--- title: "AI Token Costs: Why Long Conversations Cost More" url: https://www.insulin.dev/blog/ai-token-costs-long-conversations/ canonical: https://www.insulin.dev/blog/ai-token-costs-long-conversations/ type: Blog description: "AI token costs in Insulin: why long conversations cost more, how cached context, reasoning and failed calls are billed, and the controls that cut them." --- # AI Token Costs: Why Long Conversations Cost More > Canonical HTML version: https://www.insulin.dev/blog/ai-token-costs-long-conversations/ 1. [Home](/) 2. [Blog](/blog/) 3. AI Token Costs: Why Long Conversations Cost More # AI Token Costs: Why Long Conversations Cost More Every message in a long conversation is sent with the turns before it, and Insulin meters those turns again each time, cached or not. What adds tokens, from extended thinking to a failed call, and the controls that keep a conversation from costing more than its question. ![Max Ma](/authors/max-ma.jpg) Max Ma Oct 15, 2026 ![AI Token Costs: Why Long Conversations Cost More](/images/blog/ai-token-costs-long-conversations/hero.png) Explore AI Summary [![](/logos/company/openai.svg)](https://chat.openai.com/?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fai-token-costs-long-conversations%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Workspace. "Summarize with ChatGPT")[![](/logos/company/anthropic.svg) ](https://claude.ai/new?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fai-token-costs-long-conversations%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Workspace. "Summarize with Claude")[![](/logos/company/gemini.svg)](https://www.google.com/search?udm=50&aep=11&q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fai-token-costs-long-conversations%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Workspace. "Summarize with Gemini")[](https://www.perplexity.ai/search/new?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fai-token-costs-long-conversations%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Workspace. "Summarize with Perplexity") Table of Contents - [What are AI token costs?](#what-are-ai-token-costs) - [Why does a long conversation cost more?](#why-does-a-long-conversation-cost-more) - [Is cached context cheaper in Insulin?](#is-cached-context-cheaper-in-insulin) - [Why do extended thinking and reasoning add output tokens?](#why-do-extended-thinking-and-reasoning-add-output-tokens) - [Why does a goal cost more than one answer?](#why-does-a-goal-cost-more-than-one-answer) - [Are failed and cancelled requests billed?](#are-failed-and-cancelled-requests-billed) - [Do attached PDFs add tokens?](#do-attached-pdfs-add-tokens) - [How do you keep AI token costs down in Insulin?](#how-do-you-keep-ai-token-costs-down-in-insulin) - [Can you see what one conversation cost?](#can-you-see-what-one-conversation-cost) - [Frequently asked questions](#frequently-asked-questions) - [Takeaways](#takeaways) _AI token costs are what a model charges for the text it reads and writes on each request: input tokens for everything sent to it, output tokens for everything it produces. A long conversation costs more because every new request resends the turns before it, and Insulin meters those resent turns as input tokens in full, whether or not the provider served them from cache._ * * * Umbrella’s operations lead spends an afternoon on one problem with the Insulin assistant: a reconciliation, two PDFs attached, a few dozen follow-ups, and extended thinking switched on for the hard part. The next morning the credit balance on **Settings → Billing** is lower than anyone expected, and the question that reaches the budget owner is a fair one: why did one conversation cost so much? Because a conversation isn’t billed once. Each message is a new request, and the model is sent the conversation so far every time. Fours’ metering page puts it in one line: “A long conversation resends earlier turns on every request, and providers charge for those tokens even when they serve them from cache.” This post covers what Insulin counts as AI tokens, what adds them in a long conversation, and the control you have over each. ## What are AI token costs? **AI token costs are the charges for the tokens a model reads and writes on each request.