Someone running five agents in production shared their case on Reddit (exhibit 98). An API bill of $4,000 arrived for one month. In six weeks they cut spending by 70%.
When they measured cost per workflow, an endpoint showed up, the address of a service the agent calls. It returned malformed JSON and multiplied usage by 5. The monthly bill did not show it.
The thesis about agent costs usually comes in two sentences. First: cost is not a worry, because model prices fall every year. Second: if the agent costs less per case than a person, it pays off. Both use real data, but they look at the price of a part instead of the cost of the result.
The price of a given level of performance is falling, and fast. A benchmark is a standard test to compare models. Epoch AI tracked the cheapest model that reaches a given score on five benchmarks (exhibit 133). Since 2023, the cost of fixed performance has fallen about 47% per quarter, roughly 13 times per year.
-47% per quarter (13 times per year)
Drop in the cost of a fixed level of performance since 2023 (with other reasonable averages, 42.9% to 58.0% per quarter). Exhibit 133
GPQA Diamond is a science question benchmark. Getting 75% on it cost about 30 cents per question with o3 in January 2025. Today it costs $0.0004 with GPT-5.6 Luna, 725 times less. The authors ask readers not to take those figures as exact.
Agent costs are not a worry: model prices fall every year.
Verdict: The price of a fixed level of performance is falling fast, about 13 times per year. Its authors ask that the figure not be read as exact. But running the frontier model gets 3 to 18 times more expensive every year. And prices do not always go down: Google doubles the price of Gemini 3.8 Flash from January 1, 2027.
Rule
Budget with today's price and with a scenario where it doubles. Check list prices every quarter and write down the date you looked.
Why medium confidence. The prices come from the official pages, but they go stale fast. OpenAI's page shows that the mid tier has barely dropped since 2024. Both trend analyses measure prices on benchmarks, not an agent's cost per task. The MIT one was read as an abstract. The subsidy argument comes from a Reddit thread and from a Gartner prediction known from the press.
Show the 7 exhibits in detailHide the detail1 for, 4 against, 2 qualify
This is an update from the same team that measured price declines in 2025. They track the cheapest model that reaches a given score on five benchmarks (standard tests), and simulate how many tokens each one uses. The cost of a fixed level of performance falls very fast. They note limits: benchmarks are not useful work, real users do not always switch to the cheapest model, and the data covers few years.
Numbers
-47% per quarter (13 times per year)Drop in the cost of a fixed level of performance since 2023 (with other reasonable averages, 42.9% to 58.0% per quarter).
75 times per yearDrop in the cost of performance that has just reached the frontier; two years later it falls 4.7 times per year.
725 times lessCost per question for 75% on GPQA Diamond: about 30 cents with o3 (January 2025) versus 0.0004 dollars with GPT-5.6 Luna.
Limitation It does not measure the cost per agent task or whether the cost of using the frontier model rises. The authors ask not to read the figures as exact. Clarification added on September 23, 2026.
Gundlach, H., Lynch, J., Mertens, M., Thompson, N. arXiv (MIT FutureTech)
The price to reach a given score falls. But running the frontier model gets more expensive every year, because models are bigger and reason more. The authors ask evaluators to publish how much it costs to run the benchmarks (standard tests). Price per token (a unit of text) is not cost per task.
Numbers
5 to 10 times per yearDrop in the price of a given score in frontier models.
3 to 18 times per yearRise in the price of running frontier models.
Limitation The abstract was read. There is a version 2 (March 23, 2026) and it was not checked whether it changes the results.
“Benchmarks may therefore present a warped picture of progress in practical capabilities”
These are the Gemini list prices, checked on November 1, 2026 and reviewed on November 2. They show that prices also go up: Google doubles the price of Gemini 3.8 Flash starting January 1, 2027. Output includes reasoning tokens, which the model generates to think before it answers.
Numbers
USD 0.75 and USD 3.75Gemini 3.8 Flash until December 31, 2026: input and output per million tokens.
USD 1.50 and USD 7.50Gemini 3.8 Flash from January 1, 2027.
USD 2 and USD 12Gemini 3.1 Pro Preview, with prompts up to 200,000 tokens.
Limitation Prices expire fast, and this one already has a change date: check it before using these figures. The source research says Gemini 3.6 Flash also goes up; that was not rechecked.
