Tokenmaxxing Won't Survive Planning Season

Depending on the size of the company, planning for next year usually starts somewhere between mid-September and mid-November. Finance teams sit down, look at what was budgeted, what was actually spent, and what was the ROI of those investments. This year, AI is going to be one of the biggest line items in that conversation, and it will lead to very interesting conversations. The data is starting to make this hard to ignore. According to Ramp, the average business is spending 13 times more on AI tokens than it was in January 2025. AI is still only one to two percent of total spend even for the heaviest spenders, so on the surface this looks manageable. But the slope is what matters. Ramp's lead economist, @arakharazian, told DealBook this week that the trend is "unsustainable for most businesses" and that it is now "incumbent on firms to find ways to control that cost." Some companies are already making changes, some of them are routing work to cheaper models, switching some workflows to Chinese models for cost reasons, or starting to route different workflows to different models, putting platforms like Cursor and OpenRouter https://openrouter.ai/ between themselves and the frontier labs so they can swap without rewiring their stack.

The most public version of the problem is Uber. There is a viral snippet from an interview with an exec claiming they blew through its entire 2026 budget for Claude Code and Cursor inside the first four months of the year. The line that got picked up everywhere in social media was that the cost was getting "harder to justify" because there was no clean way to draw a line between the spend and any measurable improvement in the consumer product. 95% of Uber's engineers use AI tools every month and a reported 70% of committed code is AI-generated, and yet the executive answering for the bill could not point to what the company got for it. Microsoft is the other tell. This month, Microsoft canceled most of its Claude Code licenses in part for cost reasons and started pushing its engineers onto GitHub Copilot. I've personally heard Amazon doing the same and forcing teams to use Kiro, their own agentic IDE. Some of the largest companies in the world, both deep in the AI story, both visibly pulling back on the most expensive tools in the category, in just a couple of months. The same thing is showing up at the company level, away from the headlines.

@Aiswarya_Sankar, who works with a lot of builders, put it bluntly on X:

That number: eighteen cents of shipped product per dollar of tokens, is the real ratio behind the prettier headline charts. A lot of that 82% is not waste in the cartoon sense. It is the agent running in a loop, retrying, second-guessing itself, fixing the bug it created two turns earlier, and then producing a working diff that the human author would have written in less time. I have been running this experiment on myself. In Q1, I started looking at my own Anthropic bill and realized that, some days, I was spending over $1,000 a day in extra usage. I'm one of two building in our company. I have ten worktrees always running, building new features for grep.ai, evals, a couple of aspirational zero-to-one projects, the usual stream of bug fixes, and my own non coding automations (which grow by the day). Yes, some of that bill is probably skill issue. It is what it is. What did I do? I got smarter about usage limits and routing. I started load-balancing between Claude Code and Codex. I always use Claude Code first, but for small well-defined bug fixes I started routing Claude Code to cheaper models like Kimi 2.6. Codex /goal became a staple in evals and zero-to-one aspirational projects (read about my absurd bike coding setup here). I am spending way less today, trying not to hit extra usage, and the spend is doing more work per dollar than it was when I was just leaving the most capable model on for everything. That shift, from "always use the smartest thing" to "route the work," is the shift I think every team is about to have to make.

There are two reasons this topic will be a problem in planning meetings. One, is that most of the spend was justified at the time on vibes, not on outcomes. Coding agents have an unmistakably real productivity story to point at. Some internal automation probably did remove real manual work. Some agentic workflows probably did replace expensive operational headcount in a small scale test. But for a lot of the rest, the answer is going to be much harder to defend. If the strategy for the year was "use the smartest model everywhere," then the cost is going to be high and the value is going to be uneven. Second, smart finance teams are very numbers-driven. Measuring ROI of this spend will be very hard. Where ROI will be justifiable, it will probably be a very lagging indicator, and in most other cases, "the model is really good" is not going to fly.