** Insulin meters them as two resources, **AI input tokens** and **AI output tokens**, counted separately “because every provider charges differently for each direction, so each direction needs its own rate.” Each is recorded as it’s consumed, priced against the rate in effect for that month, and taken off your credit balance. Resource What Fours’ [metering page](https://doc.fours.com/insulin/billing/metering/#tokens) counts In a conversation, that means AI input tokens ”Everything sent to the model on each request,” cached context included Your new message, plus the earlier turns resent with it AI output tokens ”Everything the model produces,” reasoning included The reply, plus any thinking the model did first Neither cached context nor reasoning has a row of its own in the usage breakdown: cached context sits inside AI input tokens, and reasoning inside AI output tokens. A conversation that ran long, thought hard or retried shows up only as more of the same two rows. ## Why does a long conversation cost more? **Because every new request carries the conversation so far, so the input grows with each turn.** The same question costs more late in a long conversation than at the start of a fresh one, because it’s sent with the turns that came before it. The providers describe the same mechanics. Anthropic’s documentation says each turn’s input phase “contains all previous conversation history plus the current user message,” and that each response “becomes part of the input for the next turn.” OpenAI says that even when it stores the conversation for you, “all previous input tokens for responses in the chain are billed as input tokens.” In Insulin, a conversation carries more than what you typed. The agents documentation describes its short-term memory as “context from the current conversation, including earlier messages and tool results.” A long table an agent pulled from your CRM on turn three is part of that context, alongside every question and answer since. Three habits make a conversation longer than its question needs: - **Changing topic without starting over.** The new question is sent with the old topic’s turns. - **Pasting material in for one question.** The pasted report is now an earlier message. - **Pulling more data than the question needs.** A query that returns a long list adds the whole list to the conversation’s context. ## Is cached context cheaper in Insulin? **Not today. Insulin counts cached context in full, as input tokens, and prices it at the same rate as fresh input.** A provider can serve the repeated part of a long conversation from its cache, but, per Fours’ metering page, “providers charge for those tokens even when they serve them from cache,” and Fours records them as input tokens. The [billing FAQ](https://doc.fours.com/insulin/billing/faq/#is-cached-context-billed) gives the same answer: “Yes, in full, as input tokens.” The word that matters is _today_: the page says “today they are priced at the same rate as fresh input.” Plan on full price for a long conversation’s resent turns, and check the metering page before assuming a discount. ## Why do extended thinking and reasoning add output tokens? **Because thinking is output.** Fours’ metering page: “Reasoning tokens count as output. Models that think before answering bill that thinking at the output rate, and Fours records it the same way.” OpenAI says its reasoning tokens are “billed as output tokens” even though they aren’t visible through its API, and Anthropic says the same of Claude’s thinking tokens. In Insulin, the **Brain** button in the composer turns [extended thinking](https://doc.fours.com/insulin/agents/#extended-thinking) on or off. With it on, the agent shows its reasoning before it delivers a final answer, in a collapsible section above the reply. Two things follow for cost: - **The setting stays where you left it.** It “applies to the session and is remembered in your browser,” so switching it on for one hard question leaves it on for the easy ones after it. - **Collapsing it doesn’t make it free.** The reasoning was generated whether or not anyone expands it, and it’s billed as output. Some models reason before answering by design: OpenAI says its reasoning models “use internal reasoning tokens before producing a response.” Whatever reasoning a provider bills, Fours records as output. ## Why does a goal cost more than one answer? **Because a goal is many turns of work, not one reply.** For a request like “reconcile last quarter’s entitlements and write up what doesn’t match,” Insulin’s docs say “the agent may keep going across many turns until the job is done, rather than stopping at the end of its first reply.” Each round is more model work in the same conversation, and later rounds are sent with what the earlier ones added to it. The [goal card](https://doc.fours.com/insulin/agents/#watching-and-steering-a-goal) shows the goal’s status and the round it’s on. Click it to open a drawer with a timeline of rounds, a steer composer and **Cancel goal**: - **Steer early.