Gartner predicts that in customer service, the cost per case resolved with generative AI will exceed that of many outsourced human agents. One reason is that consumption is higher than expected and today's prices are subsidized. Also, new models spend more tokens per interaction, and infrastructure costs money.
Numbers
over 3 dollarsCost per resolution with generative AI that Gartner predicts for 2030.
3, 5 or up to 10 timesTokens frontier models use for similar interactions, according to the article.
Limitation A press article about a Gartner release (January 26, 2026) that returned 403. It is a prediction, not measured data.
“many organizations consume a lot more than they expect”
The author spent about USD 50 of API usage on a small weekend app. At work, a USD 20 subscription covers far more use. He concludes that today's price is subsidized, and that many agent uses stop adding up with billing for real usage. Others compare it with cheap Uber rides in 2015. In contrast, one commenter spends USD 2,000 to 5,000 a month, and his company accepts it because it multiplies his output.
Numbers
about USD 50 vs 20API spend on a small app vs a monthly subscription, according to the author.
USD 2,000 to 5,000 a monthSpend by one commenter, which his company accepts.
991 points and 457 commentsSize of the thread according to the archive. The quoted comment has 74 points.
Limitation Small, self-reported figures; the thread is an opinion about where prices are going. The quote is from the favorable counterpoint, not from the thread's thesis. It was read in the public Arctic Shift archive because reddit.com blocks automated reading. The quote was compared letter by letter with the original.
“It's been a force multiplier so my company has no problem with that additional cost.”
These are the list prices of Anthropic models, checked on November 1, 2026 and reviewed on November 2. Between the most expensive and the cheapest model of the same generation there is a 40 times gap. The page also documents costs that the table does not show. The new tokenizer splits the same text into more tokens, which are the units you pay for. US-only inference costs more, and declaring browser or computer-use tools adds thousands of input tokens per request.
Numbers
USD 4 and USD 20Claude Opus 5.5: input and output per million tokens (cache read: USD 0.20).
USD 2 and USD 10Claude Sonnet 5.5: input and output per million tokens (cache read: USD 0.10).
USD 0.10 and USD 0.50Claude Haiku 5.5, with prompts up to 100,000 tokens: input and output per million.
USD 15 and USD 75Claude Opus 4 and 4.1, the 2025 top tier: input and output per million.
about 30% more tokensEffect of the new tokenizer for the same text.
50% lessBatch API discount; it does not apply to interactive agent sessions.
1.1 timesPrice multiplier for requesting US-only inference.
Limitation Prices expire fast: check them before using these figures. Haiku 5.5 costs more with prompts over 100,000 tokens. Price per token is not cost per task: it depends on how many turns the agent takes and how much context it resends.
These are the list prices of OpenAI models, checked on November 1, 2026 and reviewed on November 2. Between the most expensive and the cheapest model of the gpt-6 family there is a 100 times gap. The mid-tier price barely changed since 2024. What improved is capability per dollar.
Numbers
USD 10 and USD 50gpt-6-astra: input and output per million tokens (cached input: USD 1).
USD 2 and USD 10gpt-6.1-sol: input and output per million tokens.
USD 0.10 and USD 0.50gpt-6-luna: input and output per million tokens.
USD 2.50 and USD 10gpt-4o, as a 2024 reference: input and output per million.
10% moreSurcharge for regional processing on models launched since March 5, 2026.
Limitation Prices expire fast: check them before using these figures. The page shows no update date. Reasoning models bill hidden tokens as output, so price per token is not enough to estimate cost per task.
Read at the source. Checked on Nov 2, 2026.
Against a person, the numbers are strong too. A SaaS support team described its case on Reddit (exhibit 99). Each ticket costs $0.18 with the agent, against about $3.50 with people.
Klarna’s CEO spoke at its third-quarter 2025 earnings presentation (exhibit 21). He said its agent does the work of more than 853 people and saved $60 million. These are the company’s figures, unaudited.
Sometimes spending more pays. Anthropic compared a single agent with a three-agent harness (exhibit 70). A harness is the code that organizes the model’s work. In this one, one agent plans, one builds and one evaluates.
