Here is a number that gets quoted in every AI strategy deck: token prices are down about 85% since March 2023.

It is true. It is also close to useless for budgeting an AI system, and it is the reason a lot of companies are about to have a very bad quarter.

My post on Anthropic’s S-1 landed earlier today and covered the document itself — the take-or-pay compute commitments, the blank boxes, the 93 pages of risk factors. This one is about the number that document quietly confirms: how the money actually gets charged, and why your bill goes up while the price list insists it should be falling.

The deflation story is real. It is also about the wrong tier.

Two indices, both current, both quoted constantly.

Silicon Data’s LLM Token Expenditure Index fell to $0.97 per million tokens on 1 September 2026 — a record low since it launched, and less than half the peak it set earlier that summer. BenchLM’s Token Price Index stood at 15 as of 26 October, against a March 2023 base of 100: frontier prices 85% below where they started, at a median blended $5.63 per million across 37 flagship models. The median across all 207 models BenchLM tracks is $0.75 per million input, $3.00 per million output.

Deflation. Real, large, and accelerating at the bottom of the market.

Here is the catch. Silicon Data’s index is usage-weighted — it blends provider list prices with actual consumption volumes. So it describes what the market pays on average, which means it is dominated by the cheap and mid-tier models doing the bulk of the boring work. It tells you the price of classification, extraction, summarising a support ticket. It does not tell you what frontier capability costs, and it certainly does not tell you what your agent costs.

And your agent is explicitly routed away from the cheap tier for anything hard. That is the whole design. You use a small model for the 90% of calls that are easy, because routing to right-sized models instead of defaulting everything to the largest frontier model available cuts cost per token by 30% to 60% with no new hardware investment.

So the deflation is real, and it lands on the part of your workload that was never the problem.

The tier your agent actually runs on went up 61%

Same BenchLM index, same methodology, twelve months to October 2026:

Tier Index Year over year Median blended $/1M Constituents
Frontier 15 +61.3% $5.63 37
Mid-tier 30.1 −44.5% $1.81 18

Frontier prices have gone up 61.3% in a year. Mid-tier prices have gone down 44.5%. Both are true. The industry quotes the three-year number, because 85% is the better headline. The twelve-month number is the one that describes the tier your hardest work runs on.

The direction reversed this year. The frontier index bottomed at 5.8 in January 2026, then climbed: 9.3 in March, 11.7 through April, May and June, 12.4 in July, 12 in August, 18.7 in September — up 16.9% in a single month — and 15 in October.

There is a reason, and it is not a mystery. The discount never applies to the frontier. Each generation ships with a capability premium — better agentic coding, longer-context reasoning, multi-step workflows — and gets priced for it. The old model gets cheap. The new one arrives expensive. The trailing tier becomes a commodity; the frontier stays a product.

Concretely, here is what the price list looked like on 1 October 2026:

Model Input $/1M Output $/1M
Claude Fable 5.1 10.00 50.00
Claude Opus 5.5 4.00 20.00
Claude Sonnet 5.5 2.00 10.00
Claude Haiku 4.5 1.00 5.00
Cached reads (all tiers) 0.10 – 0.25 —

And OpenAI cut GPT-6 prices by half in September: Luna to $0.10 input and $0.50 output, Sol to $2 and $10.

Both ends of the market got cheaper. The middle is where your agent lives. Budget off the wrong end and you will be wrong in both directions.

The unit of work changed and nobody repriced the contract

This is the actual mechanism, and it has nothing to do with price lists.

You do not buy tokens. You buy tasks. And tasks are executed by agents.

  • Gartner’s 2026 figure: agentic workflows consume 5 to 30 times more tokens per task than a standard chat interaction.
  • Anthropic’s own production data puts the multiplier at roughly 4x chat for a single agent and 15x for multi-agent systems.
  • One user request can expand into 20 to 80 separately metered model calls — planning, retrieval, tool calls, validation, summarisation, and any of those steps triggering sub-steps of its own.
  • Published token footprints run from roughly 200 to 2,000 tokens for a comparable single-turn chat interaction, up to 15,000 at the low end and over a million for complex multi-step workflows.

And the shape of that spend is worse than the multiplier suggests. In agentic workloads, the final output accounts for only 5 to 15% of the tokens consumed. Everything else is the system re-reading what it already knows.

Resent context is the invoice

A 2026 study of a multi-agent software engineering system, run across 30 real development tasks, found that input tokens made up 53.9% of total consumption — roughly double the output share — and that the single largest driver was the iterative code-review phase, which alone accounted for 59.4% of tokens on average. Agents repeatedly passed the full code context back and forth to refine it rather than generating new output. The input-to-output ratio in that work was about two to one. It is a communication tax.

A separate 2026 audit of 30 production engineering teams found that resent context alone accounted for 62% of the agent inference bill.

And from the Stanford Digital Economy Lab: agentic coding tasks burned up to 1,000 times more tokens than a simple code-chat exchange, and the cost of the same task varied by as much as 30 times from one run to the next.

So most of your bill is the model being told things it already knows, and the spread between a good run and a bad run is wider than most quarterly budgets.

McKinsey’s July 2026 analysis put it bluntly: per-token pricing has “stopped being a useful measure” of what enterprises actually pay for generative AI.

The pricing model charges you more for failing worse

Look at the multipliers again, because they line up with failure almost too neatly.

