Price Deflation & Cost per Task

The deflation is real, large, and not going to stop soon.

Start with the part that is not in dispute. GPT-4 launched in March 2023 at $30 per million input tokens and $60 per million output. In April 2026 Gemini 3.1 Flash listed at $0.10 and $0.40 — a factor of three hundred on input, for a model that beats the 2023 frontier on most things anyone deploys. Indices that track frontier list prices put August 2026 somewhere near an eighth of the March 2023 level. Whichever series you use, the direction has held for three years across every vendor.

The forecasts worth planning against are less dramatic than the history. The 10× annual declines of 2021–2025 came from a stack of one-time wins — better kernels, quantisation, mixture-of-experts routing, price wars funded by land-grab — and the credible projections put the next stretch at 3–5× per year through 2027, tapering toward 1.5–2×. Even the pessimistic end halves the price of a fixed capability roughly every eighteen months.

The unit that deflated is not the unit you consume.

Your finance team does not buy tokens. It buys resolved tickets, merged pull requests, reconciled invoices. Between the token and the completed task sits a multiplier that has been growing at roughly the rate the price has been falling.

• Agentic workflows burn 5–30× the tokens of a single chat completion for the same nominal job, on Gartner's March 2026 estimate — the loop re-sends the transcript every step, which is the quadratic arithmetic agent cost control works through.
• Reasoning models added a second multiplier. Thinking tokens are billed as output, and output is the expensive side. A task that took 800 tokens of answer in 2023 may take 8,000 tokens of deliberation plus an answer now, and it is often more accurate — you are buying a different, better product at the same line-item price.
• Coding agents are the extreme. A representative 50-turn session models out at roughly one million input tokens against forty thousand output, an input-to-output ratio near 25:1 with input carrying about 85% of the cost — and run-to-run variance on the same task class reaching 30×.

Put the two trends in the same sentence and the result is deflationary but not by the factor on the pricing page: a 10× cheaper token multiplied by a 10× hungrier workflow is flat. Real deployments land between those poles, which is why the honest planning statement is "our cost per successful task fell 40% last year" and never "tokens got 90% cheaper so our costs did."

Cost engineering depreciates; reliability engineering does not.

This is the operationally useful consequence, and it changes what your team should work on. Any optimisation whose payoff is denominated in tokens is an asset that loses value at the deflation rate. A prompt-compression project that saves 30% of spend today saves 30% of a number that will be a third of its size in two years. Meanwhile the engineering that raises the success rate keeps its value, because it moves the denominator of cost-per-successful-task and nothing about the market erodes it.

• The payback test. Fund a pure token optimisation only when its payback period is comfortably shorter than the halving time of the price it acts on — call it twelve to eighteen months at current rates. A six-week project that pays back in three months is fine. A two-quarter engineering programme to shave context is usually beaten by waiting.
• Structural savings survive; incremental ones do not. Prompt caching, routing the high-volume steps to a cheap model, and not calling the model at all are structural: they change the shape of the workload and keep paying as prices move. Hand-tuning a prompt down by 400 tokens is incremental and expires.
• Reliability compounds against deflation. Lifting a task's success rate from 70% to 85% cuts cost per successful task by about 18% and cuts the human-review and rework bill, which is the line that does not deflate at all because it is denominated in salaries.

The mirror image is worth stating plainly: deflation is not a reason to be careless. A runaway loop costs what it costs today, and denial of wallet does not wait for the next price cut. Caps and circuit breakers are reliability work, not cost work, and they belong in the first category.

A term commitment is a directional bet, and you should know which way you are betting.

Reserved capacity, provisioned throughput and annual spend commitments all convert a floating price into a fixed one. In a market where the floating price has fallen every year for three years, that is a short position against a trend — you are agreeing to pay tomorrow's volume at today's rate.

• Buy them for the things they actually deliver: capacity you can promise a customer, a latency floor under load, and data-handling terms. Those are real goods and worth paying for.
• Do not buy them as a saving, unless the discount clears both hurdles at once: the utilisation break-even that provisioned throughput and commitments works out, and the expected price decline over the term. A 30% discount on a two-year term loses to doing nothing if list prices halve in eighteen months.
• Price the model lock-in separately. The larger cost of a long commitment is usually not the money — it is that it freezes your model choice in a market that ships a materially better option every quarter, and the cheapest thing you can do for your bill is switch to a model that finishes in fewer steps.
• Prefer short terms and reopener clauses. A one-year commitment with a repricing trigger tied to public list prices costs almost nothing to ask for and removes most of the downside.

Write the price path down, because right now it is an implicit assumption.

Most agent business cases contain a price assumption nobody has ever stated: that unit cost stays at today's level for the life of the model. That is the one input we can say with confidence is wrong, and it is usually wrong in your favour — which means the business cases being rejected on cost grounds are being rejected against a pessimistic input nobody defended.

• Model three paths, not one. Flat (no further declines), central (halving every 18 months at a fixed capability), and fast (halving annually). If the project only clears the bar on the fast path, it is a bet on the market, and someone senior should agree to make it.
• Hold capability constant when you apply the path. Deflation applies to the model you have today becoming cheaper or being replaced by an equal-quality cheaper one. It does not apply to your intention to upgrade to whatever is best next year — that is a separate, roughly flat line.
• Carry the token-growth term explicitly. Your own history has it: measure tokens per successful task quarterly. In most deployments it has been rising, because agents get more steps, more tools and more context as they get better.
• Set a re-plan trigger, not a re-plan date. "We revisit when list price for our primary model moves more than 30%, or when tokens per task moves more than 30%." Both are observable from data you already collect for cost attribution.

The forecasting mechanics — task-class mixes, p95s, why the mean is biased low — are in forecasting agent spend. This step only adds the term that page holds constant: the price itself.

The dashboard that survives a price change.

Spend is the metric that will mislead you most over the next two years, because it moves for four unrelated reasons at once — volume, task mix, token growth and price — and a flat spend line can hide any of them going badly wrong. Track the ratio instead.

• Cost per successful task, by task class, monthly. The primary number. Success means passed your evaluation, not returned a response, which is the whole argument of unit economics.
• Tokens per successful task, by task class. The term you control. If it rises faster than prices fall, your engineering is losing to your own feature velocity.
• Blended price per million tokens, actually paid. Reveals how much of the market's deflation you are capturing. If it is flat while list prices fall, you are running everything on the frontier tier and have a routing opportunity worth more than any prompt edit.
• The human line, separately. Review, escalation and rework are salary-denominated and do not deflate; as tokens get cheaper this becomes the majority of the bill and eventually the only thing worth optimising.

Related: measuring ROI for the value side of the same ratio, small and local models for where the deflation is steepest, and scaling back a deployment for what to do when a task class does not clear the bar on any price path.