For a retro game, the single agent took 20 minutes and cost $9, but the game did not respond to the controls. The harness took 6 hours and cost $200, and the game was playable.
An experienced developer said on Reddit that they spend $2,000 to $5,000 a month on agents (exhibit 107). Their company accepts it because it multiplies their output.
Fixed performance gets cheaper, but the frontier model does not. According to MIT FutureTech, running frontier models gets 3 to 18 times more expensive each year (exhibit 27). They are larger and reason more.
3 to 18 times per year
Rise in the price of running frontier models. Exhibit 27
And list prices do not always fall. A token is a unit of text that the model processes and that you pay for. Google doubles the price of Gemini 3.8 Flash starting January 1, 2027 (exhibit 148). Input goes from $0.75 to $1.50 per million tokens, and output from $3.75 to $7.50.
At OpenAI (exhibit 149), the mid tier has barely changed since 2024. gpt-4o cost $2.50 for input and $10 for output. gpt-6.1-sol costs $2 and $10.
Price per token is not the cost of the task either. At Anthropic, the new tokenizer produces about 30% more tokens for the same text (exhibit 147). The tokenizer cuts text into tokens. OpenAI’s reasoning models bill hidden tokens as output.
Gartner predicts that in customer service, the cost per case resolved with AI will exceed $3 by 2030 (exhibit 51). According to Gartner, today’s prices are subsidized. And frontier models use 3, 5 or up to 10 times the tokens for similar interactions. The prediction is known through the press.
The developer on Reddit thinks the same. He spent about $50 in API costs on a small app, against $20 for a subscription.
Then there is what the token does not count. A Reddit post put it this way: you pay for the token whether the result is right or wrong (exhibit 111). With 95% accuracy per step, these are the results:
A 5-step process succeeds 77% of the time.
A 10-step one succeeds 60% of the time.
A 20-step one succeeds under 36% of the time.
That is arithmetic with independent errors, and the reliability calculator lets you try it with your own numbers.
Finding and fixing each failure costs too. In dotnet/runtime (exhibit 68), a Microsoft engineer opened nine proposed code changes with the agent in a short stretch. They generated 5 to 9 hours of review.
The Klarna case has counterweights in the same article. Its customer service and operations cost for the third quarter of 2025 was $50 million. A year earlier it was $42 million. The company started hiring people again in May 2025. The article does not separate how much of the cost comes from the agent.
If the agent costs less per case than a person, it is worth it.
Verdict: It is only part of the math. A support team reports $0.18 per ticket with the agent, against about $3.50 with people. Klarna's CEO says its agent saved $60 million. But its customer service and operations cost went up. It was $50 million in the third quarter of 2025, against $42 million a year earlier. And Gartner predicts that the cost per resolution with AI will pass $3 by 2030.
Rule
Put a quality measure and the cost of the team that maintains the agent next to the cost per case. If after a few months it does not pay for all of that, turn it off.
Why low confidence. The supporting figures are self-reported: a small Reddit thread and unaudited statements by Klarna's CEO. The rising support cost comes from the same press article. That article does not say how much of it is due to the agent. The Gartner figure is a prediction known from the press. The rule about turning the agent off is an opinion from a Reddit thread.
Show the 4 exhibits in detailHide the detail1 for, 1 against, 2 qualify
A mid-size SaaS describes how it built a support triage agent, which sorts requests and decides who handles them. It started with passwords and billing. The agent only answers if it can cite an article or a fact from the database. People reviewed 100% of answers for the first 30 days. Refunds, cancellations, large customers and low-confidence answers go to a person. The first version without review made up a feature that did not exist. The team did not shrink: it moved to the hard cases.
Numbers
47% and 22%Tickets resolved with no person, and tickets drafted by the agent and sent by a person, at 90 days.
USD 0.18 vs about 3.50Cost per ticket with the agent vs the previous human cost.
0.85Confidence threshold: below it, the answer goes to a person.
Limitation Self-reported numbers in a small thread (3 points and 7 comments in the archive), without independent verification. The original post text was read in the public Arctic Shift archive because reddit.com blocks automated reading.
Gartner predicts that in customer service, the cost per case resolved with generative AI will exceed that of many outsourced human agents. One reason is that consumption is higher than expected and today's prices are subsidized. Also, new models spend more tokens per interaction, and infrastructure costs money.
