Token prices held flat in 2026. The bills climbed anyway.
For anyone running agentic workloads, the math underneath the bill changed in 2026, and the architecture call matters more than the rate card.
Uber reportedly burned its entire 2026 AI budget in four months on Claude Code.1 The published per-token rate had barely moved. Microsoft is working through the same arithmetic. Both ran into a structural problem most cost models have not accounted for: the bill follows the tokens, and the way we use models now spends them at a scale and variance nobody really budgeted for.
Two forces are at work. The first is mostly accounting. Vendors raised effective prices while holding the published rate steady: Anthropic's new tokenizer can emit up to 35% more tokens for the same input,2 subscription session limits drain faster at weekday peak by design, Microsoft raises 365 commercial pricing July 1, and OpenAI repriced tool and container sessions earlier in the year.3 The subsidy era is over. Providers are now optimizing margins.
The second is structural and harder. A Stanford and Michigan study measured agentic coding at roughly 1,000x the tokens of a chat session.4 The driver is input, not output. The agent re-reads its entire accumulated context before every action, so cache-read input tokens dominate the cost. Same agent, same task, run twice: the expensive run is typically about 2x the cheap one, with a tail that stretches toward 30x on the hardest problems. Agents cannot predict their own cost in advance and systematically underestimate it. If you promise a customer a fixed price on agent work, you are underwriting a bet you cannot price.
Conventional software gets cheaper per unit as it scales. Agentic software gets more expensive per task as the context grows, because the input you carry forward keeps climbing. That inversion likely breaks a standard software P&L.
Hygiene before architecture
Routing layer. A gateway like OpenRouter gives you one OpenAI-compatible endpoint into 300-plus models across 60-plus providers, with unified cost visibility, automatic fallback when a provider degrades, and the freedom to move workloads off a vendor that reprices. Pass-through pricing plus about 5.5% on credit top-ups, so it is convenience and resilience, not arbitrage. Regulated and data-residency workloads need their own scrutiny.5
Tier the model to the task. Frontier-for-everything is the single most expensive default. Across billions of enterprise calls, tiered runs near $2.31 per million tokens blended; frontier-for-everything runs about $18.40.6 Reserve the expensive models for genuine ambiguity, and offload high-volume, low-stakes work to a local or budget tier.
Cache and compress the context. Caching the stable material cuts cached input by roughly 85 to 90%. Put your system prompt, tool definitions, and long reference documents at the front of the request and mark that prefix for caching. Keep volatile values (timestamps, session IDs) out of it, or you bust the cache on every call. Beyond caching, most of what reaches the model is redundant boilerplate; tools like Headroom, Token Company, RTK, and LeanCTX compress it before it is sent.7
FinOps at the workflow level. Know your cost per resolution, per pull request, per document. Most of the real spend hides in orchestration, retries, retrieval, and guardrails.
Context, not rate, is the lever
The lever that lowers the bill is the same one that makes the agent more accurate, and it is not the per-token price. It is the context. Send less, send it precisely, and you pay less and get better answers at the same time. Input accounts for about 76% of token consumption: you pay, overwhelmingly, to feed the model.8 Feeding it more degrades quality. Chroma measured what they call context rot across eighteen models, with performance growing less reliable as input grows.8 More context is often a more expensive, dumber agent.
Two engineers are reaching the same discipline from opposite ends. Tejas Chopra (Netflix) open-sourced Project Headroom after a $287 Claude bill, on the premise that up to 90% of what gets shipped to a model is redundant.7 Headroom losslessly compresses that material before it reaches the model; he reports roughly $700k saved across users since January. Mario Zechner built pi, a deliberately minimal coding agent against the maximal-harness instinct: exact control of what enters the context yields better output, and most harnesses make that impossible by injecting things behind your back.9
I read both of them the same way. Slow down. Build the tighter thing.
The architecture call
Hygiene slows the bleeding. Architecture decides whether you bleed at all. The decision follows from the variance: a cost distribution you cannot forecast has no business on a path a customer touches on every transaction.
Here is how I am thinking about it for my own work. Read it as one worked example, not a rule. The things a customer sees and depends on, the interfaces and the core trust-workloads, should run on deterministic runtimes, not stochastic agents. Selling software is selling trust, and trust means a bill and a user experience that do not surprise anyone.
The agentic work goes where being wrong or variable is cheap: devops, monitoring, build-time tooling, internal automation. A 2x swing or a tail run lands on an internal pipeline rather than on a customer or a per-transaction P&L line. Use the agent at build time and watch time. Ship the deterministic artifact at runtime.
The principle is the only part I would defend as general: match autonomy and stochasticity to blast radius and budget predictability. Where you need a defensible action and a predictable cost, deterministic. Where variance is cheap and the blast radius small, let the agent explore.
Each wave of technology makes one layer cheap and increases the value of the adjacent layer. AI is compressing the cost of producing language and software. It is not compressing the thought and judgment about where to spend the compute, what to make deterministic, and what to let wander. The people who win the next few years are not the ones chasing the lowest per-token rate. They are the ones who decided, deliberately, where the tokens go.
Watch the rate card if you like. Just do not mistake it for the bill, and do not let hype make your architecture decisions for you.
Sources & notes
1. Uber reportedly exhausting its 2026 AI budget in four months on Claude Code: Forbes. Microsoft's token-cost problem: Fortune.
2. Anthropic Opus 4.7 tokenizer emitting up to 35% more tokens at unchanged per-token rate: finout. Peak-hour limit depletion and the end of flat-rate pricing: age-of-product, Kingy AI.
3. Microsoft 365 commercial price increase effective July 1, 2026: AppDirect. OpenAI tool/container session repricing: OpenAI pricing.
4. Bai, Huang, Wang, Sun, Mihalcea, Brynjolfsson, Pentland, Pei. "How Do AI Agents Spend Your Money?" arXiv 2604.22750; summary via Stanford Digital Economy Lab. The ~1,000x figure is the abstract's round number; the body reports 1,200x vs chat and 3,500x vs reasoning. Same-task run-to-run variance is ~2x typical, heavy-tailed to ~30x on the hardest tasks.
5. OpenRouter mechanics, pass-through pricing, ~5.5% credit fee, fallback billing, and limits: OpenRouter docs, TokenMix, ofox. LiteLLM and ZenMux are comparable gateways worth a look.
6. Blended tiered vs frontier-routed cost across enterprise calls: Optimum Partners. Caching and batch discounts per provider documentation.
7. Project Headroom, by Netflix engineer Tejas Chopra (used by several Netflix teams, not an official Netflix project): The Register, GitHub, Chopra's writeup. ~90% redundant tokens, reversible compression, ~$700k saved across users since January. Adjacent tools: Token Company, RTK, LeanCTX.
8. Input as ~76% of token consumption: arXiv 2601.16746. "Context rot" across 18 models: Chroma. Models skipping the middle of long context: Liu et al., TACL 2024.
9. pi, by Mario Zechner: build notes, pi.dev, source. Context engineering paramount; minimal system prompt and toolset; deliberately no plan mode or sub-agents.
June 2026. Figures are directional and move fast; verify against primary sources before quoting. Written for operators who pay the bill, not for the rate card.