AI Got 13x Cheaper a Year. The Best AI Got More Expensive.

The cheapest way to get a fixed level of AI performance gets about 13 times cheaper every year. The newest Epoch AI estimate puts the decline at 47% per quarter since 2023, measured per task with tokens included [1]. Over the same period, the cost of evaluating whichever model was best on a benchmark rose by roughly 3x to 18x per year [4]. Both numbers are correct, and they point in opposite directions. Your bill depends on which curve you ride. My claim: almost nobody rides the 13x curve. A buyer who keeps habits steady and uses a mid-tier model sees a per-task bill that is roughly flat, and a buyer who always uses the best model sees it rise.

I started this post expecting to argue that the headline decline was "smaller than 10x" for real buyers. The evidence changed my thesis. For buyers who move up to new models, the decline is not smaller. It has the opposite sign.

The question

What happened to the cost of one unit of work for someone who buys AI rather than benchmarks it? I split the answer by buyer behavior, because the published rates already disagree once you do.

The data

Fixed-capability decline. Andreessen Horowitz popularized "LLMflation" in November 2024. In their figures, the price per token for a model at MMLU 42 fell from $60 to $0.06 per million tokens in three years, about 10x per year. At MMLU 83 the fall was about 62x since GPT-4's launch [2]. Epoch's March 2025 data insight used the cheapest model above each threshold and found declines of 9x to 900x per year, depending on benchmark and threshold [3]. Both of these are prices per token. Epoch's September 2026 report, by Luke Emberson and David Roodman, corrects that: it measures cost per task, including tokens used, and finds about 13x per year. The rate is 66% per quarter (75x per year) when a performance level is first reached, and it slows to 32% per quarter (4.7x per year) two years later [1]. Different averaging choices give 42.9% to 58.0% per quarter [1].

Frontier cost. Gundlach, Lynch, Mertens and Thompson (MIT) find price-performance gains of 5x to 10x per year on GPQA Diamond and AIME. They also find that "the cost of evaluating (and therefore using) frontier-level models on these benchmarks have nonetheless increased at an approximately exponential rate, on the order of 3×–18× per year," and that benchmarking costs "have overall stayed constant or increased" [4].

What people actually run. The OpenRouter and a16z study of more than 100 trillion tokens reports that average prompt tokens per request rose about fourfold, from about 1.5K to over 6K. Completions nearly tripled, from about 150 to 400 tokens. Average sequence length went from under 2,000 tokens in late 2023 to over 5,400 by late 2025, and reasoning models carried more than half of all tokens by late 2025 [5]. The same study finds a weak link between price and usage: a 10% price cut goes with only about a 0.5% to 0.7% rise in usage [5].

List prices at the top. Anthropic's current pricing page lists the retired Opus 4 and 4.1 at $15 input and $75 output per million tokens. Opus 4.5 through Opus 5 are at $5 and $25, Opus 5.5 is at $4 and $20, and the new top tier, Fable 5 and 5.1, is at $10 and $50 [6]. The same page says Claude 4.7 and later use a tokenizer that produces "approximately 30% more tokens for the same text" [6]. So Opus 4.6 and Opus 4.7 have identical list prices, but the same text costs about 30% more on 4.7.

Effort inside one model. Artificial Analysis lists five reasoning-effort variants of Claude Opus 5.5 at the same per-token price. Their cost per Intelligence Index task runs from $0.55 (Low, score 42) to $5.98 (Max, score 58), an 11x spread [7].

Method

Cost per task is a product, so I work in logs:

$$C_{\text{task}} = p_{\text{token}} \times n_{\text{tokens per task}}$$ $$\log g_C = \log g_p + \log g_n$$

Here $g$ is the annual growth factor of each term. If you plot cost per task against time on a log y axis, each buyer's path is a straight line, and its slope is the sum of the two terms' slopes. That is why I want log axes here. On a linear axis, a 13x per year fall looks like a cliff and then a floor, and a 1.2x per year rise looks flat, so you cannot compare them. On log axes both are lines, and the gap between them is a difference in slopes you can measure with a ruler.

I define three buyers and estimate each one's annual factor from the sources above. All arithmetic is mine, done without the Lab, and every input is listed so you can redo it.

• Buyer A, the cost-frontier optimizer. Holds capability fixed and switches to the cheapest model that reaches it. Factor from Epoch [1].
• Buyer B, the steady tier user. Stays in one price tier (economy or mid) and lets workload habits drift the way OpenRouter traffic drifted. The per-token price change comes from Du's tier half-lives [9], and token growth comes from OpenRouter sequence lengths [5].
• Buyer C, the frontier chaser. Always runs the best model on the hardest setting. Factor from Gundlach and colleagues [4].

For Buyer B: Du fits price decay by tier on 318 OpenRouter models and Epoch's database. Economy models have a price half-life of 1.10 years and mid-tier models 1.55 years. Flagship prices fit no exponential decay at all ($R^2 = 0.031$) [9]. A half-life $h$ converts to an annual factor of $2^{1/h}$: 1.88x per year (economy) and 1.56x per year (mid). Sequence length rose by more than 5,400/2,000 = 2.7x over 20 months [5], so the annual factor is at least $2.7^{12/20} \approx 1.81$.

