Story

AI Coding Costs: Why Indie Hackers Now Budget Tokens Per App

AI coding costs picked up a per-app budget on 14 November 2026. Tony Dinh wrote: “I needed to ration my tokens for each product now.” One app is “set to $500 worth of claude code token per month for development & maintainance”. The wording is his, including the spelling. He also wrote: “NO WASTED TOKENS!!!”.

The app is a small reader. On 13 November, his launch post said: “I built an app so you can read HN offline while on a flight.” He says he accidentally priced it at $19.99. Sam Whitaker’s Indie Hackers report the next day names it Offline Hacker News and says it now sells for $10 on the App Store.

The other end of the fortnight is a free tool. On 12 November, Theo wrote that T3 Code is “a free open source tool that has cost me $800k+ to make”. On 17 November, OpenAI published a lower API price on one new model, Microsoft pitched a coding model at a lower claimed cost, and a 27,280-run eval said most of the spend went to runs that did not pass.

This was written on 18 November 2026. Claude’s pricing page was read that day. Plan prices exclude tax and can change. Founder figures are what those people reported. Benchmark percentages are the claims of the companies that published them.

AI coding costs are now a per-product line item

Whitaker’s piece is the write-up of that post and of the launch the day before. On that thread, Zane Berg ran the store math. At $10, with Apple’s 15% small-business cut, his comment puts $500 of tokens at about 59 sales a month, just to break even on the agent. That arithmetic is his.

The same day as the launch, a different Dinh post was about servers: “Expecting to lower my infra cost by at least half.” The expectation is about infrastructure, separate from the token line.

The $10 figure is the App Store price in Whitaker’s report. How long new iOS apps have been waiting in review is another record from the same season. Payment-verified revenue from AI-built apps is a different ledger: apps that published a number, beside a founder who published a monthly token cap.

How much does Claude Code cost per month?

Claude’s pricing page, read on 18 November 2026, lists Pro at $20 a month billed monthly, or $17 a month when billed annually, with $200 paid up front. Pro includes Claude Code. Max is $100 a month, with 5x Pro usage and $100 a month in API credits. Ultra is $200 a month, with 10x Pro weekly usage, no 5-hour limit for verified accounts, and $200 a month in API credits. Prices on that page exclude tax and are subject to change.

Anthropic’s page on managing Claude Code costs describes enterprise deployments. Across those deployments, the average is around $13 per developer per active day, and $150 to $250 per developer per month. Costs sit below $30 per active day for 90% of users. The same page says /usage shows the session’s token usage and cost. /clear starts a fresh session, and the docs say that “costs nothing”. /compact is itself a large request. Each subagent sends its own requests.

Dinh’s $500 sits above the Ultra price, on one app. The daily and monthly averages are Anthropic’s, from enterprise deployments. Claude Code is already one of the tools indie hackers name in public. His post writes a monthly amount onto a single product.

Is a subscription cheaper than paying per token for coding agents?

A flat plan and a token cap send different bills. Pro includes Claude Code. Max and Ultra add API credits. The enterprise figures are averages from deployments. Dinh’s $500 is a cap he chose for one product. Which costs less depends on the month you actually run.

T3 Code puts the subscription on the user. Whitaker’s report on 15 November says the app runs coding agents such as Claude Code and Codex from one interface, users bring their own subscriptions, and the project’s GitHub page says it “does not resell tokens”.

On 29 October 2026, OpenAI’s Tibo posted that OpenAI is doubling usage limits for Pro tiers.

Noah Salas, on the T3 Code thread, described the maker’s side. Heavy users cannot run up the project’s bill. Usage that grows with adoption earns the project nothing. A subscription is a set bill with a ceiling. A token budget caps one product. Under bring-your-own-subscription, the user pays.

Are AI coding models getting cheaper?

OpenAI’s post says GPT-6.2 Astra’s API input and output token prices are 20% lower than GPT-6 Astra’s, with a 60% discount on cached input. The prices in the post are $8 input, $0.40 cached input, and $40 output, per million tokens. GPT-6.2 Sol keeps GPT-6.1 Sol’s API token prices. On Cognition’s FrontierCode 2.0, OpenAI says Sol comes within 0.5 points of Astra “at about one-fifth the cost”. That sentence is OpenAI’s comparison, on that benchmark. The Hacker News thread was on the front page, with more than 840 points. Counts on that thread move.