Numbers
over 3 dollarsCost per resolution with generative AI that Gartner predicts for 2030.
3, 5 or up to 10 timesTokens frontier models use for similar interactions, according to the article.
Limitation A press article about a Gartner release (January 26, 2026) that returned 403. It is a prediction, not measured data.
“many organizations consume a lot more than they expect”
At the third quarter 2025 earnings presentation, Klarna's CEO gave two figures. He said its AI agent does the work of more than 853 full-time people and saved 60 million dollars. The same article gives data against this: customer service and operations cost rose versus the prior year. Also, the company went back to hiring people in May 2025, and an analyst calls it an example of bad deployment.
Numbers
over 853 people and 60 million dollarsEquivalent work and savings according to the CEO (it was 700 people in early 2025).
50 versus 42 million dollarsCustomer service and operations cost in the third quarter of 2025 and a year earlier.
Limitation The savings figures are the company's own and unaudited. The article was read; automated download now returns 403, so it carries no verbatim quote.
A thread asks how to measure the value of an agent. One comment asks to go back to the product's main metric. If the agent does not pay for itself and for the team that maintains it, the comment asks to shut it down. Another asks to measure the user's outcome rather than the number of agent actions. For example, did the user act on the answer, or send it without editing.
Numbers
3 times its costRevenue an agent product should generate, according to the rule one commenter proposes.
Limitation A low-scoring thread (25 points and 32 comments in the archive). The 3 times rule is an opinion, not data. It was read in the public Arctic Shift archive because reddit.com blocks automated reading. The quote was compared letter by letter with the original.
“an agentic product should generate 3x revenue for the cost of the agent”
Quote checked against the original
Read at the source. Checked on Nov 2, 2026.
What about always using the most capable model, so you do not pay for mistakes? That does not add up either. In the same Anthropic generation, the most expensive model costs 40 times the cheapest one. Claude Opus 5.5 costs $4 per million input tokens, and Claude Haiku 5.5, $0.10. In OpenAI’s gpt-6 family, the gap is 100x.
HAL is an evaluation with 21,730 runs (exhibit 9). It found that more reasoning effort lowered accuracy in most runs. In a Hacker News thread, a commenter shared their experience (exhibit 97). With work split into small pieces, cheaper models perform like frontier ones.
The better model has its argument too. With Opus 4.6, Anthropic was able to remove pieces of the harness.
Always use the most capable model: going cheap costs more in the end.
Verdict: Not always. There are gaps of 40 and 100 times between the most expensive and the cheapest model. They are models of the same generation or family. In HAL, more reasoning effort lowered accuracy in most runs. Sometimes the better model does simplify the system. With Opus 4.6, parts of the harness, the code around the agent, could be removed.
Rule
Start with the cheap model and move to the expensive one only when a validation fails. Before paying for the expensive one, measure whether it improves your tasks.
Why medium confidence. HAL is a paper with many runs, but on public benchmarks and 2025 models. The supporting figure is a vendor case with no control group. The cheap-model-first strategy comes from Reddit and Hacker News, with no independent measurement.
Show the 6 exhibits in detailHide the detail1 for, 3 against, 2 qualify
It compares a single agent with a three-agent harness (planner, generator and evaluator) for building full applications. The harness gave clearly better results, but cost over 20 times more. Agents that grade themselves are too generous. The separate evaluator also failed until they tuned it to be skeptical and to test with an external tool. Each harness piece assumes something the model cannot do alone, and with a better model (Opus 4.6) they removed pieces.
Numbers
20 minutes and 9 dollars versus 6 hours and 200 dollarsA retro game with Opus 4.5: single agent (the game did not respond to controls) versus full harness (playable).
3 hours 50 minutes and 124.70 dollarsA digital audio workstation with Opus 4.6 and the simplified harness.
Limitation The vendor's own cases, with no systematic control group.
“agents reliably skew positive when grading their own work”
Someone who runs five agents in production describes how they cut spend after a USD 4,000 bill in one month. Using a cheap model first, and the expensive one only when validation fails, worked. They also trimmed the history resent on every call and cached repeated system prompts. They also used structured output and measured cost per workflow. An endpoint returning malformed JSON multiplied usage, and it only showed up when they measured per workflow.