** A steer typed while the goal is running “is picked up on the goal’s next round,” so correcting course on round two saves the rounds that would have followed the wrong plan. - **Stop what isn’t converging.** **Cancel goal** ends the work, and so does writing “Stop.”, “Cancel”, “pause” or “drop it” in the conversation. [How to keep a long-running AI task on track](/blog/keep-a-long-ai-task-on-track/) covers when a piece of work deserves a goal, and how to steer one. The balance has a backstop too. A long response that’s already streaming is checked every ten seconds and cut off once the balance is known to have run out, because, in Fours’ words, “a deep research run or a multi-tool agent loop can otherwise spend far more than the balance it was allowed against.” If that happens, [which of the four pause reasons applies, and its fix](/blog/why-ai-is-paused/) tells you what to do next. ## Are failed and cancelled requests billed? **Yes.** Fours’ metering page: “A generation that fails partway has already been run, and the provider charges Fours for it whether or not the answer reached you. Whatever token count the provider reports for the failed call is what is metered.” That paragraph opens with “Failed and cancelled requests are still billed,” so it covers a reply you stop partway too. That makes the recovery steps worth knowing: - **Failover adds a request.** If a model call fails, from a provider outage, a usage limit or a rejected key, the turn retries on your next connected provider. The second call is metered like any other, and the first for whatever its provider reported. - **Retry is a new request.** A failed turn’s **Retry** re-sends the message that failed, as a request of its own. - **Reload a turn that already finished.** An [interrupted card](https://doc.fours.com/insulin/agents/#picking-the-turn-back-up) is provisional, because the turn may still have finished. If the card is replaced by the note “This turn finished, but its reply could not be loaded here,” reload the conversation rather than retrying: the work is done, and a retry would run it again. ## Do attached PDFs add tokens? **They can, and how much depends on the model the turn runs on.** Insulin’s [attachment rules](https://doc.fours.com/insulin/agents/#file-attachments) split the models in two: Model the turn runs on What it gets from an attached PDF Anthropic, OpenAI and Gemini models They “read a PDF itself, layout, images and charts included” The Fours-hosted models and the open-source providers They “get only a PDF’s text, up to 50,000 characters”; a scanned PDF has no text to give OpenAI’s documentation says what reading the whole file means on its side: “PDF parsing includes both extracted text and page images in context, which can increase token usage.” Fours’ documentation doesn’t put a token figure on an attachment, so treat a long PDF the way you’d treat a long conversation: attach the document the question needs, not every document that might be relevant. ## How do you keep AI token costs down in Insulin? **Keep each conversation to one topic, turn thinking on only for the questions that need it, choose the default model for the work, and watch goals and retries.** Each control answers one of the things that drive the cost: What drives the cost The control in Insulin Earlier turns resent with every request Start over when the topic changes: `/clear` in the web workspace, a new thread in Slack Extended thinking and reasoning The Brain button, on for hard questions and off after them The rate of the model you’re on The Default model of the Insulin assistant or a custom agent A goal’s many rounds Steer early; **Cancel goal** when it isn’t converging Failed calls and retries Reload, rather than retry, a turn that already finished A PDF read in full Attach the document the question needs **Start over when the topic changes.** In the web workspace, send [`/clear`](https://doc.fours.com/insulin/agents/#resetting-a-conversation) in the composer. It stops any reply still being written, then “deletes the current conversation and its messages and opens a fresh, empty one for the same agent or channel — the old conversation is not kept.” Copy what you need before you send it. An organization channel’s conversation is shared by every member, so it can’t be cleared this way. Outside the web workspace, the boundary is set for you. [In Slack](https://doc.fours.com/insulin/agents/#slack-and-teams), each direct message and each thread you talk to Insulin in is “a conversation of its own,” so a new thread in your DM is a fresh start. In Microsoft Teams, the one-on-one chat is “the same conversation as your Insulin chat in the web workspace,” so time spent in either makes the one conversation longer. **Choose the default model for the work.