Agents Just Became AI's Biggest Customer — Now the Industry Is Racing to Make That Affordable

Three stories from the past two weeks don't share a company or a model release, but they share a number: tokens. One shows just how much more expensive an agent turn is than a chat turn. One shows two vendors racing to close that gap with cheaper, purpose-built agent infrastructure. And one shows a Fortune 10 company already running the economics at scale. Read together, they're less "new model dropped" and more "the industry just noticed what agents actually cost, and is scrambling to fix it."

1. Agents Have Become AI's Biggest Customer

OpenRouter's platform data shows agentic token consumption jumped roughly 14x since February 2026, rising from about 0.51 trillion to 7.3 trillion tokens a month, while ordinary human chat usage grew only about 2.8x over the same stretch. Agent traffic crossed human traffic in raw token volume around February 1 and hasn't looked back. The reason isn't subtle: a single agent task — plan, call a tool, read the result, call another tool, revise — burns somewhere between 5x and 15x the tokens of one human turn, because every intermediate step re-reads context that a chat reply never has to. The one bit of good news in the data is that nearly 70% of that volume is cached prompts billed at a steep discount, so real spend isn't rising as fast as the raw token count implies — but it's still rising, and it's rising because of a workload type that barely existed at this volume a year ago.

2. The Industry's Answer: Purpose-Built, Cheaper Agent Infrastructure

Two vendors shipped direct responses to that cost curve in the same week. Writer released Palmyra X6, a flagship model built on top of Z.ai's open-weight GLM-5.2, paired with a rebuilt agent "harness" and new governance tooling aimed squarely at runaway token spend from marketing and revenue-team workflows. The company claims 52% lower cost and 48% faster output on multistep agent tasks versus its prior stack — notable less for the specific number than for the fact that a model vendor is now marketing "cheaper per agent task" as the headline feature, not raw benchmark scores.

OpenAI moved on a different lever: it's sunsetting the Assistants API (retirement landed August 26) in favor of the Responses API, which it's positioning as the permanent home for all new agent capabilities. The pitch is architectural, not cosmetic — better cache utilization across multi-step calls, and automatic routing to a cheaper model tier when a lighter one can handle the step. OpenAI's builder's guide for GPT-5.6 makes the same point from a different angle: on Agents' Last Exam, GPT-5.6 Sol at "low" reasoning effort beat GPT-5.5 at "high" effort with the same harness, meaning the cheapest configuration is sometimes also the better one.

// Assistants API: persistent thread state, one call per step
const thread = await openai.beta.threads.create();
await openai.beta.threads.messages.create(thread.id, { role: 'user', content: query });
const run = await openai.beta.threads.runs.create(thread.id, { assistant_id });

// Responses API: chainable, cache-friendly, model tier is a parameter
const response = await openai.responses.create({
  model: 'gpt-5.6-sol',
  reasoning: { effort: 'low' },       // cheaper tier, often equal or better on agent tasks
  input: previousResponse.output,     // chains directly, reuses cached context
});

If your stack still calls the Assistants API, the sunset date has already passed — this is a live migration, not a future one.

3. Proof the Economics Work at Production Scale

The clearest evidence this isn't just vendor marketing is Toyota North America, which has moved past 50 production agents built on LangChain's Deep Agents and LangSmith. The most concrete example, GearPull, sits on top of terabytes of manufacturing data in vector databases and serves every North American plant: an engineer types a problem description and gets a troubleshooting answer in about 10 seconds, a workflow the company says has already produced millions of dollars in documented savings. That's the counterweight to the token-cost story above — the same agentic workloads driving up OpenRouter's numbers are, at least in this case, generating returns that comfortably outrun the extra tokens they consume.

What This Means for Builders

None of these three stories is a reason to panic, but together they're a reason to check your assumptions. If you're running agent loops in production, your token bill is probably growing faster than your chat traffic would suggest, and it's worth confirming how much of that spend is actually hitting cache versus recomputing context on every step. If you're still on the Assistants API, migrate now rather than after something breaks. And if you're building an internal agent business case, Toyota's GearPull is a useful reference point: the payoff shows up not from the agent being smarter, but from it being fast and narrow enough that the token cost stops mattering.

— Maya