Where Token Prices Are Going: What to Budget for LLM Costs Through 2027

Two things are true at once, and holding both in your head is the whole skill of budgeting for this.

The price of a token has fallen further and faster than the price of almost anything in the history of computing. A unit of capability that cost a dollar in early 2024 costs somewhere between one and five cents today, depending on which benchmark you anchor to. And yet nearly every company we work with is spending more on AI this year than last, often several times more.

Both facts are consequences of the same thing, and if you budget from only one of them you will be wrong by a factor that shows up in a board meeting. Here is what is driving the price curve, where it plausibly lands by the end of 2027, and how to write a number down that survives contact with reality.

The curve so far

Measured properly, which means holding capability constant rather than tracking a named model, the cost of a fixed level of intelligence has been falling somewhere between 10x and 900x per year depending on the benchmark, with a median around 50x. That is not a typo and it is not a marketing figure. It comes from tracking the cheapest model that clears a given score threshold and watching that price over time.

The clearest recent illustration is at the top. In mid-2026, GPT-4-class capability, which was the frontier eighteen months earlier at $30 per million output tokens, is available from several providers at under $1. DeepSeek V4-Pro ships strong reasoning at $0.435 input and $0.87 output, a level that would have cost dozens of times more a year and a half ago, and Qwen3.8-Max launched in August claiming near-frontier capability at $2 and $6.

The other thing worth noting is that the decline is not smooth. It comes in steps, and the steps are triggered by open-weight releases far more often than by closed-lab price cuts. Kimi K3, DeepSeek V4 and the Qwen open-weight releases each forced repricing somewhere in the closed tier within a quarter. That is the mechanism, and it tells you where to look for the next step down.

Four forces still pushing prices down

Hardware improves and the improvement is passed on. Each GPU generation delivers more tokens per second per dollar, and inference is competitive enough that the gains reach the price list rather than the margin. This alone accounts for a steady annual decline before anything else happens.

Serving software keeps finding free throughput. Continuous batching, paged attention, speculative decoding and quantisation have collectively multiplied the number of requests a given card can serve, and none of that required a new model. We covered the mechanics in the continuous batching explainer, and the practical point is that this line of work is not close to finished.

Models are getting smaller for the same capability. Distillation, better data curation and architectural sparsity mean a 2026 model at 30 billion parameters does what a 2024 model needed 400 billion to do. Serving cost tracks active parameters, so this converts directly into price.

Open weights set a floor and drag everything toward it. This is the strongest of the four, and it works on the closed tier from below rather than by direct substitution. Anthropic's Opus line moved from $15 and $75 to $5 and $25 across recent releases. Qwen3.8-Max launched at $2 and $6 while claiming near-frontier capability. A closed provider can hold a premium for support, latency, indemnity and the last few points of quality, but each open release makes that premium harder to defend, and the labs reprice rather than lose the middle of the market.

Three forces pushing the other way

Reasoning tokens. Modern models spend hidden tokens thinking before they answer, and you pay for all of them at the output rate. A model that costs half as much per token but thinks three times as long is more expensive. We worked through the arithmetic in the reasoning token piece. The reason this matters for forecasting is that reasoning length has been growing faster than per-token prices have been falling on the hardest tasks, which means the price of an answer on those tasks has been roughly flat even as the price of a token collapsed.

Context windows. A million-token window invites million-token prompts. Teams that used to send 2,000 tokens of retrieved context now send 200,000 because they can, and the per-request cost rises even as the per-token cost falls. This is the single most common cause of a bill that grows while prices drop.

Agents. One user question used to be one model call. Now it is a plan, six tool calls, three retries and a verification pass. Agentic workloads routinely consume twenty to a hundred times the tokens of the chat interaction they replaced. That multiplier has swamped every price cut of the last two years for teams that adopted agents, and it will keep doing so.

Put the two lists together and you get the thing that confuses budget holders. Unit prices fall roughly 5x to 10x a year. Consumption per useful outcome rises roughly 10x to 30x a year for teams building agents. The bill goes up.

What to budget through 2027

Here is the forecast, stated as ranges because anyone giving you a point estimate on this is guessing with more confidence than the data supports.

Read the last row against the ones above it. Per-token prices fall by roughly a factor of three over the period. Cost per completed agentic task falls by maybe half, because tasks get more ambitious as they get cheaper. That gap is where budgets break.

The frontier tier is the least predictable line. There is no economic law saying the most capable model must get cheaper, and Anthropic has demonstrated repeatedly that it will hold price on Opus-class models across releases. What falls reliably is the price of yesterday's frontier, which becomes this year's mid-tier at a tenth the cost. If your work genuinely needs the top of the board, budget for flat pricing and plan to spend more as your usage grows. If it does not, budget for the decline.

What this means for how you build

Four consequences, in rough order of how much money they represent.