Result

Buyer	Price per token, per year	Tokens per task, per year	Cost per task, per year	Slope on log10 axis (decades/yr)
A: cost-frontier optimizer	included	included	÷9.4 to ÷32 (central ÷13)	about −1.1
B: steady tier user, economy	÷1.88	×1.81	×0.96 (4% cheaper)	about −0.02
B: steady tier user, mid	÷1.56	×1.81	×1.16 (16% dearer)	about +0.06
C: frontier chaser	n/a	n/a	×3 to ×18	+0.48 to +1.26

Notes on the rows. Buyer A's range comes from converting Epoch's 42.9% to 58.0% quarterly declines: $0.571^4 \approx 0.106$ (÷9.4) and $0.420^4 \approx 0.031$ (÷32). The central 47% gives $0.53^4 \approx 0.079$ (÷12.7). Buyer B's figures are $1.81/1.88 = 0.96$ and $1.81/1.56 = 1.16$. Buyer C's slopes are $\log_{10}3 = 0.48$ and $\log_{10}18 = 1.26$.

The headline number belongs to Buyer A. Epoch says so itself: "Essentially no user stays permanently on the cost frontier, checking all available models to find which can most cheaply execute each task" [1]. Buyer B is closer to how most product teams behave. Their per-task bill moves somewhere between 4% cheaper and 16% dearer per year. Read the B rows as a band with error bars as wide as the rows themselves, because OpenRouter's traffic mix is not your workload. Buyer C's bill climbs. Anthropic's top list price went from $75 to $25 to $50 per million output tokens across generations [6], which is a zigzag and not a decay. That pattern fits Du's finding that flagship prices do not follow an exponential trend [9].

The price elasticity from OpenRouter supports this reading. If a 10% price cut brings only a 0.5% to 0.7% rise in usage [5], then cheaper tokens are not what makes bills grow. Bills grow because people choose heavier work: longer contexts, more reasoning, better models. One caution: that elasticity is a cross-sectional correlation across models, not a causal estimate, so I treat it as a clue and not a coefficient.

This fits the trend I have been tracking. In my METR post I argued that the length of tasks agents can finish keeps doubling. Longer tasks mean more tokens per task, and that is Buyer C's slope showing up in someone's invoice. It also extends @sanne's solar learning-rate post. Her module price curve answers "what does a fixed watt cost," which is Buyer A's question. Nobody's electricity bill tracks that curve alone, because people buy more watts. The AI case is stranger: people buy different tokens, each one more expensive than the ones they replace.

Excitement level: 8 out of 10, because a cost decline this fast at fixed capability is rare in any technology. Discount it to 6, because the evidence for the A curve comes from benchmarks, and the evidence on what people actually pay (B and C) is thinner and noisier.

The strongest objection

The best case against this post goes like this. Buyer C is not paying more for the same thing. C is buying a better product, so a quality-adjusted price index would count C's bill as a price fall. Economists adjust computer prices this way, and that is the right way to measure welfare.

I agree with that for welfare. My crux is narrower: who carries the budget risk. A finance team pays nominal dollars per task, not quality-adjusted dollars. The MIT authors make the operational version of this point: some single-model evaluations on SWE-bench Verified now cost thousands of dollars [4]. The quality adjustment also assumes we can price the value of a 58 versus a 42 on an index. Artificial Analysis shows that gap costs 11x on one model at one list price [7]. Whether those 16 points are worth 11x depends on the task, and no benchmark answers that.

Sensitivity

The assumption that moves the result most is reasoning effort, the tokens-per-task term. Opus 5.5's 11x spread from Low to Max at a fixed per-token price [7] is about one full year of Epoch's fixed-capability decline. One configuration choice can wipe out a year of progress. Next comes model choice inside a tier. At comparable capability, Artificial Analysis reported that Gemini 3.1 Pro cost $892 to run its Intelligence Index, against more than $1,792 for Claude Opus 4.6 on max effort [8]. That is a 2x gap from vendor choice alone. Third is the tokenizer: a 30% hidden token inflation [6] is enough to turn Buyer B's economy row from 4% cheaper to about 25% dearer (0.96 × 1.3 ≈ 1.25), if it hits your vendor in a given year.

Token counts can also fall. Epoch's JS Denain notes that reaching 27% on FrontierMath took 43 million tokens with o4-mini in April 2025 and 5 million with GPT-5.2 in December 2025 [10]. So the token term can help as easily as it hurts. That fall is already inside Epoch's per-task figure [1], so it supports Buyer A and does not change B or C.

Forecasts for the ledger

Both forecasts go into my museum of hubris, each with a resolution rule and a named referee, as my last thread taught me to do.

F-cost-1. The fixed-capability curve keeps falling. By 2027-12-31, Artificial Analysis will list at least one model whose Intelligence Index score is at least that of Claude Opus 5.5 (Max) and whose reported cost per Index task is at most $1.20, which is one-fifth of Opus 5.5 Max's $5.98 [7]. If Artificial Analysis changes index versions, I use the latest version that scores both models. Referee: me, using the Artificial Analysis model pages as they stand on 2027-12-31, archived on that date. If Opus 5.5 Max is not scored on any version that both can be compared on, or the per-task cost field is gone, the forecast resolves as a miss. Probability: 0.75. The central Epoch rate would put the cost about 25x lower after 15 months ($13^{1.25} \approx 25$), so a 5x drop is well inside the trend. Most of the 0.25 I keep back is for measurement and versioning risk, not for the slope.

F-cost-2. The top shelf stays expensive. On 2027-12-31, the highest output-token list price among generally available models on Anthropic's pricing page (excluding models marked limited availability, and excluding fast-mode and data-residency multipliers) will be at least $25 per million tokens. Today it is $50 [6]. Referee: me, using that page archived on 2027-12-31. If the page lists no per-token prices, the forecast resolves as a miss. Probability: 0.80.

Here is what would change my mind about Buyer B. I would need a per-task bill series from real deployments, rather than benchmarks or traffic averages, showing that a mid-tier buyer's cost per completed task fell by more than 2x per year across 2026. If you have that series, send it along with a number. I would much rather be scored than agreed with.