Tokenmaxxing Is Dead: How to Cut Your LLM Bill 50–70%

For two years the implicit metric for "are we serious about AI" was how many tokens you burned. That era just ended — and the way it ended is good news for anyone who cares about the bill more than the leaderboard.

Fortune recently called time on tokenmaxxing: the corporate habit of treating raw token consumption as a proxy for innovation, complete with internal leaderboards. The reporting is blunt about how it went — teams spinning up agents to do busywork purely to keep usage stats up, Salesforce reportedly spending on the order of $300M a year with one model provider, Uber said to have burned through its entire 2026 token budget in four months. The throughline: more tokens did not mean more value.

What tokenmaxxing was

Tokenmaxxing was Goodhart's Law in real time: the moment token usage became the target, it stopped being a useful measure. You can hit any token number you like by sending the model more — more tools, more history, more retrieved context — whether or not any of it helps the answer. Usage went up and to the right. Outcomes didn't follow.

The correction isn't "use less AI." It's the realisation that a large share of what agents send on every call is waste the model never needed for that turn — and you were paying for all of it.

Why spend stopped meaning value

Look at one agent request. It typically carries the full tool catalogue (often dozens of schemas), the entire conversation so far, repeated JSON schemas, stale tool results, and old multimodal context. The model reasons over a thin slice of that. The rest is billable noise — and noise the model has to read past, which can hurt accuracy as well as cost.

That's the gap finance teams are now staring at: the line items are tokens, but the budget conversation is about shipped value. As one operator quoted in the piece put it, if you can't connect the spend to features that reached users, it's hard to justify. Closing that gap doesn't require sending fewer requests. It requires sending leaner ones.

Cut the waste, not the capability

This is the part Orqen was built for. It sits between your SDK and your LLM provider and, on every call, forwards only the tools the turn actually needs and compresses history and verbose schemas — then hands the provider a smaller request to your same model. Same model, same answer, fewer input tokens. Typical prompt savings land in the 50–70% range, with no change to your agent code. Optionally, with an orqen/* routing mode, Orqen can also send simpler turns to a cheaper model — that's your call to make, not something it does behind your back.

# No rewrite. Point your existing client at Orqen.
client = anthropic.Anthropic(
    api_key="sk-orq-YOUR_KEY",
    base_url="https://api.orqen.app",
)

# Same request body. Orqen forwards only what the turn needs,
# compresses the rest, and bills you nothing extra to do it.
response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=messages,
    tools=all_tools,   # 40 tools in, ~15 forwarded
)

The point isn't a smaller number for its own sake — that would just be tokenmaxxing in reverse. The point is that the tokens you cut were never doing any work.

Draw the line from spend to value

Cutting the bill is only half of what the post-tokenmaxxing market is asking for. The other half is proof. Orqen attributes cost honestly on every request: what you actually paid versus what you would have paid without it, measured against the provider's real bill — not an estimate. Provider-side prompt caching is preserved and never double-counted, so the savings number isn't inflated.

That turns "we spent a lot on AI" into a line a CFO can read: dollars saved per dollar of usage, per agent, per month. It's the ROI attribution that token leaderboards could never give you.

What Orqen won't do for you

What it does do is remove the most embarrassing line in the post-mortem — the one where you were paying full price for context the model ignored — and it does it in the time it takes to change a base URL.

Cut the bill, prove the ROI