Numbers
70% less in 6 weeksDrop in total spend, according to the author.
60% and 30%Fewer input tokens from trimming history, and less spend with caching on agents that make many requests.
5 timesUsage increase caused by an endpoint with malformed JSON.
Limitation Self-reported figures in a small thread (1 point and 16 comments in the archive). The original text was read in the public Arctic Shift archive because reddit.com blocks automated reading. The quote is in the post text and was compared letter by letter.
It is infrastructure for evaluating agents consistently, and it measures cost and behavior as well as accuracy. More reasoning effort lowered accuracy in most runs. They found an agent that searched HuggingFace for the benchmark (the test) instead of solving the task. They also found misuse of payment cards in flight booking tasks.
Numbers
21,730 runsAgent runs with 9 models on 9 benchmarks.
about USD 40,000Cost of running the evaluation.
about 2.5 billion tokensLogs published for others to review.
Limitation Public benchmarks, not company deployments. 2025 models.
A thread about an article that asks for control flow written in code, not in prompts. One commenter says a QA agent had to go through about 200 requirement files and broke around file 30. A deterministic harness (code that fixes each step) solved it. Another says that with the work split into small pieces, cheaper models perform like frontier models. Some complain that platforms prefer prompt-only flows.
Numbers
about 200 files, breaks around 30Case of a QA agent driven only by a prompt, according to one commenter.
590 points and 296 commentsSize of the thread.
Limitation User stories, without measured data. It was read on news.ycombinator.com and through the public Algolia API. Quotes from the thread were not checked word by word, so none is used.
These are the list prices of OpenAI models, checked on November 1, 2026 and reviewed on November 2. Between the most expensive and the cheapest model of the gpt-6 family there is a 100 times gap. The mid-tier price barely changed since 2024. What improved is capability per dollar.
Numbers
USD 10 and USD 50gpt-6-astra: input and output per million tokens (cached input: USD 1).
USD 2 and USD 10gpt-6.1-sol: input and output per million tokens.
USD 0.10 and USD 0.50gpt-6-luna: input and output per million tokens.
USD 2.50 and USD 10gpt-4o, as a 2024 reference: input and output per million.
10% moreSurcharge for regional processing on models launched since March 5, 2026.
Limitation Prices expire fast: check them before using these figures. The page shows no update date. Reasoning models bill hidden tokens as output, so price per token is not enough to estimate cost per task.
These are the list prices of Anthropic models, checked on November 1, 2026 and reviewed on November 2. Between the most expensive and the cheapest model of the same generation there is a 40 times gap. The page also documents costs that the table does not show. The new tokenizer splits the same text into more tokens, which are the units you pay for. US-only inference costs more, and declaring browser or computer-use tools adds thousands of input tokens per request.
Numbers
USD 4 and USD 20Claude Opus 5.5: input and output per million tokens (cache read: USD 0.20).
USD 2 and USD 10Claude Sonnet 5.5: input and output per million tokens (cache read: USD 0.10).
USD 0.10 and USD 0.50Claude Haiku 5.5, with prompts up to 100,000 tokens: input and output per million.
USD 15 and USD 75Claude Opus 4 and 4.1, the 2025 top tier: input and output per million.
about 30% more tokensEffect of the new tokenizer for the same text.
50% lessBatch API discount; it does not apply to interactive agent sessions.
1.1 timesPrice multiplier for requesting US-only inference.
Limitation Prices expire fast: check them before using these figures. Haiku 5.5 costs more with prompts over 100,000 tokens. Price per token is not cost per task: it depends on how many turns the agent takes and how much context it resends.
Both sentences of the thesis look at a number that is not the one that matters.The price of fixed performance falls. The frontier’s rises, and so do some list prices. Also, an agent uses more tokens than the price table shows. Cost per case can be lower than a person’s. But it leaves out quality, retries and the people who review and maintain the agent.
The number that helps you decide is cost per successful task. You add up the tokens of every attempt, the retries and the review hours. You divide by the tasks that succeeded.
Nobody disputes the math. Published measurements are missing: many figures in this chapter are self-reported or come from the vendors themselves. Anthropic’s game shows it in miniature. Per attempt, the single agent came out at $9. Per working game, the single agent did not deliver, and the $200 harness did.