  • Fast mode: 2x.
  • US-only inference: 1.1x.
  • Batch processing: 50% off.
  • Cached reads: $0.10–$0.25, against normal input pricing of $1–$10. A 10 to 50x gap.

Now think about what your agent does when it gets stuck. A tool call fails, so it retries, often with additional context. Retry loops can consume many multiples of the tokens a single linear pass requires. The context grows. And the two levers you reach for under time pressure — switch to fast mode, send the whole transcript again — both cost multiples.

Fast mode at 2x during a thrash is a tax on failure. Re-sending context you could have cached at fifteen cents is a tax on amnesia.

The structural version of the same point, from IDC: output tokens run 3 to 5 times more expensive than input tokens, and a multi-agent pipeline that is not monitoring its input-to-output ratio can understate its true inference cost by 3x or more.

Nobody is coming to help, because the incentives point the other way

Anthropic’s S-1 is explicit about the business model. Revenue comes from token-based usage fees and subscriptions. About 80% of revenue is consumption-based, and contracts allow customers to cut usage or walk away with limited or even no notice.

Your consumption is their revenue. The metric everyone monitors is cost per token, which is the one metric that always improves — on both sides of the transaction.

Let me be precise, because this gets overstated. Your rising bill is not a conspiracy. It is what happens when a vendor’s cost is amortised over volume. Anthropic spent $8.9 billion on compute and infrastructure in H1 2026 against $16.3 billion of revenue. Their unit cost falls as volume rises. Your bill going up is their unit cost going down. Same fact, two sides of the trade.

What I will say is this: there is no mechanism in consumption pricing that rewards a vendor for making cost per outcome fall. Their revenue rises when you do more, even if each individual outcome gets cheaper. Which means the deflation arrives as a discount on tokens, and the rising bill arrives as a consequence of your own agents working. Both are true. Only one of them is in your business case.

The inversion nobody budgets for

Every productivity investment you have ever made had one property: the machine that works better costs less per unit of output.

AI does not have that property. Consumption pricing means a better agent does more work, and more work is more tokens. Your AI bill scales with your success. The agent that handles forty tickets instead of twenty is not saving you money — it is charging you double for the privilege.

Which is why the finance conversation always goes badly. The business case assumes a fixed saving: a fraction of a salary per headcount, every month, forever. The actual shape is a line that goes up as adoption rises. Those two curves cross, and the crossing point is where somebody in a spreadsheet has to explain why the productivity tool now costs more than the salaries it replaced.

AI did not give you a cheaper worker. It gave you a worker whose wages are metered per step.

What to actually do

Six things, in the order they move the number.

1. Measure cost per completed outcome, not per call. Include failed runs and the human escalation. An agent that burns forty model calls and then hands off to a person cost you forty model calls and delivered nothing. And do not budget the token line alone — the all-in operating cost of an agentic system (orchestration, vector stores, observability, evaluation, security controls, workflow monitoring, human review, enterprise integration, audit logging, plus the engineers maintaining all of it) runs two to five times the raw token cost.

2. Kill resent context before you do anything else. It is 62% of the bill in one audit and 59.4% in another. Cached reads at $0.10–$0.25 against $1–$10 normal input is a 10x to 50x lever on the input side, and it is entirely a design decision. Summarise between steps. Cache the system prompt. Do not re-send a transcript you could have summarised. Paying for context you already sent is not intelligence, it is amnesia with an invoice.

3. Route deliberately — which means you need evals. Routing cuts cost per token by 30% to 60% with no new hardware. But you can only route if you can classify the task, and classifying the task means knowing which calls were hard, which means an evaluation layer. That layer is the actual moat and almost nobody builds it. It is also why open models keep failing to displace the frontier despite being dramatically cheaper: Menlo Ventures data shows the open-source share of enterprise workloads actually fell. Enterprises buy the frontier anyway, because nobody built the router.

4. Cap the task, not the model. Put a ceiling on steps and on dollars per task, and fail loudly instead of looping. Every uncapped retry loop is an unbounded line item with a 2x fast-mode surcharge attached to it.

5. Price the frontier in the business case, not the median. Your hardest 10% of tasks determines whether the workflow works at all. If that runs on Fable or Opus class pricing — $10 and $50 per million — then $10 per million is your real number, not the $0.97 usage-weighted index that describes the easy 90%.

6. Ask for the metric nobody publishes. Cost per successful outcome, with the definition of success written down and agreed. If a vendor cannot commit to it, that tells you they do not control your step count — which means the cost of your AI system is set by your workflow’s shape and not by anyone’s price list.

The bottom line

IDC says it more bluntly than I can: the cost per agent action will fall about 95% from 2025 to 2030, while total token delivery costs rise 487-fold.

That is not a contradiction. It is what happens when the unit price collapses and the constraint that was keeping volume in check disappears with it. Cheap is the trend, not the guarantee.

The industry sold a deflation story and the bill went up anyway. The reason is not that the price lists are lying. It is that the price list prices a token, you are buying a task, a task is made of many tokens, and the number of tokens is set by your agent’s architecture — not by your procurement department.

The cheapest AI system in the world is not the one with the cheapest tokens. It is the one that does the most work in the fewest steps, sending the fewest things twice.

Nobody is selling you that, because it is your job to build it.

Sources