Do not architect around today's prices. A design that only works because a model costs $0.50 per million will be fine in 2027. A design that only works because a model costs $30 per million is a design that will be beaten by somebody who waited six months. Assume the capability you need will be a tenth of the price by the time you are at scale, and build for the volume that implies rather than the volume today's price allows.

Measure cost per outcome, not per token. The token price is the vendor's number. Cost per resolved ticket, per document processed, per completed agent run is yours, and it is the only one that connects to revenue. Teams tracking tokens routinely celebrate a price cut while their cost per outcome quietly doubles.

Keep the switching cost near zero. The whole value of a falling price curve is captured by whoever can move. If it takes a quarter to change models, you will bank a fraction of the decline; if it takes an afternoon, you bank most of it. That is an architectural property, not a procurement one, and it is the argument for a routing layer that most teams only make after being burned.

Spend the savings on specialisation, not on a bigger model. As general capability approaches free, it stops differentiating anything, because your competitors have the same model on the same day you do. What does not commoditise is a model shaped around data only you have, which is the case we made in fine-tuning on your own data.

The list price is not the price you pay

Anyone forecasting from published rates is forecasting the wrong number. Four discount mechanisms sit between the pricing page and the invoice, and all four are widening.

Batch tiers. Submit work asynchronously with a deferred deadline and most providers halve the rate. Anything without a user waiting on it, which includes nightly enrichment, backfills, evaluation runs and most document processing, qualifies. We went through the mechanics in batch inference cost optimisation. Teams routinely leave this on the table because the work started life as an interactive feature and nobody revisited it.

Prompt caching. Repeated prefixes are discounted steeply, often by 90%, provided you build prompts so the stable part comes first in a fixed order. Retrieval systems that concatenate chunks in relevance order defeat their own cache on every request. Sorting by a stable key instead costs nothing and can halve an input bill.

Committed spend. Above a threshold, every provider negotiates. The discounts are real, in the twenty to forty per cent range, and the trade is a commitment that locks you out of the falling curve for the term. On a market dropping several-fold a year, a three-year commitment at a 30% discount is usually a bad trade dressed as a good one.

Self-hosting above a volume line. Weights cost nothing and serving costs plenty, so this only pays above the point where a saturated GPU cluster undercuts the API. That line sits somewhere north of a billion tokens a month for most shapes of work, and it moves down as serving software improves. The full model is in the cloud versus on-premise analysis.

Applied together these regularly take 50% to 70% off a bill without changing a single model choice, which is a larger saving than most teams get from a year of the market falling around them.

A budgeting template that survives the year

Stop budgeting a single AI line and split it three ways, because the three parts behave completely differently.

Committed baseline. The steady production traffic you can forecast within twenty per cent. Price it at today's rates minus an expected 30% decline over twelve months, and revisit quarterly. This is the part of the budget that gets cheaper on its own.

Growth allocation. New features and expanding usage on existing ones. Price this at today's rates with no decline assumed, because new workloads land at whatever the market charges when they launch and they tend to launch on the more expensive tier before anyone optimises them.

Experimentation. Evaluation runs, prompt iteration, agent development, model comparisons. This is small in absolute terms and consistently underestimated, and starving it is how teams end up unable to answer whether a cheaper model would have worked. Ring-fence a fixed monthly figure and stop reconciling it.

Put a per-route budget alarm on all three rather than a monthly total, because a monthly total tells you about the overspend after it happened. The failure mode that produces a surprise invoice is almost never gradual growth. It is one agent loop that started retrying, or one context window that quietly grew from 4,000 tokens to 180,000 after somebody improved the retriever.

The scenario that would break this

Forecasts should say what would falsify them, so here is the one that would.

Compute scarcity. Every projection above assumes inference capacity keeps expanding faster than demand. If frontier training runs absorb a sufficiently large share of available accelerators, or if a supply shock hits advanced packaging or memory, inference prices could flatten or rise for a period regardless of algorithmic progress. There were brief episodes of exactly that in 2024 and 2025, when capacity constraints produced queueing and rate limits rather than headline price rises, which is how a shortage usually shows up first.

A second, smaller risk is regulatory fragmentation raising the cost of serving models across jurisdictions, which would show up as regional price divergence rather than a global rise.

Neither would reverse the long trend. Both could easily produce a year that looks nothing like the table above, and a budget with no slack in it would not survive either.

The practical summary

Token prices are going down and your AI bill is going up, and both will continue. Plan for unit costs to fall by roughly a factor of three by the end of 2027, for open weights to keep setting the floor, and for your own consumption to rise faster than prices fall if you are building anything agentic.

The teams that come out of this well are not the ones who picked the cheapest model. They are the ones who kept the ability to change their minds cheaply, measured cost against outcomes instead of tokens, and put the savings into something a price cut cannot hand their competitors. Swfte Connect exists for the first of those. The other two are yours.

---

Related: the efficiency race, AI API pricing trends, and the token cost calculator.