Microsoft’s post the same day claims MAI Code 2 gives “98% of the performance of GPT 6.1 Sol at 40% of the cost”, and 96% of Claude Opus 5.5’s performance at 5% of the cost. Those percentages are Microsoft’s benchmarks. A footnote says MAI Code 2’s cost is shown at its 50% promotional price. The post also says the models are “explicitly not yet tuned for” harder problems such as very long horizon agentic work. Microsoft says MAI Code 2 is available in GitHub Copilot, and that it is coming to OpenRouter, Vercel, OpenCode, and Kilo. The Hacker News thread sits beside the announcement.

On 10 November, Dinh’s list of predictions for 2030 included “Tokens will cost more, not less.” That line is his prediction. The Astra numbers are a published API price. The 98% line is a vendor benchmark at a promotional price.

Where the money actually goes

Runta’s FrontierHarness Eval v1.1, by Shilin Zhu, is also dated 17 November, and was run with Baseten. It covers 83 billion tokens across 27,280 agent runs: 22 harness configurations, 5 models, and 62 tasks.

The model explains 92.8% of the variance in cost per task. With the model held fixed, the most expensive harness cost 1.9 to 4.9 times as much per pass as the cheapest. “Most of the cost, 64.5%, went to runs that did not pass.” A higher cost did not buy a higher pass rate. The note to harness developers is to “cut steps, not just prompt tokens”, because every step sends the whole context again. The Hacker News thread is the discussion of that eval.

On the Indie Hackers thread about Dinh, Mo, posting as AppsByMo, reported his own logs: a reply late in a long Claude Code session cost about 3.3 times a reply in a fresh session, so he starts a new session for each separate job. That multiple is self-reported, from his logs.

When tokens aren’t the bottleneck

Theo’s $800k+ is his report of what the free, open-source tool cost him to make. On 2 November he posted a shorter line: “We’re broke and can’t pay more people”.

Whitaker’s 15 November report adds the usage and the constraint that followed. On 11 November, per a chart Theo posted, agents run through T3 Code used 27.9 trillion tokens. The project passed 1,000 GitHub contributors. The new guidelines say: “We are much more bottlenecked by review capacity than we are creativity, tokens or code.”

Ilia Pomirchi, on that thread, runs a few dozen Claude Code sessions. He says tokens stopped being the limit months ago, and that reading what the sessions produced is the limit. He also says Google and Microsoft startup credits do not cover Claude, and that AWS credits do. That credit claim is his.

Who owns AI-generated code is the separate argument about the file those runs write. The state of indie hacking in 2026 is the wider year these bills sit in. A coding tool can also change hands. What people posted after the Conductor acquisition is that record. The token line in this one stays with the person paying the model.

How do you cut AI coding costs?

The moves below are what those builders, the docs, and the 17 November posts already said.

Dinh’s is a ceiling per product: one app at $500 of Claude Code tokens a month. Mo’s, from his logs, is a new session for each separate job. The Claude Code docs say /usage shows session cost, /clear starts fresh and “costs nothing”, /compact is a large request, and each subagent sends its own requests.

Runta’s advice to harness developers is to “cut steps, not just prompt tokens”, because every step resends the context. Most of the cost in that eval went to runs that did not pass, and the harness still changed the cost per pass with the model held fixed. The comparison ran across 62 tasks.

OpenAI’s lever is the lower Astra token price, with Sol’s API prices left where they were. Microsoft’s lever is a cheaper-model claim, shown at a 50% promotional price, on models the post says are “explicitly not yet tuned for” very long horizon agentic work. A cheaper model for a routine edit is their pitch. The pitch is the vendor’s.

A short timeline

When What is on the record
2 November 2026 Theo: "We're broke and can't pay more people".
10 November 2026 Dinh's 2030 list includes "Tokens will cost more, not less."
11 November 2026 A chart Theo posted: 27.9 trillion tokens through T3 Code.
12 November 2026 Theo says T3 Code has cost him $800k+ to make.
13 November 2026 Dinh's offline Hacker News reader. He expects infra cost down by at least half.
14 November 2026 Dinh sets $500 of Claude Code tokens a month on one app.
15 November 2026 Indie Hackers: the $800k+ cost, the chart, and review as the bottleneck.
17 November 2026 GPT-6.2 prices, MAI Code 2, and Runta's 27,280-run eval.

Theo's posts are 2 November and 12 November. Dinh's are 10 November, 13 November on infra, the launch, and 14 November. The chart, the contributors, and the guidelines are in the 15 November report. The $10 price is in the 14 November report. The 17 November pages are OpenAI, Microsoft, and Runta.

What do you think? What do you spend on AI coding in a month, and do you budget that spend per product?

