Tokens keep getting cheaper and AI bills keep binding. Count verified tasks instead.

The first week of November 2026 produced an unusual cluster of AI spending news. Bloomberg reported that Apple cut the daily AI budget for some engineers from $300 to $200 in October and is weighing cuts to $100 for some teams and $25 for others. Two days earlier, Bloomberg reported that Goldman Sachs had limited Claude Code and GitHub's AI coding tools mainly to its roughly 12,000 engineers, taking them away from most traders and bankers. And Revolut said it had cut its AI coding inference bill by about 90% by running open-weight models itself. Our thesis is simple. Per-token prices are falling fast, but agentic and reasoning workloads multiply tokens faster, so budgets now bind. Caps and self-hosting are the first responses. Neither answers the real question, which is what a finished, checked piece of work costs.

Start with what is claimed versus measured. Apple has not commented, and employees told Bloomberg that few people used the full $300, so the savings may be modest. Apple is also asking some teams to move to cheaper models or on-device AI. Goldman framed its change as productivity and risk allocation, not cost. A spokesperson said the bank made the tools available to "determine the best use cases" and then placed them with staff "best suited to leverage them for the long term." Quant "strats" kept access. So the Goldman case is evidence that firms are now rationing AI tools by role, not that the bill forced the move.

The price side of the ledger is not in dispute. Epoch AI found that the price of reaching a fixed capability level fell between 9x and 900x per year depending on the benchmark, and about 40x per year for GPT-4-level performance on GPQA Diamond. a16z estimated a 10x annual decline for models of equal quality. But those are prices per token at constant capability. Nobody holds capability constant. Teams move to the newest model, and the newest models think out loud.

Token volume is the other side, and it is growing in the places enterprises care about. Epoch measured reasoning models' response lengths growing about 5x per year, with a 90% interval of 2x to 13x, against 2.2x for non-reasoning models. In the OpenRouter and a16z study of more than 100 trillion tokens, reasoning models went from a negligible share of tokens in early 2025 to more than half. Programming rose from about 11% of token volume to more than 50%. Average prompts grew from about 1,500 to over 6,000 tokens, and coding requests often exceed 20,000 input tokens. An agent that rereads a large repository context on every tool call spends most of its budget on input it has already seen. At Claude Opus 5.5 list prices of $4 per million input tokens and $20 per million output tokens, $200 buys 50 million input tokens or 10 million output tokens before caching discounts. A few parallel agents working long sessions can reach that in a day.

Self-hosting is the second response, and Revolut is the clearest case. David Abitbol, who leads its AI operations and infrastructure, told the AI Engineer conference in New York that some engineers had been spending $1,000 to $10,000 a month on commercial tools with little return. Revolut started with Z.ai's GLM-5.1 and now serves GLM-5.3 on 56 Nvidia B200 GPUs to about 2,000 engineers. Self-hosted GLM-5.1 came to about $0.30 per million tokens, roughly a tenth of the cost of Claude Code or Codex. For scale, Z.ai's own API lists GLM-5.3 at $1.40 input and $4.40 output per million tokens. Revolut's gains came from NVFP4 quantization, separating prefill from decode, and tuning, not just more GPUs. The caveats matter. A self-hosted per-token figure depends on how busy the cluster is. Revolut has not published full hardware, staff and operations costs. GLM is text-only, so vision work still goes to Gemini or GPT. And the reported 60% improvement in coding time and 70% in lead time are internal comparisons, not a controlled study.

Supply is moving the same way. On November 5, Thinking Machines released Inkling 2 Small, a mixture-of-experts model with 594 billion total and 17 billion active parameters, aimed at agentic coding and long-horizon work. It is available on Tinker with closed weights, and the company says it intends to release the weights under Apache 2.0 after external safety testing. Low active parameter counts cut the compute per token. They do not cut the memory needed to hold 594 billion weights, which is why self-hosting remains a capital decision. Meanwhile Meta acquired Conductor, whose app runs several coding agents in parallel workspaces. Terms were not disclosed. Tools that run more agents at once are tools that spend more tokens at once.

Does the spend buy output? The best controlled evidence is thinner than the budgets suggest. In METR's 2025 randomized trial, 16 experienced open-source developers working on 246 real issues took 19% longer when AI was allowed. They had expected a 24% speedup and still believed afterward that AI had sped them up by about 20%. METR's late-2025 follow-up with 57 developers and more than 800 tasks pointed the other way: an 18% speedup for returning developers (95% interval of 38% faster to 9% slower) and 4% for new recruits. METR calls this only very weak evidence of the size of the change. Many developers declined to work without AI, and 30% to 50% withheld tasks they did not want to do unaided, so the true gain may be larger. The honest summary is that gains are probably positive now, poorly measured, and far smaller than self-reports.

That is why cost per token is the wrong metric, and cost per seat is not much better. The number that matters is cost per verified task: total spend on an attempt, including retries, reasoning tokens and the human review it triggers, divided by the share of attempts that pass an independent check such as tests, a reviewer or a reproducible benchmark. A model at a tenth of the token price that passes half as often, or produces work that needs twice the review, is not a 90% saving. A cap of $200 a day is not a policy on its own either. It rations attempts without asking which attempts succeed.

Verification has its own trap. Sampling more and letting a checker pick the best answer raises the bill without raising the evidence when the checker shares the generator's blind spots, as we argued in our piece on test-time compute. Our verifier-budget tool estimates how many samples a verifier with a measured blind-spot rate can safely support. For enterprises the practical steps are plain. Log tokens per task, not per user. Report pass rates against checks the agent cannot game. Route easy work to cheap or self-hosted models and keep frontier models for tasks where they measurably pass more often. Set caps after measuring cost per verified task, not before.