** Each model has its own input and output rates, so routine questions cost less on a model with lower ones. There’s no model picker in the composer: you [choose once per agent](https://doc.fours.com/insulin/agents/#where-you-set-the-model), not per message. For the built-in Insulin assistant, click the pencil in the conversation header (its label reads _Insulin model settings_), find the model in the **Models** section and click its row. It takes the **Default** badge, and Insulin runs it first, failing over to the others from there. For a custom agent, **Default model** is a field in the agent’s own settings, changed by anyone who can edit the agent. ## Can you see what one conversation cost? **Not as a single figure.** The Usage panel on **Settings → Billing** groups a month’s consumption by resource, user, app, model or provider, and Fours’ [usage documentation](https://doc.fours.com/insulin/billing/usage/#what-this-panel-does-not-do) says “the finest grain available here is a grouped row for the period,” with no per-request drill-down. The closest view of one person’s long afternoon: switch to the **User** grouping so the **Users** filter fills, pick that person, then switch to **Model**. The filter stays applied as you change groupings. Expect the two figures on the page to disagree for a while. The credit balance moves “within about a minute of a request,” while the usage breakdown “is re-aggregated every couple of hours,” so a heavy conversation shows in the balance before it shows in the table. [AI usage by person, app and model](/blog/who-spent-the-ai-budget/) walks through each grouping, the unattributed rows and the filters. ## Frequently asked questions ### Why does a long AI conversation cost more? Because every new request resends the earlier turns. Fours’ metering page says a long conversation “resends earlier turns on every request,” and those turns are metered as input tokens each time. The same question costs more late in a long conversation than in a fresh one. ### Is cached context cheaper in Insulin? Not today. Insulin counts cached context in full, as input tokens, and prices it at the same rate as fresh input. Fours’ docs say providers charge for those tokens even when they serve them from cache. ### Are reasoning tokens billed as input or output? As output. Fours’ metering page says models that think before answering bill that thinking at the output rate, and Fours records it the same way. In Insulin, the Brain button turns extended thinking on or off, and your browser remembers the setting. ### Are failed or cancelled AI requests billed? Yes, both. A generation that fails partway has already run, and the provider charges for it whether or not the answer reached you. Insulin meters whatever token count the provider reports, and a Retry is billed as a new request. ### Does /clear keep the old conversation? No. `/clear` stops any reply still being written, deletes the current conversation and its messages, and opens a fresh, empty one for the same agent or channel. Copy anything you need first. An organization channel’s conversation is shared by every member, so it can’t be cleared. ### Do PDF attachments use more tokens? They can. Anthropic, OpenAI and Gemini models read the PDF itself, images and charts included, while Fours-hosted models and open-source providers get only its text, up to 50,000 characters. OpenAI says page images in context can increase token usage. ## Takeaways - **A conversation isn’t billed once.** Every new request resends the earlier turns, and Insulin meters them as input tokens each time, cached or not. - **Thinking is output.** The Brain button’s setting is remembered in your browser, so turn it off after the hard question. - **Goals, failed calls and retries all add requests.** Steer a goal early, and reload rather than retry a turn that already finished. - **Start over when the topic changes:** `/clear` in the web workspace, which deletes the old conversation, or a new thread in Slack. Usage-based pricing charges for the work a conversation does, and a long conversation does more work than it looks. For what Insulin meters, what it never bills, and how that compares with paying per seat, read [per-seat vs usage-based AI pricing](/blog/per-seat-vs-usage-based-ai-pricing/). ## Sources Primary sources for the platform rules cited above. Last verified October 15, 2026. Cloud providers change fees, eligibility, and program terms without notice — check the source before relying on a figure. - [Insulin What Is Metered — Fours