Money and markets · AI economics
Where the AI Money Goes
The price of a given level of AI capability falls more than tenfold a year, and company AI bills keep rising. The reason is simple once you see what a task actually consumes, and it changes how you should budget, buy and price.
In March 2023, the cheapest model that could match GPT-4 on a standard general-knowledge test cost $37.50 per million tokens. Less than two years later, a model clearing the same bar cost 18 cents. That is about 200 times cheaper.
In a later window, from January 2025 to April 2026, the businesses tracked by Ramp, a company that sells spend-management software, used about 11 times more tokens and spent about 6 times more money on AI.
Both facts are true, and together they say something most AI budgets still miss. The price of any fixed level of AI capability collapses. The bill rises anyway, because nobody keeps buying the fixed level. Companies buy the newest capability, and they use it for tasks that consume tens or hundreds of times more tokens than a chat answer. A bill is price times quantity, and quantity is winning.
Three consequences follow. Token price is the wrong number to budget on. The way AI is sold is moving from seats to meters and outcomes. And the money is drifting toward the parts of the work that don't get cheaper when the token does.
The price that falls
A language model reads and writes in tokens, chunks of text a bit shorter than a word, and providers charge per million of them. The price of a token at a fixed level of quality has fallen remarkably fast, and it is one of the best-measured facts in the field. Independent studies disagree on the exact rate, and agree on at least fivefold a year. The GPT-4-level series above works out to about 17 times a year. Cheaper chips, better algorithms, small models that match yesterday's large ones, and competition between sellers all push the same way.
Look closely at what that measurement holds constant, though. It follows the cheapest model that clears a bar set years ago. That is last year's frontier, sold at clearance prices. If a 2023-quality answer is what you need, you can now have it almost free.
Almost nobody wants a 2023-quality answer.
The bill that rises
Ramp's numbers show what buyers actually did. Tokens grew 11 times and dollars 6 times, so the price they paid per token did fall, but only by about 1.6 times a year, a sliver of the benchmark decline. Meanwhile premium models went from about 6 percent of these businesses' AI token spend to more than half.
The price a company pays follows the models it buys today, and current models get cheaper far more slowly than old ones. At the very top, the trend can run backwards. One academic estimate has the price of running the newest frontier models across a benchmark rising 3 to 18 times a year, because those models are larger and think longer. Anthropic's price list now carries a tier above its top Opus model at two and a half times the price.
Nobody buys last year's capability at this year's price. They buy this year's capability and give it bigger jobs.
In 1865, William Stanley Jevons argued that using coal more efficiently would make Britain burn more of it, because cheap power opens uses that weren't worth paying for before. The Jevons paradox is a good label for what is happening to AI spending, but a label explains nothing. The explanation is in what a modern task actually consumes.
Where the tokens go
A bill has four kinds of token, and most people picture only the first.
Input tokens are what you send: instructions, documents, the conversation so far. Output tokens are what the model writes, and they cost five times as much as input on current price lists. The reason is physical. The model reads your input in one parallel pass, but it writes one token at a time, and every single token needs its own trip through the whole network (Why Long Answers Cost More Than Long Questions). Cached input is input the provider already processed for an earlier request that started the same way, and it is sold at a deep discount. Reasoning tokens are output the model writes to think before it answers.
Guess first
On the same model, a chat answer costs about a cent. What does an agent that fixes one bug in a codebase cost: about 2 cents, about 10 cents or about 50 cents?
Now price three jobs on the same mid-priced model, Claude Sonnet 5.5, at $2 per million input tokens and $10 per million output. A chat answer costs about a cent. An agent that fixes one bug in a codebase costs about 54 cents, roughly 60 times as much. A long analysis of a stack of documents, where the model thinks before writing, costs about 67 cents.
Same price list, same model, a sixtyfold spread. Three mechanisms produce it.
Thinking is billed as output, and you never see it. In the analysis job, 40,000 of the 43,000 output tokens are the model reasoning with itself. They are about 60 percent of the bill and buy no words the reader reads. That can be money well spent, when the thinking is what makes the answer right (What Thinking Buys). It is still the biggest line, and it varies wildly: on the same set of evaluation questions, one deliberately concise model used about seven times fewer output tokens than the typical model. Comparing price per token across models without knowing tokens per task tells you very little.