It is worth thinking about the incentives behind all of this. The frontier labs are not optimizing for your company's cost structure. OpenAI, Anthropic, Google, and the rest are chasing AGI, which is the right thing for them to do. Their job is to build the most capable model in the world. Most capable, is in fact, turning into "more expensive". While Opus 4.7 price per token matches 4.6, the model "thinks" more, producing more tokens per job. Anthropic acknowledged this by increasing usage limits when it released 4.7. GPT 5.5 is 2X more expensive than 5.4. Google's Gemini 3.5 family of models (which arguably could benefit from adding some price pressure to competitors) is significantly more expensive than 3.1. This is at the model layer. Once you account for the rapid adoption of "agentic" and "long-running" workflows, long context windows, tool calls, retries, and the multi-step loops an agent in real-world scenarios run, the cost is not one prompt, it is the planning, the searching, the reading, the writing, the verification, the retries, and sometimes a process that runs for minutes or hours. Then, there comes the second order effects. Those who adopted AI tools first on some sort of seat-based model, are now being surprised with massive bills as they do their own reckoning and move to usage-based models (looking at you AI code review companies). Furthermore, you have companies incentivizing this tokenmaxxing. Jensen Huang has said publicly that he expects an Nvidia engineer on a $500,000 salary to spend at least $250,000 a year on tokens. That is the cultural water everyone has been swimming in. Half your salary, in tokens, as a sign of seriousness. Tokenmaxxing was not an accident. It was encouraged from the top of the value chain.

At the same time, the smaller models are getting much better. Open-source and cost-effective releases: Kimi, MiniMax, GLM, DeepSeek, Qwen, and other Chinese open-source models have become genuinely useful for narrow, well-scoped work. US-based labs are also fueling an open source renaissance: DeepMind's Gemma 4 and Poolside's Laguna are two recent examples I stumbled upon in my timeline as models that are surging in usage and the promise of a cheaper / better future.

@natolambert has been making a sharper version of this argument:

the open ecosystem still lags the frontier on agentic capability by something like a year, but the gap is closing where it matters most for cost-sensitive work, and open models are going to be used more and more for "automated, enterprise agents and low-cost domains.

That is the shape of the next architecture. It is not "one giant model does everything." It is smart model routing, task-specific harnesses, maybe fine-tuned smaller models, evaluation-driven workflows, and a lot more discipline about where the expensive tokens actually need to go. We have learned a version of this at Parcha Labs the hard way over the last couple of years, and I have written about this before (in fact, two years ago: Agents Aren't All You Need. Even before "agentic" was all the rage, we built the frameworks that enabled us to always the most capable models for the parts where capability is actually the bottleneck: planning, ambiguous reasoning, synthesis, the judgment calls where the cost of being wrong is high. And then relying on cheap, heavily prompted and orchestrated models for extractions, web summaries, formatting classification, etc with a tight evaluation loop behind it. I am pretty sure we were one of the biggest fans of Haiku 3.5 until the day it was sent to a farm upstate.

We are seeing the same thing today, except what was the model then is the "harness" now. This is the part of the next twelve months that I think most teams are about to be forced into. The internal metric of "tokens consumed" is going to stop being a sign of progress. The better question, and the one that is going to start showing up in QBRs, is whether the team produced the same or better output, with the same or better reliability, at a lower cost. That shift is going to require teams to do actual engineering instead of wrapping a frontier model in an "automation" box. They will need evaluations. They will need to know which steps require frontier and which steps are mostly pattern matching. They will need to measure cost per successful outcome instead of cost per token, they will need fallback paths, model comparisons, routing logic, and test sets that tell them whether the cheaper option is actually good enough. And they will need to be smarter than throwing an out-of-the-box harness into every problem.

None of this is an argument against spending on AI. If anything, I think the spend keeps going up. Companies are going to keep pushing into automation, keep building agents, and keep using AI in coding, operations, support, compliance, sales, research, and finance. The first wave proved that LLMs can do useful work. The second one proved that when surrounded by a well-engineered harness, AI can be a powerful replacement for many processes that were driven by very smart humans (e.g. coding). The next wave is about making that harness efficient and smart enough to run everywhere, and that is the wave where most of the leverage actually shows up. Planning season is going to expose the gap between those two things. Some of this year's AI spend created real leverage. A lot of it was tokenmaxxing dressed up as innovation, and that part is going to come out in the wash.

My prediction is that the next twelve months bring a reversal, but not the one some people are bracing for. It will not be a reversal away from AI. It will be a reversal away from undisciplined AI usage. More companies will move toward smarter harness leveraging smaller models, open-source models, smart-routing platforms, and even back to purpose-built workflows. More teams will ask, honestly, whether they really need the most expensive model for a given task. More vendors are going to be forced to show ROI beyond a usage dashboard. And more AI products are going to compete on cost per useful unit of work, not on raw intelligence. Who knows... we may even see Anthropic giving Haiku more love and OpenAI investing more in their mini and nano models (we can only hope).

That is the part I am actually excited about. The companies that come out ahead are not going to be the ones that burned the most tokens. They are going to be the ones that figured out how to turn intelligence into real leverage without lighting their AI budget on fire.