Measured this way, the agent can come out ahead. The support team with $0.18 tickets started with simple tasks. It sends a person whatever it cannot close with confidence.
The Hacker News thread shows another path. A QA agent, for quality control, had to go through about 200 requirements files. It broke down around file 30. A deterministic harness fixed it: code that sets the steps and leaves only each piece to the model.
In a Reddit thread on product management, one comment proposed the final yardstick (exhibit 92). Go back to the product’s main metric. If the agent does not pay for itself and for the team that maintains it, switch it off. The same commenter wants it to generate 3 times its cost in revenue. That number is an opinion. The idea of switching it off if it does not pay stands on its own.
What you need to measure is the cost of each task that succeeds, not the cost of each model call.
Verdict: Yes. Tokens, the units of text the model charges for, cost the same whether the result is right or wrong. Human review is a cost too. A game built by a single agent cost $9 and did not respond to the controls. With a three-agent harness it cost $200 and was playable.
Rule
Compute the cost per successful task. Add up the tokens for every attempt, the retries and the review hours. Divide that by the tasks that came out right. Measure it per workflow, not from the monthly bill.
Why medium confidence. The math is simple and nobody disputes it. What is missing is published measurements. The game example comes from the vendor itself, with no control group. The review figure comes from a Microsoft engineer writing about a Microsoft tool. The rest are small Reddit threads with self-reported figures. The qualifying sources show something more. The math can come out in the agent's favor when the work is split into small steps.
Show the 7 exhibits in detailHide the detail5 for, 0 against, 2 qualify
The post argues that price per token is the wrong metric, because you pay the same whether the result is right or wrong. Reliability falls as the chain of steps gets longer, so the cost of checking has to be added. The author concludes that serious work cannot be automated cost-effectively. A comment proposes something more useful: measuring cost per successful outcome.
Numbers
77%, 60% and under 36%Reliability of a 5, 10 and 20 step process with 95% success per step, if errors were independent.
Limitation A very small thread (2 points and 6 comments in the archive) and the post promotes an article by the author. The math is arithmetic, not a measurement. It was read in the public Arctic Shift archive because reddit.com blocks automated reading. The quote was compared letter by letter with the original.
“Cost per successful outcome seems like a far more useful metric than cost per token”
It compares a single agent with a three-agent harness (planner, generator and evaluator) for building full applications. The harness gave clearly better results, but cost over 20 times more. Agents that grade themselves are too generous. The separate evaluator also failed until they tuned it to be skeptical and to test with an external tool. Each harness piece assumes something the model cannot do alone, and with a better model (Opus 4.6) they removed pieces.
Numbers
20 minutes and 9 dollars versus 6 hours and 200 dollarsA retro game with Opus 4.5: single agent (the game did not respond to controls) versus full harness (playable).
3 hours 50 minutes and 124.70 dollarsA digital audio workstation with Opus 4.6 and the simplified harness.
Limitation The vendor's own cases, with no systematic control group.
“agents reliably skew positive when grading their own work”
Ten months of data on the Copilot coding agent in the dotnet/runtime repository. More than half of its pull requests were merged and almost none were reverted, because each one went through expert review. It did best on mechanical, well-scoped tasks (cleanup, tests) and worst on those that need judgment (performance). An instructions file and build setup almost doubled success. The bottleneck moved from writing code to reviewing it.
Numbers
878 PRs, 67.9% successAgent pull requests and success rate (87.1% for PRs by Microsoft staff; populations not comparable).
0.6% vs 0.8%Merged PRs later reverted: the agent's vs the rest.
from 38.1% to 69%Success before and after adding instructions and setting up the build.
86.2% vs 55.1%Success with and without commits from a person in the PR.
5 to 9 hoursReview work created by nine PRs the author opened in a short time.
Limitation Written by a Microsoft engineer about a Microsoft tool. It is a repository with very expert reviewers. The author notes that the agent and the people do not do the same kind of tasks.
“we cannot compromise on review quality, on expert and experienced eyes validating the changes”
Someone who runs five agents in production describes how they cut spend after a USD 4,000 bill in one month. Using a cheap model first, and the expensive one only when validation fails, worked. They also trimmed the history resent on every call and cached repeated system prompts. They also used structured output and measured cost per workflow. An endpoint returning malformed JSON multiplied usage, and it only showed up when they measured per workflow.