Agents re-read everything, every step. An agent works in a loop. The harness, the software around the model, sends the model the task, it decides on an action, a tool runs, and the result is appended to the conversation. Then the whole conversation, from the first instruction to the latest tool output, goes back to the model for the next step. In the bug fix, the context never exceeds about 90,000 tokens, yet over 30 calls the model reads about 1.5 million. Because each step re-reads a longer history, the reading grows with the square of the number of steps. Double the steps and the bill rises about 2.4 times. Anthropic's engineers have reported that agents typically use about four times the tokens of a chat, and multi-agent systems about fifteen times. On one browsing benchmark, how many tokens a system spent explained 80 percent of the differences in how well it performed. Spending tokens is how this kind of capability gets delivered.
Caching is the reason agents are affordable at all. Because the start of each agent step repeats the previous step, the provider can keep that prefix processed and sell the re-read at a twentieth of the normal input price. Without prompt caching, the 54-cent bug fix would cost $3.26. Real usage looks like the example: Anthropic reports that Claude Code averages about $13 per developer per active day, and its own sample session reads nearly 800 cached tokens for every fresh one.
One more detail shows how little a price list fixes. Anthropic's newer models use a tokenizer that cuts the same text into about 30 percent more tokens. Two models with the same price per token can charge noticeably different amounts for the same document.
What it does to how AI is sold
When the cost of a task depends on how long the model thinks and how many loops it runs, customers stop looking alike. In Ramp's April 2026 data, the median business spent about $2,200 a month on AI tokens. The 99th percentile spent about $830,000, some 370 times more.
A flat seat price is a bet on the average customer. When one customer can cost a hundred times another, heavy users lose the vendor money and light users quietly pay for them. So vendors are moving to meters. On Anthropic's Team and Enterprise plans, each member's Claude Code use draws on a seat allowance, and use beyond it is billed as credits under limits the company sets. Others price the result instead. Intercom's support agent, Fin, charges $0.99 per outcome, and counts a resolution when the customer asks for no further help after its last answer. Under that kind of contract, how an outcome is defined is the real pricing decision.
Each model moves a different risk onto the buyer. Seats make budgets predictable and pay for idle licenses. Usage tracks value and produces bills nobody can forecast without ceilings. Outcomes push the model and loop risk onto the vendor and put all the weight on who counts the outcome. Most companies will end up holding all three at once, and their finance teams will need one unit that spans them.
Where the money flows
That unit is the cost of an accepted task: what the attempt costs in tokens, plus the cost of checking it, divided by the share of attempts that pass.
Take the 54-cent bug fix. Suppose a developer spends 8 minutes reviewing each attempt, at $60 an hour, and accepts 70 percent of them. Each accepted fix costs about $12.21, and the tokens are about 6 percent of that. Cut the review to 2 minutes and the accepted fix costs $3.63, with tokens at about 21 percent. Now remove the person. Run the same agent 1,000 times a day with no human checking each run, and the token bill is about $12,000 a month and the whole variable cost. Both situations can live in the same company.
Where a person checks the work, the token price is rarely the line that matters. Where nobody does, it is the only line.
This is where I think most companies get AI budgeting wrong. They negotiate hard on the number they can see, the price per million tokens, and leave alone the numbers that actually decide the cost: review minutes, acceptance rate, how many steps the agent takes, whether the cache is hit. The second-order effect matters more. As tokens get cheaper, the share of every AI project spent on the work around the model grows: connecting the model to company data and systems, getting permissions and security approved (Data Access Is the Real Integration Project), building evaluations, and paying for the people who check the output. None of that falls with the token. Saved time doesn't vanish either; it piles up at whatever step comes next, often review (Where the Saved Time Goes).
The public spending surveys mostly miss this shift. Menlo Ventures estimated $37 billion of enterprise generative AI spending in 2025, more than triple the year before, but its categories count purchases. Internal hours spent on integration, access reviews, evaluation and checking don't appear in them, and those are the lines I expect to grow fastest.
