Uber's engineering org adopted Claude Code between December 2025 and March 2026. Claude Code usage jumped from 32% to 84% of Uber's 5,000-engineer organization in that window. By April, the entire annual AI budget was gone.
Monthly API costs per engineer were running between $500 and $2,000. Token prices had been falling the whole time. That didn't help.
The paradox resolves once you understand one number. According to Gartner's March 2026 analysis, agentic models require between 5 and 30 times more tokens per task than a standard chatbot. Enterprises that scaled past the pilot phase discovered this multiplier only after their production bills arrived.
The per-token price is not the relevant price. The per-task price is. And most teams have never measured it.
Why agents burn tokens so differently from chatbots
A reasoning agent doesn't just send a prompt and receive a completion. It sends the full accumulated context - system prompt, conversation history, all intermediate results - to the model at every step. By step 20 of a multi-step task, the agent is paying for the same context 20 times over.
According to research from the Stanford Digital Economy Lab, re-sent context accounts for 62% of total agent inference bills. Most of what you are paying for is the model re-reading what it already knows.
Agentic coding workflows average 1-3.5 million tokens per task, including retries and self-correction loops. A single debugging session with a frontier model can consume 500,000+ tokens. That is not a ceiling; it is an average.
When an agent hits a test failure, tries a fix, fails again, and tries another approach, each retry cycle is a full round-trip at the current, inflated context size. Three failed attempts at turn 40 don't just cost 3× a single turn - they cost 3× a turn that's already carrying 30,000+ input tokens of context.
The billing model most teams were using treated this like a chatbot. It was not.
Anthropic's billing split: what happened and what it signals
On May 13, 2026, Anthropic announced it would separate its subscription billing into two pools starting June 15.
The new credit would cover programmatic usage - Claude Agent SDK, claude -p, Claude Code GitHub Actions, and third-party apps built on the SDK. Chatting with Claude on the web, desktop, or mobile, and using Claude Code in the terminal, would stay on existing subscription limits.
The new monthly Agent SDK credit was sized at $20 for Pro, $100 for Max 5x, and $200 for Max 20x - at full API rates, with no rollover. For heavy agentic users, this represented a significant effective cost increase.
On June 15, Anthropic confirmed that the planned move is no longer happening - the change was paused the same day it was supposed to take effect. Developer backlash was immediate and pointed. Anthropic said they are revising the plan and will share advance notice before any future change takes effect.
The pause doesn't change what the announcement revealed. The framing isn't a price increase; it's the end of a structural subsidy.
The credit system was designed to end the compute arbitrage era, where a $20 Pro subscription could run agent workflows that would cost $500 on a direct API key. That arbitrage will close. The only open question is when.
The routing gap that's costing teams the most
Analysis of 2.4 billion enterprise API calls in Q1 2026 found that organizations running a tiered model architecture achieved a median blended cost of $2.31 per million tokens. Organizations routing every workload to frontier models paid $18.40 per million tokens. That 87% gap is the direct financial consequence of one architectural decision made at the start of the deployment process.
Frontier AI is priced for frontier tasks: complex multi-step reasoning, long-context synthesis, judgment under genuine ambiguity. Classification, extraction, intent detection, document summarization, and the routine logic that makes up the majority of most enterprise agentic workflows do not require frontier capability.
One concrete example from a 35-engineer SaaS team: they had been running Claude Code, Cursor, and a custom autonomous bug-triage agent for four months. Their April 2026 bill was $87,000. After routing triage tasks to a smaller model and escalating only hard cases to Opus, adding context pruning, and setting $100/day hard caps per developer, their May bill was $24,000.
Engineering productivity was unchanged, measured by sprint velocity.
No product changes. No capability loss. The savings came entirely from routing discipline and a per-developer cap that stopped runaway sessions before they finished.
What a team should actually measure
A simple chatbot query triggers one inference call, but an agentic workflow - calling external tools, verifying outputs, self-correcting - can trigger 10 to 20 model calls for a single user-initiated task. That changes the business model: the relevant unit is no longer cost per prompt, but cost per completed task.
Most teams have never calculated that number. The four things worth measuring, in order of leverage:
Cost per task, by workflow type. Separate agentic coding sessions from RAG lookups from inline autocomplete. The distributions are completely different.
Re-sent context as a share of the bill. If it's above 50%, prompt caching is the first fix, not model selection.
Per-user distribution. If 20% of your engineers are driving 80% of the agentic token volume, that might mean they've found genuinely productive workflows worth spreading - or it might mean they're running expensive models on tasks where a cheaper option would work fine. You can't tell without the data.
Model tier per step. Not per workflow, per step. Some steps in an agent loop need a frontier model. Most don't.
Anthropic's own admin guidance points in the same direction: the Agent SDK credit is "sized for individual experimentation and automation," and teams running shared production automation "should use Claude Platform with an API key for predictable pay-as-you-go billing." That's the company saying directly: if you're running production agents, subscriptions were never the right unit.
A tool like Beagle that runs inside Slack and Teams can surface this kind of per-task attribution for conversational workflows specifically, without requiring a separate observability stack. But the measurement habit matters more than the tool. The deployments that didn't produce budget surprises shared one characteristic: token volume was modeled per workflow type before the architecture was finalized - not with precision, but separately from the chatbot assumption, with a realistic loop count and realistic context depth. The teams that skipped that step are the ones reconciling unexpected spend after the fact.
The Gartner forecast says token prices will fall another 90% by 2030. But as token costs fall, token consumption is expected to rise faster. Overall inference costs are expected to increase. Cheaper tokens do not make undisciplined agents affordable. They make undisciplined agents easier to scale - and harder to catch on the bill.