FAQ

How much does Claude Code cost per month?

Claude's pricing page, read on 18 November 2026, lists Pro at $20 a month billed monthly, or $17 a month billed annually with $200 paid up front. Pro includes Claude Code. Max is $100 a month, with 5x Pro usage and $100 a month in API credits. Ultra is $200 a month, with 10x Pro weekly usage, no 5-hour limit for verified accounts, and $200 a month in API credits. Those prices exclude tax and are subject to change. Anthropic's cost docs describe a different group, enterprise deployments: around $13 per developer per active day and $150 to $250 per developer per month, with costs below $30 per active day for 90 percent of users. Tony Dinh, on 14 November 2026, set one of his apps to $500 of Claude Code tokens a month for development and maintenance. That $500 figure is his.

Are AI coding models getting cheaper?

On 17 November 2026 OpenAI said GPT-6.2 Astra's API input and output token prices are 20 percent lower than GPT-6 Astra's, with a 60 percent discount on cached input. The prices in that post are $8 input, $0.40 cached input, and $40 output, per million tokens. GPT-6.2 Sol keeps GPT-6.1 Sol's API token prices. On Cognition's FrontierCode 2.0, OpenAI says Sol comes within 0.5 points of Astra at about one-fifth the cost. That comparison is OpenAI's. The same day Microsoft said MAI Code 2 gives 98 percent of the performance of GPT 6.1 Sol at 40 percent of the cost, and 96 percent of Claude Opus 5.5's performance at 5 percent of the cost. Those are Microsoft's benchmarks. A footnote says the MAI Code 2 cost is shown at its 50 percent promotional price, and the post says the models are explicitly not yet tuned for harder problems such as very long horizon agentic work. On 10 November, Tony Dinh's list of predictions for 2030 included the line that tokens will cost more, not less. A lower published price and that prediction are both on the record.

Is a subscription cheaper than paying per token for coding agents?

The Claude plans are monthly prices. Pro includes Claude Code. Max and Ultra add a monthly amount of API credits. Anthropic's averages of about $13 per active day and $150 to $250 a month describe enterprise deployments, which is a different figure from a consumer plan. Tony Dinh set a $500 token amount on one app, above the Ultra price on the page, as a cap he chose. T3 Code, in Sam Whitaker's Indie Hackers report on 15 November 2026, has users bring their own subscriptions, and its GitHub page says it does not resell tokens. On 29 October 2026 OpenAI's Tibo posted that OpenAI is doubling usage limits for Pro tiers, which is more room inside a subscription. These sources do not publish one price that fits every builder. A subscription is a set bill with a usage ceiling. A token budget is a cap that can sit on one product. Which costs less depends on how much the agent runs.

How do you cut AI coding costs?

The steps on the record come from builders and from vendor pages. Tony Dinh rations tokens per product and set one app to $500 of Claude Code tokens a month. Mo, commenting as AppsByMo on the Indie Hackers thread about that post, said his own logs showed a reply late in a long Claude Code session costing about 3.3 times a reply in a fresh session, so he starts a new session for each separate job. That multiple is from his logs. Anthropic's docs say /usage shows session token usage and cost, /clear starts fresh and costs nothing, /compact is itself a large request, and each subagent sends its own requests. Runta's FrontierHarness Eval v1.1, 27,280 agent runs, found that 64.5 percent of the cost went to runs that did not pass, and that a higher cost did not buy a higher pass rate. With the model fixed, the most expensive harness cost 1.9 to 4.9 times as much per pass as the cheapest. The advice to harness developers is to cut steps, not just prompt tokens, because every step sends the whole context again. OpenAI and Microsoft published lower or lower-claimed prices on 17 November 2026. Microsoft's comparison uses a promotional price, and the post says the models are not yet tuned for very long horizon agentic work. These are reports of what people and pages said.

Where does the money in an AI coding run actually go?

Runta's FrontierHarness Eval v1.1, by Shilin Zhu on 17 November 2026, run with Baseten, covers 83 billion tokens across 27,280 agent runs: 22 harness configurations, 5 models, and 62 tasks. The model explains 92.8 percent of the variance in cost per task. With the model fixed, the most expensive harness still cost 1.9 to 4.9 times as much per pass as the cheapest. Most of the cost, 64.5 percent, went to runs that did not pass. A higher cost did not buy a higher pass rate. On a smaller scale, Mo's self-reported Claude Code logs put a late reply in a long session at about 3.3 times a reply in a fresh session. That is his log, not the eval.