The strongest objection is that prices are falling so fast that none of this will matter. For a workload frozen in time, that's true. At ten to seventeen times cheaper a year, today's chat answer and today's bug fix will cost 1 percent of their current price or less within two years. But companies don't freeze workloads. They move to the better tier, add steps, add reasoning, and automate tasks they couldn't afford before. That is exactly what Ramp's numbers show.
Will volume keep outrunning price? I'd plan on it for the next year, with humility: Ramp's September 2026 index showed its effective price per token down 41 percent from a March peak, and its economist warned that volume may not grow enough to offset it. The signal to watch is spending per employee falling for two or more quarters while the model mix stays the same.
What to do
Budget per accepted task. Token price is an input. Cost per accepted task, including review and failed attempts, is the number that tells you whether an AI process pays (Cost per Task, Not Cost per Token).
Measure before you optimize. For each task, record input, cached input, output and reasoning tokens, plus calls per task, cache hit rate and acceptance rate. Without those, optimization is guesswork aimed at whatever line is most visible.
Pull the levers in the order the bill suggests. Caching first: keep the start of every request stable, and remember that even changing the thinking or effort setting between requests breaks the cache. Then model tier: Anthropic's smallest current model lists at a twentieth of the mid-tier price, so test your top use cases on it, and check the cost per accepted task, because a cheap model that needs more retries can end up costing more. Then thinking effort, lowered where the task doesn't need it. Batch anything that can wait, at half price. And give every agent run a step limit and a cost ceiling.
Forecast in ranges, and negotiate the drops. Ramp advises its customers to plan for AI spend growing 50 to 100 percent a year, with a 40 to 50 percent buffer for monthly swings. Treat that as a starting range until your own data replace it. Plan around the tier your teams will actually use, and write the right to the vendor's future price cuts into the contract.
Put money where the tokens can't reach. Integration, evaluation and review capacity are where many AI projects stall, and they are the budget lines that grow as tokens get cheaper. Fund them on purpose.
Start small and metered. One workload, a spending ceiling, a month of real data. That month will tell you more than any forecast, including this one.
The engine keeps getting cheaper, and that part will take care of itself. The bill is decided by what you ask the engine to do and by everything you build around it.
Check yourself
Why does the AI bill rise when the price of a fixed level of capability falls?
Nobody keeps buying the fixed level. Companies move to the newest models and use them for tasks that consume far more tokens, and a bill is price times quantity.
Why does an agent cost so much more than a chat answer?
It re-reads the whole conversation at every step, and its thinking is billed as output. Caching cuts the cost of the re-reading.
Why is cost per accepted task a better budgeting unit than token price?
Where a person checks the work, review minutes and the acceptance rate drive the cost, and those do not fall when tokens do.
Notes, sources and method
Prices are list prices read from vendor pages on October 11, 2026. Market data carry their own dates. Numbers in brackets refer to the source list at the end.
Key numbers
| Number used in the article | Value | Source and date |
|---|---|---|
| Cheapest model at GPT-4 level on MMLU (86 or higher) | $37.50 per million tokens (March 2023) to $0.18 (February 2025), about 208 times; 16.6 times per year at the endpoints | Epoch AI price tables, data updated November 20, 2025 [6] |
| Same, GPT-3 level (MMLU 43.9 or higher) | $60 (November 2021) to $0.07 (October 2024), 857 times; 10.5 times per year | Epoch AI [6] |
| Epoch's fitted declines across six benchmarks | 9x to 900x per year, median 50x; median 200x when pre-2024 models are dropped | Epoch AI, March 12, 2025 [6] |
| a16z "LLMflation" | about 10x per year, 2021 to 2024, MMLU | Andreessen Horowitz, November 12, 2024 [7] |
