Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Start here · Getting started

What an agent costs, in plain figures

Written October 27, 2026 · Edited November 13, 2026

For anyone about to pay for an agent × the real monthly or per-use figure × the vendors' own pricing pages, ending with a number you can defend: which meter you are on, what it costs at the tier you need, and the two extras most likely to move it. Every price below comes from the vendor's own pricing page, in US dollars before tax.

Before you start

There are three ways to pay, and mixing them up is the most common budgeting mistake. A subscription is a fixed monthly price for a plan with usage limits: you pay the same whether you use it lightly or heavily, until you hit the limit. An API account is metered: you pay per million tokens, input and output billed separately, and there is no monthly cap except the spend limit you set. Cursor is a hybrid: a monthly plan with a usage pool, and third-party models billed separately at API list price on top. Which meter you are on decides which table below applies to you, and a subscription never covers API use — OpenAI's Help Center says API usage is billed separately from a Plus subscription, and Anthropic's pricing page says the same pool is shared across chat and Claude Code but API credits are a separate purchase.

Prices change. The figures on this page are already out of date the day a vendor changes its page — re-check the vendor's pricing page on the day you pay, not the day you read this. The agent cost estimator on this site turns the subscription figures below into a monthly estimate for one plan.

What "done" looks like

At the end of one sitting you can state your figure in one sentence: the plan or model, the monthly or per-job price, and the one extra that could move it. If you cannot name the meter you are on, you are not done — go back to step 1.

Steps

  1. Name which meter you are on. If you will chat and code inside the vendor's own app (ChatGPT, Claude, Cursor), you are on a subscription — go to step 2. If you are building something of your own that calls a model over the internet, you are on the API — go to step 3. If you will code in Cursor but run Claude or GPT models inside it, you are on the hybrid: the Cursor plan price plus third-party model use at API list price, billed separately. Cursor's pricing page says third-party model usage is not included in the monthly prices shown and must be purchased separately.
  2. Price the subscription route from the vendor's page. These are the plan prices as shown on the vendors' own pricing pages on 2026-10-27. ChatGPT: Free $0; Plus $20 a month, billed monthly, with no annual billing option; Pro 100 $100, Pro 200 $200, and Pro 500 $500 a month, all billed monthly only. Claude: Free $0; Pro $20 a month billed monthly, or $17 a month on the annual plan ($200 billed up front); Max $100 a month, with 5x Pro's usage and $100 a month in API credits; Ultra $200 a month, with 10x Pro's weekly usage, no five-hour limit for verified accounts, and $200 a month in API credits; Enterprise $20 per seat a month plus usage billed at API rates, billed annually. Claude Code is included in all paid Claude plans, and the usage pool is shared across web, desktop, mobile, and Claude Code, resetting on a rolling five-hour window with weekly limits on top — so heavy chat use eats the same allowance as heavy coding. Cursor: Free $0; Plus $8 a month (Grok models only); Premium $30 a month (4x the Plus usage); Super $100 a month (15x); Ultra $200 a month (50x); Teams from $30 per user a month; Enterprise priced on request. Cursor's usage pool refreshes weekly, and you can buy top-offs when you reach the included usage.
  3. Price the API route per million tokens, input and output separately. API prices are quoted per million tokens, and output always costs more than input — usually five times as much on the current Claude and GPT-6 models. OpenAI's API pricing page: GPT-6 Astra $10 input / $50 output; GPT-6.1 Sol $2 / $10; GPT-6.1 Terra $1 / $5; GPT-6 Luna $0.10 / $0.50; Codex model gpt-5.3-codex $1.75 / $14. Anthropic's API pricing docs: Claude Fable 5.2 $8 / $40; Claude Fable 5.1 $10 / $50; Claude Opus 5.5 $4 / $20; Claude Sonnet 5.5 $2 / $10; Claude Haiku 5.5 $0.10 / $0.50 for prompts up to 100,000 tokens, stepping up to $0.50 / $2.50 for longer prompts. Work one real job through the arithmetic before you budget anything: 100,000 input tokens and 10,000 output tokens costs $0.30 on Sonnet 5.5 (0.1 × $2, plus 0.01 × $10) and the same $0.30 on GPT-6.1 Sol; the same job on Opus 5.5 costs $0.60, on Fable 5.1 or GPT-6 Astra $1.50, on Terra $0.15, and on Haiku 5.5 with a short prompt about 2 cents. The model choice, not the vendor, is what moves the per-job figure most. The same arithmetic settles the meter choice: 67 of these jobs at $0.30 each cost $20.10 — just over a month of Claude Pro at $20 a month. Fewer than about 67 such jobs in a month, and the API route costs less than Pro on price alone; more than that, and Pro costs less — before Pro's usage limits, which still apply.
  4. Apply the three things that change the API bill. First, repeated context is cheaper than new context: prompt caching bills a cache read at a fraction of the input price — 10% of the input rate generally, 5% on Opus 5.5 and Sonnet 5.5, 2.5% on Fable 5.1 — so a repeat request carrying 100,000 cached input tokens costs $0.01 on Sonnet 5.5 instead of $0.20, before any output. Writing to the cache costs extra: 1.25 times the input rate for a 5-minute cache, or 2 times for a 1-hour cache — on Sonnet 5.5 that write costs $0.25, so one repeat read brings the total to $0.26 against $0.20 for a fresh request, and caching only starts saving money after the first repeat read on a 5-minute cache, or the second on a 1-hour cache. Second, work that can wait is cheaper: Anthropic's Batch API bills input and output at half price, so the $0.30 Sonnet job above costs $0.15 as a batch. Third, the fine print moves the other way: Claude models from the 4.7 series onward use a newer tokenizer that produces about 30% more tokens for the same text, so a rate cut does not cut a per-task bill by the same percentage; US-only inference carries a 1.1x multiplier on all token categories; and Haiku 5.5's higher prices apply to the whole request once the prompt passes 100,000 tokens. On the OpenAI side, one price on the page is explicitly temporary: GPT-5.6 Sol's promotional $4 / $20 rate is available at least through November 30, 2026 — budget the standard rate, not the promotion, for anything running past that date.
  5. Add the extras the headline price leaves out. Tax is excluded from the prices on both the Claude and Cursor pricing pages, so the figure on your card will be higher where tax applies. Web search on OpenAI's API is $10 per 1,000 calls plus the search content tokens at model rates — an agent that searches on every turn pays for the searches as well as the tokens. Usage credits and top-offs (Claude usage credits past plan limits, Cursor top-offs past the weekly pool, ChatGPT credits past a Pro allowance) are all billed separately and none of them is in the tables above. And a subscription's "unlimited" wording always sits next to usage limits: Claude's page says there is no fixed message count and limits depend on conversation length, model, and features, so a heavy week can still end at a limit screen.
  6. Check the figure on the vendor's page, on the day you pay. Open the pricing page linked in the Sources section below and confirm three things: the plan name (Cursor's current plan names are Plus, Premium, Super, and Ultra — older names you may see in articles no longer apply to new customers), the billing period (monthly or annual, and what is billed up front), and whether the figure you are budgeting is the standard rate or a promotion with an end date. If any of the three differs from this page, the vendor's page wins — this page is a snapshot, and the vendor's page is the price you will actually pay.

Check your own figure

What can go wrong

Limits

Related playbooks

Sources

Sources: Anthropic pricing, Anthropic API pricing, OpenAI API pricing, ChatGPT Plus, ChatGPT Pro, Cursor pricing.

Back to: Start here · Also: Agent cost estimator · Token cost calculator · Pick the cheapest model for the job