Numbers
70% less in 6 weeksDrop in total spend, according to the author.
60% and 30%Fewer input tokens from trimming history, and less spend with caching on agents that make many requests.
5 timesUsage increase caused by an endpoint with malformed JSON.
Limitation Self-reported figures in a small thread (1 point and 16 comments in the archive). The original text was read in the public Arctic Shift archive because reddit.com blocks automated reading. The quote is in the post text and was compared letter by letter.
These are the list prices of Anthropic models, checked on November 1, 2026 and reviewed on November 2. Between the most expensive and the cheapest model of the same generation there is a 40 times gap. The page also documents costs that the table does not show. The new tokenizer splits the same text into more tokens, which are the units you pay for. US-only inference costs more, and declaring browser or computer-use tools adds thousands of input tokens per request.
Numbers
USD 4 and USD 20Claude Opus 5.5: input and output per million tokens (cache read: USD 0.20).
USD 2 and USD 10Claude Sonnet 5.5: input and output per million tokens (cache read: USD 0.10).
USD 0.10 and USD 0.50Claude Haiku 5.5, with prompts up to 100,000 tokens: input and output per million.
USD 15 and USD 75Claude Opus 4 and 4.1, the 2025 top tier: input and output per million.
about 30% more tokensEffect of the new tokenizer for the same text.
50% lessBatch API discount; it does not apply to interactive agent sessions.
1.1 timesPrice multiplier for requesting US-only inference.
Limitation Prices expire fast: check them before using these figures. Haiku 5.5 costs more with prompts over 100,000 tokens. Price per token is not cost per task: it depends on how many turns the agent takes and how much context it resends.
A mid-size SaaS describes how it built a support triage agent, which sorts requests and decides who handles them. It started with passwords and billing. The agent only answers if it can cite an article or a fact from the database. People reviewed 100% of answers for the first 30 days. Refunds, cancellations, large customers and low-confidence answers go to a person. The first version without review made up a feature that did not exist. The team did not shrink: it moved to the hard cases.
Numbers
47% and 22%Tickets resolved with no person, and tickets drafted by the agent and sent by a person, at 90 days.
USD 0.18 vs about 3.50Cost per ticket with the agent vs the previous human cost.
0.85Confidence threshold: below it, the answer goes to a person.
Limitation Self-reported numbers in a small thread (3 points and 7 comments in the archive), without independent verification. The original post text was read in the public Arctic Shift archive because reddit.com blocks automated reading.
A thread about an article that asks for control flow written in code, not in prompts. One commenter says a QA agent had to go through about 200 requirement files and broke around file 30. A deterministic harness (code that fixes each step) solved it. Another says that with the work split into small pieces, cheaper models perform like frontier models. Some complain that platforms prefer prompt-only flows.
Numbers
about 200 files, breaks around 30Case of a QA agent driven only by a prompt, according to one commenter.
590 points and 296 commentsSize of the thread.
Limitation User stories, without measured data. It was read on news.ycombinator.com and through the public Algolia API. Quotes from the thread were not checked word by word, so none is used.
Measure what each successful task costs, per workflow, with review included. Now:
Calculate the cost per successful task for each workflow with the cost calculator. Record it on the cost sheet, with the date of the prices.
Budget with today’s price and with a scenario where the price doubles. Check list prices every quarter.
Start with the cheap model. Move to the expensive one only when a check on the result fails. Before you pay for the expensive one, measure whether it improves your tasks.
Put a quality measure and the cost of the team that maintains the agent next to the cost per case. If after a few months the agent does not cover all of that, switch it off.
Set a spending cap per task and per day. The cap stops the agent and alerts a person.
OWASP’s list for agents calls for cost, rate or token caps with automatic cutoff (exhibit 33). OWASP is a security foundation.
Cost per hour rises near the edge of what the model can do. A Hacker News thread discussed a 2025 essay by Toby Ord (exhibit 81). According to Ord, o3 costs about $350 per hour on the longest tasks it can reach. There it still fails about half the time.
about USD 350 per hour
Cost of o3 at its full horizon, where it still fails about half the time. Exhibit 81
A stuck task that retries without stopping can produce a bill of $4,000. The cap does not have to be low: the developer who spends thousands a month has the support of their company. But it has to exist and it has to cut. If a task hits the cap, it stops and goes to a person.