| Gundlach et al. | 5x to 10x per year at fixed frontier-level performance; about 3x per year from algorithmic efficiency; price of running the newest frontier models on benchmarks rising 3x to 18x per year | arXiv:2511.23455 v2, March 23, 2026 [8] |
| Ramp, businesses with connected AI | token usage +1,001 percent (11.01x), spend +497 percent (5.97x), January 2025 to April 2026 | Ramp, June 8, 2026 [10] |
| Ramp premium models | 5.7 percent of AI token cost in June 2025; 55.9 percent of cost and 45.8 percent of tokens in April 2026 | Ramp [10] |
| Ramp spend distribution, April 2026 | median $2,246 a month, mean $140,842, 99th percentile $831,338 (370x the median) | Ramp [10] |
| Ramp planning guidance | assume 50 to 100 percent annual growth in AI spend plus a 40 to 50 percent volatility buffer; test top three use cases on a lightweight model | Ramp [10] |
| Ramp AI Index | effective price $0.68 per million tokens, down 41 percent from a March 2026 peak of $1.15; slowing adoption growth; top-1-percent firms' spend per employee down 9.7 percent in August; volume "may not be large enough" to offset lower prices | Ramp, September 9, 2026 [11] |
| Google tokens processed per month | 9.7 trillion (2024), about 480 trillion (2025), over 3.2 quadrillion (May 2026), 330x in two years, across consumer and other surfaces | Google I/O, May 19, 2026 [9] |
| Menlo Ventures enterprise generative AI spend | $37 billion in 2025, $11.5 billion in 2024 (3.2x); $19 billion applications, $18 billion infrastructure ($12.5 billion model APIs, $1.5 billion data and orchestration) | Menlo Ventures, December 9, 2025 [12] |
| Agent token multipliers | agents about 4x chat tokens, multi-agent systems about 15x; token usage explains 80 percent of BrowseComp performance variance (95 percent with tool calls and model choice) | Anthropic Engineering, June 13, 2025 [13] |
| Claude Code costs | about $13 per developer per active day; below $30 for 90 percent of users; agent teams about 7x tokens in plan mode; sample session 940,000 cache-read tokens vs 1,200 fresh input | Anthropic documentation, read October 11, 2026 [5] |
| Output tokens to run one evaluation suite | 11 million for one concise June 2026 model vs a median of 82 million, about 7x | Artificial Analysis, read October 11, 2026 [14] |
| Tokenizer | models from Claude 4.7 onward produce approximately 30 percent more tokens for the same text; Opus 4.7 announcement gives roughly 1.0 to 1.35 times | Anthropic [1], [16] |
| Intercom Fin | $0.99 per outcome, 50-outcome monthly minimum; resolution counted when no further help is requested after the agent's last answer | fin.ai, read October 11, 2026 [15] |
| Jevons | efficient use of fuel tends to increase consumption | The Coal Question, 1865 [17] |
List prices used (USD per million tokens, October 11, 2026)
| Model | Input | Cache read | Output | Batch in / out |
|---|---|---|---|---|
| Claude Haiku 5.5 (prompts up to 100k tokens) | $0.10 | $0.01 | $0.50 | $0.05 / $0.25 |
| Claude Sonnet 5.5 | $2 | $0.10 | $10 | $1 / $5 |
| Claude Opus 5.5 | $4 | $0.20 | $20 | $2 / $10 |
| Claude Fable 5.1 | $10 | $0.25 | $50 | $5 / $25 |
| OpenAI gpt-6.1-sol | $2 | $0.10 | $10 | $1 / $5 |
| OpenAI gpt-6-astra | $10 | $1 | $50 | $5 / $25 |
OpenAI rows are short-context rates. Enterprise discounts, regional endpoints, long prompts and fast tiers change these prices. Other pricing details:
- Output costs five times input on every row above, largely because output is generated one token at a time, each needing its own pass through the network, while input is processed in parallel.
- Cache reads cost 5 percent of base input on Sonnet 5.5 and Opus 5.5, 10 percent on Haiku 5.5, 2.5 percent on Fable 5.1. Cache writes cost 1.25 times base for a five-minute lifetime or 2 times for an hour, so a five-minute cache pays for itself after one read. OpenAI lists cached input at 5 to 10 percent of input across its gpt-6 rows.
- Reasoning tokens are billed as output on both vendors. Anthropic bills the full thinking generated even when only a summary is shown; on newer models, thinking blocks from earlier turns that stay in context are billed again as input. OpenAI says reasoning can run from a few hundred to tens of thousands of tokens per request. Effort settings steer thinking and guarantee no particular count.
