Token Prices Fell 280x but Your AI Bill Still Tripled

The short answer to why your bill rose while prices fell: your organization is buying far more tokens than it was a year ago, and the extra volume comes from agents, not from people typing more prompts. Your invoice is the product of two numbers, price per token and tokens consumed, and they are moving in opposite directions at different speeds.

Agentic AI is software that pursues a goal semi-autonomously by planning steps, calling tools, and acting on the results across many model calls, rather than answering a single prompt. A token is the unit models read and write, roughly three-quarters of a word, and the unit you are billed on. That autonomy is the whole value proposition, and it is also the whole cost problem: every plan, tool call, and retry is more tokens through the meter.

Decompose the bill and the mystery evaporates:

Two of the three forces push the bill up, and both live on the volume side of the equation. The one force pushing down, price, is real but is being outrun. This is why the honest framing of "LLM inference cost is falling but our spend is rising" is not a contradiction at all. It is arithmetic.

Per-token prices really did fall by roughly the amounts you have seen quoted, but only if you attribute the figures correctly and respect the time window. The headline numbers come from two different analyses that get blurred together, and getting them straight matters for forecasting.

The first is a16z's LLMflation analysis. LLMflation is a16z's name, coined in November 2024, for the rapid and sustained fall in the price of a given level of model performance. According to a16z (November 2024), the cost of an equivalent-performance LLM has been decreasing by about 10x every year, a factor of roughly 1,000 over the three years from late 2021 to late 2024. Their concrete example: at the MMLU-42 capability tier, price fell from $60 per million tokens for GPT-3 in late 2021 to $0.06 per million for Llama 3.2 3B served on Together.ai. At a higher capability tier (MMLU-83), a16z found the equivalent-performance price dropped roughly 62x since the GPT-4 launch in March 2023.

The second is the 280x in this post's title. According to Stanford HAI's 2025 AI Index, the inference cost for a system performing at the level of GPT-3.5 dropped more than 280-fold between November 2022 and October 2024. That is a tighter, capability-anchored claim than a loose "prices are down 280x," and the window matters: both the a16z and Stanford figures are historical analyses that end in late 2024, not live 2026 spot prices.

The strategic point is that this deflation is not stopping. Gartner predicted in March 2026 that by 2030, performing inference on a one-trillion-parameter LLM will cost generative-AI providers over 90% less than in 2025. So the price side of the paradox is durable: tokens keep getting cheaper. If you are still hunting for savings purely on the price line, the deeper playbook is in our guide to reducing LLM token costs through optimization, but price is the half of the equation that is already working in your favor.

Agents cost more because they consume more tokens per task, on the order of 4x for a single agent and 15x for a multi-agent system. According to Anthropic's engineering team (June 2025), agents typically use about 4x more tokens than chat interactions, and multi-agent systems use about 15x more tokens than chats. The same writeup found that token usage by itself explained 80% of the variance in agent performance, which tells you the extra tokens are not waste. They are where the capability comes from.

Where do those tokens go? Four places, and they compound:

Multi-agent systems add a fifth: parallel subagents, each maintaining its own context, is exactly how you get from a 4x multiplier to a 15x one.

So "how much do AI agents cost per task" has a clean first-order answer: multiply your per-chat cost by roughly 4 for a single agent and 15 for a multi-agent design, then adjust for your own model mix and cache hit rate. The multiplier is the reason a 280x price drop does not show up as a 280x cheaper invoice. You moved your workload onto the 4x to 15x side of the meter at the same time.

The metric that exposes the real cost is dollars per resolved task: total token spend over a period divided by the number of jobs the agent actually completed. This unit is more honest than cost per million tokens because it prices in the retries, failed runs, and multi-step overhead that per-token math hides.

Cost per resolved task is a unit-economics measure, meaning it ties spend to a unit of business value (a completed job) rather than to a unit of infrastructure (a token). The difference is decisive. A model that is 30% cheaper per token but fails 40% of runs, forcing re-runs or human cleanup, can cost more per finished task than a pricier model that succeeds on the first attempt. Per-token dashboards make the cheaper-per-token option look like the obvious win. Cost per resolved task tells the truth.

Building the metric is mostly instrumentation:

Across the production agent systems we audit at Particula Tech, teams that switch from cost per million tokens to cost per resolved task usually find their real unit cost is two to five times what their per-token dashboards implied, because the dashboards never counted the failed runs. Picture an invoice-processing agent that resolves 6,000 of 10,000 attempts in a month: the 4,000 unresolved attempts still burned tokens, so your true denominator is 6,000, not 10,000, and your cost per resolved task is meaningfully higher than the naive average suggests. For the full accounting discipline around this, our AI FinOps framework for token budgeting and chargeback lays out the ledger a CTO needs.

Per-token pricing hid the increase because the unit finance was watching (price per million tokens) was falling on exactly the line item that was exploding (total tokens consumed). Everyone was staring at the number going down while the number going up did the damage.

This is a classic measurement trap. Vendor pricing pages, procurement negotiations, and internal cost models are all denominated in dollars per million tokens, and that figure has been dropping for three years straight. A leader who benchmarks the vendor's price sheet quarter over quarter sees steady progress and reasonably concludes costs are under control. Meanwhile the workload silently migrated from single chat calls to 4x and 15x agentic workflows, and total token volume climbed a curve that no one had a chart for, because no dashboard tracked it.

The failure is not that anyone chose the wrong vendor. It is that per-token price is an input cost, and nobody was measuring the output: cost per unit of work delivered. This is the same gap that lets AI pilots pass a demo and then detonate a budget in production, which is why we argue for an ROI gate before agents leave the pilot. If the pilot's success metric is per-token price rather than cost per resolved task, the pilot is measuring the one number guaranteed to look good.

Four levers bend the curve, in this order of return: prefix caching, model routing, spend controls, and unit-economics tracking. Pull them in sequence, because each one changes the numbers you need to make the next decision.

Skip the exotic fixes until these four are in place. They cover most of the achievable savings at the least operational risk, and they are the same short list we install first on almost every engagement.

The bottom line: treat agentic AI as a shift from fixed per-seat software spend to variable, usage-based token spend, and manage it with the same rigor you apply to cloud FinOps. According to Gartner (July 1, 2026), $234 billion in enterprise application-software spend is at risk from agentic AI, as autonomous agents take over work that today flows through licensed applications. That budget does not vanish. It moves onto a meter, where volume, not a negotiated seat price, sets the bill.

Here is what to do with that, concretely and in priority order:

What to skip: do not chase a cheaper base model as your primary cost strategy. Price is the one variable already moving in your favor, and swapping models without fixing volume and resolution rate just reshuffles a bill that agents, not the model, inflated. For where this fits in a broader operating model, our AI for business pillar connects cost discipline to adoption and ROI decisions.

If your token bill is climbing while the price sheet keeps improving, the gap is almost always the metric you are watching. Particula Tech's AI FinOps audits trace spend to cost per resolved task and stand up the caching, routing, and spend-control levers that turn a runaway agentic bill back into a curve you can forecast. The 280x price collapse is real. Whether it reaches your invoice depends entirely on how you count.