Doc](https://doc.fours.com/insulin/billing/metering/) — Tokens: input and output counted separately because every provider charges differently for each direction; AI input tokens as everything sent to the model on each request and AI output tokens as everything it produces; cached context counted in full, a long conversation resending earlier turns on every request, recorded as input tokens and today priced at the same rate as fresh input; reasoning tokens counted as output; failed and cancelled requests billed for whatever token count the provider reports; the six Resource rows of the usage breakdown, none of them for cached or reasoning tokens - [Insulin Billing FAQ — Fours Doc](https://doc.fours.com/insulin/billing/faq/) — Is cached context billed: yes, in full, as input tokens, today at the same rate as fresh input; Are failed requests billed: yes, because the provider charges for work it performed whether or not the answer reached you - [Insulin Agents — Fours Doc](https://doc.fours.com/insulin/agents/) — Memory: short-term memory as context from the current conversation, including earlier messages and tool results; File Attachments: Anthropic, OpenAI and Gemini models reading a PDF itself, and the Fours-hosted models and open-source providers getting only its text, up to 50,000 characters; Extended Thinking: the Brain button in the composer, its setting applying to the session and remembered in the browser; Where you set the model: no picker in the composer, the Insulin model settings pencil, the Models section and the Default badge, and a custom agent's Default model; Model Selection: a failed call retrying on the next connected provider; Picking the turn back up: Retry re-sending the failed message, and reloading rather than retrying a turn that finished; Long-running work: a goal kept going across many turns, the round it is on, a steer picked up on the next round, Cancel goal and the stop phrases; Slack and Teams: each Slack DM and thread a conversation of its own, and the Teams chat the same conversation as the web workspace; Resetting a Conversation: /clear deleting the current conversation and its messages, organization channels excluded - [Insulin Reading Your Usage — Fours Doc](https://doc.fours.com/insulin/billing/usage/) — The Usage panel grouping a month by resource, user, app, model or provider; the finest grain a grouped row for the period, with no per-request drill-down; the credit balance moving within about a minute of a request while the breakdown is re-aggregated every couple of hours - [Insulin When AI Pauses — Fours Doc](https://doc.fours.com/insulin/billing/when-ai-pauses/) — What happens to a request in flight: a long response already streaming checked every ten seconds and cut off once the balance is known to have run out, because a deep research run or a multi-tool agent loop can otherwise spend far more than the balance it was allowed against - [Context windows - Claude Platform Docs](https://platform.claude.com/docs/en/build-with-claude/context-windows) — Each turn's input phase contains all previous conversation history plus the current user message, and each response becomes part of the input for the next turn; thinking tokens are billed as output tokens - [Conversation state | OpenAI API](https://developers.openai.com/api/docs/guides/conversation-state) — Even when using previous\_response\_id, all previous input tokens for responses in the chain are billed as input tokens - [Reasoning models | OpenAI API](https://developers.openai.com/api/docs/guides/reasoning) — Reasoning models use internal reasoning tokens before producing a response; reasoning tokens are not visible via the API but are billed as output tokens - [File inputs | OpenAI API](https://developers.openai.com/api/docs/guides/file-inputs) — On models with vision capabilities, the API extracts both text and page images from a PDF and sends both to the model; PDF parsing includes extracted text and page images in context, which can increase token usage ## Keep reading - [WorkspaceWhy AI Can't Read Your PDF, and How to Fix ItOct 15, 2026](/blog/why-ai-cant-read-your-pdf/) - [IntegrationsAI API Spend Limits When You Bring Your Own KeyOct 14, 2026](/blog/ai-api-spend-limits-byok/) - [IntegrationsClaude API Credits From Team and Max PlansOct 14, 2026](/blog/claude-api-credits-team-max-plans/) - [SecurityAI Acceptable Use Policy for an AI WorkspaceOct 12, 2026](/blog/ai-acceptable-use-policy/) [Browse every post on the Insulin Blog](/blog/) ### Stay Updated New posts, product updates and marketplace strategy are shared on LinkedIn as they publish. [Follow Fours on LinkedIn](https://www.linkedin.com/company/suger-inc)