- Batch processing takes 50 percent off input and output and stacks with caching. OpenAI's Fast tier lists gpt-6.1-sol at twice the standard rate and Ultrafast at six times. Haiku 5.5 charges five times more per token above 100,000 tokens of prompt.
- Changing the thinking configuration or effort level between requests invalidates the cache (Anthropic documentation).
- Frontier list prices: Opus 4.1 at $15 and $75 (retired except on two cloud platforms) against Opus 5.5 at $4 and $20, 73 percent lower per token and about 65 percent lower per unit of text once the newer tokenizer is counted. Fable 5.1 is two and a half times Opus 5.5. Sonnet 5 launched at $2 and $10 with a scheduled rise to $3 and $15 on September 1, 2026, which Anthropic's page now says will not occur. Opus 4.6 and 4.7 share a list price of $5 and $25, so the tokenizer change made the same document about 30 percent more expensive on average, up to about 35 percent for some content.
- On Anthropic's Team and Enterprise plans, each member's Claude Code use draws from a seat allowance that resets on a rolling window; use beyond it is billed as usage credits with organization-set spend limits.
Method and assumptions
All costs are tokens times price per category, divided by 1,000,000. The worked bill uses Claude Sonnet 5.5 (input $2, five-minute cache write $2.50, cache read $0.10, output $10) and Claude Opus 5.5 (input $4, cache write $5, cache read $0.20, output $20). Token counts are illustrative assumptions, in each model's own tokens; real counts depend on tokenizer, harness and task. Replace them with your own.
Job 1, chat answer. 2,000 fresh input tokens (instructions, short history, question) and 500 output, no extended thinking. Sonnet 5.5: $0.004 + $0.005 = $0.009. Opus 5.5: $0.018.
Job 2, agent fixes a bug, 30 calls. Call 1 writes a 15,000-token prefix (instructions, tool definitions, task) to the cache. Each later call reads the whole context so far from the cache and writes 2,500 new tokens, of which 600 are model output and the rest tool results. Every call outputs 600 tokens, and all calls fall within the cache lifetime.
- Cache reads: sum over k = 2 to 30 of (15,000 + 2,500 x (k - 2)) = 1,450,000.
- Cache writes: 15,000 + 29 x 2,500 = 87,500.
- Output: 30 x 600 = 18,000.
- Sonnet 5.5: $0.145 + $0.21875 + $0.18 = $0.54375. Opus 5.5: $1.0875.
- Without caching, 1,537,500 base-price input tokens: $3.255 on Sonnet and $6.51 on Opus, 5.99 times the cached bill.
- At 60 calls: reads 5,162,500 (3.56x), writes 162,500, output 36,000, giving $1.2825 on Sonnet, 2.36 times the 30-call bill. Reads grow with the square of the number of calls because each call re-reads a context that grows by a roughly constant amount; writes and output grow linearly.
- Harnesses push back with compaction (summarizing old context) and by clearing old tool results. Both help and both cost something: compacting a large context is itself a large request, and rewriting the start of the context invalidates the cache.
- Check against Anthropic's documented Claude Code session (1,200 input, 50,000 cache write, 940,000 cache read, 5,300 output on Sonnet 4.6 at $3, $3.75, $0.30, $15): $0.5526, matching the documented $0.55. Cache reads outnumber fresh input about 783 to 1.
Job 3, reasoning-heavy analysis. One call with 120,000 input tokens and 43,000 output (40,000 thinking, 3,000 answer). Sonnet 5.5: $0.24 + $0.43 = $0.67. Thinking costs $0.40, 59.7 percent. Batch halves the total to $0.335. Opus 5.5: $1.34.
Cost per accepted task. Review on every attempt, 70 percent acceptance, Sonnet agent run $0.544. Eight minutes of review at $60 an hour is $8.00: ($0.544 + $8.00) / 0.7 = $12.21, with tokens 6.4 percent. Two minutes: $3.63, tokens 21.4 percent. Review time, hourly cost and acceptance rate are illustrative. Volume without per-run review: 1,000 runs x 22 working days x $0.54375 = $11,962 a month. With review on every run at 8 minutes, the same volume would carry about $176,000 of review, so the reversal to tokens-as-the-whole-variable-cost holds only where no person checks each run.