An agent in production needs a spending cap that shuts it off.
Verdict: Yes. OWASP's risk list for agents calls for cost, rate or token caps with an automatic cutoff. A team running five agents got a $4,000 bill in one month. The cost per hour climbs right near the edge of what the model can do, where it fails the most. The cap does not need to be low. Some people spend $2,000 to $5,000 a month, and their company accepts it.
Rule
Set a spending cap per task and per day that stops the agent and alerts a person. When a task hits the cap, have it stop and hand off instead of retrying.
Why medium confidence. OWASP is an expert consensus list, not a measurement. The spending cases come from Reddit and from a Hacker News thread about a 2025 essay, with self-reported or estimated figures. The opposing source does not argue against a cap: it shows that high spending can be worth it.
Show the 4 exhibits in detailHide the detail3 for, 1 against, 0 qualify
It is OWASP's first consensus list built for agents. The ten risks run from agent goal hijacking (ASI01) to rogue agents (ASI10). They include tool misuse (ASI02) and identity and privilege abuse (ASI03). They also include the agentic supply chain (ASI04) and cascading failures (ASI08). It adds the principle of least agency: do not grant autonomy that is not needed. Its mitigations include least-privilege profiles per tool, human approval for destructive actions and controls outside the model. It also asks for cost, rate or token caps with automatic cutoff.
Numbers
10 risks (ASI01 to ASI10)Size of the list. ASI02 and ASI03 extend excessive agency (LLM06) from the 2025 list to multi-step flows.
Limitation It is a voluntary consensus framework: it binds no one and measures nothing. In the agent list, unbounded consumption is no longer an item of its own: it appears as a scenario and mitigation inside ASI02. The full PDF (57 pages) was read.
“deploying agentic behavior where it is not needed expands the attack surface”
Someone who runs five agents in production describes how they cut spend after a USD 4,000 bill in one month. Using a cheap model first, and the expensive one only when validation fails, worked. They also trimmed the history resent on every call and cached repeated system prompts. They also used structured output and measured cost per workflow. An endpoint returning malformed JSON multiplied usage, and it only showed up when they measured per workflow.
Numbers
70% less in 6 weeksDrop in total spend, according to the author.
60% and 30%Fewer input tokens from trimming history, and less spend with caching on agents that make many requests.
5 timesUsage increase caused by an endpoint with malformed JSON.
Limitation Self-reported figures in a small thread (1 point and 16 comments in the archive). The original text was read in the public Arctic Shift archive because reddit.com blocks automated reading. The quote is in the post text and was compared letter by letter.
A thread about an essay by Toby Ord that uses METR data to estimate the hourly cost of agents. That cost rises sharply as the task nears the limit of what the model can do, right where it fails most. The strongest counterpoint in the thread is that cheaper open models complete the same set of tasks for much less.
Numbers
about USD 350 per hourCost of o3 at its full horizon, where it still fails about half the time.
306 points and 137 commentsSize of the thread.
Limitation The essay is from 2025 and uses that year's models. The thread was read on news.ycombinator.com and through the public Algolia API, not the full essay. Quotes from the thread were not checked word by word in the source verification, so none is used.
The author spent about USD 50 of API usage on a small weekend app. At work, a USD 20 subscription covers far more use. He concludes that today's price is subsidized, and that many agent uses stop adding up with billing for real usage. Others compare it with cheap Uber rides in 2015. In contrast, one commenter spends USD 2,000 to 5,000 a month, and his company accepts it because it multiplies his output.
Numbers
about USD 50 vs 20API spend on a small app vs a monthly subscription, according to the author.
USD 2,000 to 5,000 a monthSpend by one commenter, which his company accepts.
991 points and 457 commentsSize of the thread according to the archive. The quoted comment has 74 points.
Limitation Small, self-reported figures; the thread is an opinion about where prices are going. The quote is from the favorable counterpoint, not from the thread's thesis. It was read in the public Arctic Shift archive because reddit.com blocks automated reading. The quote was compared letter by letter with the original.
“It's been a force multiplier so my company has no problem with that additional cost.”