though Decrypt notes Meta likely pays far less. The take for agent builders: the leaderboard measured burning tokens, not getting results. An agent with a wallet and a loop has the same failure mode, and it runs without anyone watching the scoreboard.
The spend problem, per Decrypt
The article's context is a pile of reported figures. All of it is Decrypt's reporting, so treat it as claimed, not checked by us:
- ▸Uber reportedly burned through its 2026 AI tools budget in about four months.
- ▸Bain & Company's Jue Wang told the Associated Press that "so many companies, so many users, default to using Opus for everything, including generating emails."
- ▸An Amazon senior vice president told staff: "Please don't use AI just for the sake of using AI."
- ▸Per Silicon Data's Token Expenditure Index, per-token prices have fallen more than 90% since 2023, yet LLM spending has doubled since late 2025. Decrypt calls it the Jevons paradox: cheaper units, more units.
Cheaper tokens did not shrink the bill. They removed the friction that used to make people ask whether a call was needed.
Why agents make it worse
A chat user wastes tokens one message at a time. An agent loop thinks, acts, reads the result and repeats, resending the whole conversation each pass, as Decrypt describes it. Add orchestration and the multiplier shows up in the vendor's own numbers: Anthropic says Claude Code agent teams can use about seven times more tokens than a standard session when teammates run in plan mode.
Now give that loop a wallet. Inference is the cost you can see in a bill. The onchain cost, such as gas, swaps and x402 payments, is the one an agent can trigger by itself. If the model bill needs a cap, so does everything the agent can sign.
The tips that transfer to builders
Decrypt's guide runs from beginner to advanced. These are the ones that map straight onto an agent stack.
1. Route by job, not by habit. Test five recent tasks on a small model and a large one, and keep the small one if it passes. Decrypt's worked example: 1,000 invoices at 2,000 input and 200 output tokens each cost about $30 on its top-tier model and about $0.30 on its smallest, before any optimization. Routing alone cut the bill to $3.90, even after the larger model redid 300 failures. Those are the article's figures and its pricing chart, which we could not independently check.
2. Cache what never changes. Put fixed instructions first, mark them with cache_control, and confirm cache_read_input_tokens shows up after the first call.
3. Batch what can wait. Decrypt reports a 50% batch discount from both OpenAI and Anthropic, with OpenAI promising results within 24 hours.
4. Trim your tools. Anthropic reportedly cut one agent's usage from 150,000 tokens to 2,000 by loading only the tool definitions it needed. Decrypt's advice is to check /context and /mcp and switch off servers you don't use.
5. Let code do the math. Ask for a script, run it, and hand back only the result. Models miscount.
6. Put the agent on a leash. Set step limits and dollar caps. Claude Code supports a --max-budget-usd flag, and Decrypt says Shopify uses circuit breakers that cut off developers whose spending spikes within a day. It also warns that OpenAI's alerts don't stop traffic, so use hard project limits.
The Sato turn: a budget is a wallet control
Item 4 is the one that bites onchain builders. Every MCP server you attach adds tool definitions to context, and a crypto agent's stack tends to collect them: a wallet, a DEX skill, a price feed, a payment rail. More tools is not more capable. It is more tokens per loop.
So build the stack deliberately:
- ▸Pick the smallest set of MCPs and skills the job needs. Describe the agent at satohub.ai/build and Sato maps the stack, so you start from a short list instead of a pile.
- ▸Treat the model budget and the wallet budget as one design. A per-run dollar cap on inference and spend limits on the wallet are the same idea at two layers.
- ▸Check what you're attaching. The Sato Score on a listing is a transparency and liveness signal, meaning how open, active and verifiable something is. It is not a quality or safety grade, and it won't tell you how many tokens a tool definition costs. Read the repo.
What to watch
- ▸Whether per-token price drops keep outrunning usage. Silicon Data's index and Bain's June brief, which Decrypt cites as token costs halving while consumption grew 450%, say they haven't so far.
- ▸Budget flags as defaults. A hard dollar cap should be a first-class setting in agent runtimes, not a flag you find after the invoice.
- ▸Agent payments with limits built in. As agents pay per call over rails like x402, the same discipline has to reach the wallet side.
The prompt is the cute part. The stack, and the cap on what it can spend, is where the work lives.