Frozen-workload decline. At 10 to 17 times cheaper per year, two years brings a fixed workload to between 1/100 and 1/289 of its price, so 1 percent or less.
Price-decline rates. Epoch endpoints: 208x over 1.90 years = 16.6x per year (GPT-4 level); 857x over 2.87 years = 10.5x per year (GPT-3 level). These differ from Epoch's regression slopes.
Ramp ratios. Implied price per token 5.97 / 11.01 = 0.542 of the starting level; annualized over 15 months, 0.542^(12/15) = 0.613, about 1.6 times cheaper per year. Premium versus other tokens in April 2026: (55.9 / 45.8) / (44.1 / 54.2) = 1.50. If that ratio held, moving from 5 to 45.8 percent premium tokens raises the average price by about 1.20.
Other ratios. Google 3,200 / 9.7 = 330x. Menlo 37 / 11.5 = 3.2x. Ramp 99th percentile to median 831,338 / 2,246 = 370. Opus 4.1 to Opus 5.5: 4 / 15 = 0.267 per token, times 1.30 for the tokenizer = 0.347 per unit of text.
The Python that recomputes every figure above is in notes/A24.notes.md.
Caveats and open questions
Why the decline estimates differ. Epoch, a16z and Gundlach et al. use different thresholds, time windows and model sets, and agree on direction and order of magnitude: at least five times cheaper per year at fixed capability. Epoch blends input and output prices three to one and excludes reasoning models, because their many extra tokens make price per token misleading. A benchmark threshold says little about whether a specific task is solved, and MMLU is old and open to contamination, as the a16z author notes. The series measure the price of capability already reached, which is last year's frontier. Gundlach et al. were read at abstract level only.
Ramp's data. The figures come from users of Ramp's Token Spend Management product, a self-selected group of active AI buyers, and Ramp sells spend-management tools. Ramp does not say whether the token and spend growth figures cover the same, fixed set of businesses; the implied price assumes they do. Ramp does not define "premium". The premium mix shift explains at most about a fifth of the gap between benchmark declines and the price buyers paid (Method). Most of the gap comes from what is measured: the benchmark series follow the cheapest model matching an old level, while businesses buy current models. A changing cohort, with heavier users joining, could produce the same pattern, and the data cannot separate these. Ramp's June post ($0.72 average effective price in April 2026) and its September index ($1.15 March peak) are different weightings and are not combined.
Will volume keep outrunning price? Spend rises only when quantity grows faster than price falls, and nothing guarantees that. Menlo's prediction that net spend keeps rising is a forecast. Ramp's September index is one analyst's reading of one dataset over a few months; the spend-per-employee series is volatile and has a seasonal dip. A counter-force exists: Anthropic reported that moving from Sonnet 3.7 to Sonnet 4 gained more performance than doubling Sonnet 3.7's token budget, so better models can need fewer tokens for a result. No public data settle the balance. The observation that would change the article's conclusion is dollars per employee falling for two or more quarters at a constant model mix, in a dataset that observes real spend.
Google's volume includes consumer products and says nothing about dollars.
Menlo's scope. Menlo restated earlier years to its current definition, which excludes chips, inference and model serving, and AI features inside existing software. Its categories are purchases; none is labeled internal hours spent on integration, access reviews, evaluation or checking. That statement rests on the category list as read, not a full read of the report.
Price, cost and value. A buyer sees only the price. The provider's cost to serve (hardware, power, utilization, model size) is not public, and value to the customer is a third number. A price cut can reflect cheaper serving, competition or a decision to buy market share, and from outside these cannot be separated. List prices have moved in both directions; a plan that assumes steady decline for the tier you use is a bet.
Artificial Analysis figures are for one model page and an index version that changes; an earlier snapshot gave different counts. Treat as volatile.
Scenarios for where budgets move (conceptual, October 2026, not forecasts). A, volume wins: use per worker keeps growing faster than price falls; token spend grows, seats give way to metered use, and metering and cost control become a budget line; signal: dollars per employee rising at a constant model mix. B, price wins: workloads plateau and prices keep falling; token spend shrinks as a share and money moves to integration and evaluation; signal: dollars per employee falling at a constant model mix. C, review is the limit: agents produce more than teams can check; token spend flattens while review capacity, evaluation sets and verification tooling grow; signal: agent output queuing for review. Integration and evaluation gain share in all three. Real budgets will mix them. These scenarios rest on reasoning: there is no independent public data on how large integration, review and evaluation costs are relative to token spend.
Illustrative counts. The worked bill's token counts are assumptions. Real deployments vary with tokenizer, harness, language and task; count tokens on the model you will buy before comparing offers.
Sources
- Anthropic. Pricing. Claude Platform documentation, read October 11, 2026. https://platform.claude.com/docs/en/about-claude/pricing — model prices, cache multipliers, batch discount, Haiku 5.5 long-prompt rate, tokenizer note (approximately 30 percent more tokens), Sonnet 5 pricing footnote, Opus 4.1 retirement note.
- OpenAI. Pricing. OpenAI developer documentation, read October 11, 2026. https://developers.openai.com/api/docs/pricing — gpt-6 family prices, cached input, batch, Fast and Ultrafast tiers.
- Anthropic. Steering thinking. Claude Platform documentation, read October 11, 2026. https://platform.claude.com/docs/en/build-with-claude/thinking-steering-and-cost — thinking billed as output, billing independent of display, prior thinking blocks as input, effort levels, cache invalidation.
- OpenAI. Reasoning models. OpenAI developer documentation, read October 11, 2026. https://developers.openai.com/api/docs/guides/reasoning — reasoning tokens billed as output, usage field.
- Anthropic. Manage costs effectively (Claude Code). Documentation, read October 11, 2026. https://code.claude.com/docs/en/costs — $13 per developer per active day, below $30 for 90 percent of users, example session, agent teams about 7 times tokens, seat allowance plus usage credits.
- Cottier, B., Snodin, B., Owen, D., Adamczewski, T. LLM inference prices have fallen rapidly but unequally across tasks. Epoch AI, March 12, 2025 (data updated November 20, 2025). https://epoch.ai/data-insights/llm-inference-price-trends
- Appenzeller, G. Welcome to LLMflation: LLM inference cost is going down fast. Andreessen Horowitz, November 12, 2024. https://www.a16z.com/llmflation-llm-inference-cost/
- Gundlach, H., Lynch, J., Mertens, M., Thompson, N. The Price of Progress: Price Performance and the Future of AI. arXiv:2511.23455, v2 March 23, 2026. https://arxiv.org/abs/2511.23455
- Pichai, S. I/O 2026: Welcome to the agentic Gemini era. Google, May 19, 2026. https://blog.google/innovation-and-ai/sundar-pichai-io-2026/
- Hajjar, R. How much do AI tokens cost businesses? Ramp, June 8, 2026. https://ramp.com/blog/ai-token-cost-for-businesses
- Kharazian, A. Cracks in the AI Thesis, Part 2 (Ramp AI Index, September 2026). Ramp, September 9, 2026. https://ramp.com/data/ai-index-sept-2026
- Tully, T., Redfern, J., Das, D., Xiao, D. 2025: The State of Generative AI in the Enterprise. Menlo Ventures, December 9, 2025. https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/
- Anthropic. How we built our multi-agent research system. Anthropic Engineering, June 13, 2025. https://www.anthropic.com/engineering/multi-agent-research-system
- Artificial Analysis. GPT-5.5 Instant (June 2026) model page, read October 11, 2026. https://artificialanalysis.ai/models/gpt-5-5-instant-06-26
- Intercom. Fin pricing, read October 11, 2026. https://fin.ai/pricing
- Anthropic. Introducing Claude Opus 4.7. April 16, 2026. https://www.anthropic.com/news/claude-opus-4-7
- Jevons, W. S. The Coal Question, chapter VII, "Of the Economy of Fuel" (1865). Library of Economics and Liberty. https://www.econlib.org/library/YPDBooks/Jevons/jvnCQ7.html