You pay for everything the model reads and writes, and a conversation is re-sent each turn, so the bill is driven by context, agent steps and model choice more than by prompt wording.
- Long threads cost more with each turn. Each turn re-sends the whole conversation, so starting fresh when the topic changes saves money and helps accuracy.
- Attachments are charged again each turn. A file attached once stays in context for the rest of the session.
- Agents and retries multiply the bill. Each step reloads a growing trace, each handoff adds more, and abandoned runs are still billed.
- Match the model to the task. Classification, extraction, formatting and routing suit a small model; reasoning, synthesis and long-context work are worth the frontier price.
- Advertised voice rates are a floor. Speech-to-text, the LLM, text-to-speech and telephony sit on top, so price client work from the full stack, not the platform rate.
Every price here was correct when this site last checked it, between 17 Sep and 23 Sep 2026. Prices change often and this site no longer updates them, so check the vendor’s own page before you rely on one: Anthropic pricing · Retell pricing (CloudTalk’s breakdown) · Vapi pricing (Retell’s review).
You are billed for everything the model reads and everything it writes — not just your question. A long conversation means re-sending the whole conversation every turn.
That is why costs climb without your usage changing.
Where the money actually goes
Context is re-sent every turn
Turn twenty of a conversation sends turns one through nineteen again. So what a thread has cost in total grows faster than the thread does: if every turn is about the same size, doubling the number of turns roughly quadruples the tokens sent. That is arithmetic, not a measured bill — and nothing in the interface tells you it is happening.Anthropic, Using the Messages API, read at source 23 Sep 2026: “The Messages API is stateless, which means that you always send the full conversational history to the API.” The doubling-quadruples figure is this page’s own arithmetic (Reasoning, September 2026): n turns of equal size send about n²/2 turn-sized chunks in total. Prompt caching, below, lowers the price of the repeated part but not the amount re-read.
The fix is unglamorous: start a fresh conversation when the topic changes. It is also better for accuracy: in one test of 18 models, performance became less reliable as the input grew, even on simple tasks.Chroma, Context Rot: How Increasing Input Tokens Impacts LLM Performance, technical report, 14 Jul 2025, read at source 23 Sep 2026: “Our results reveal that models do not use their context uniformly; instead, their performance grows increasingly unreliable as input length grows.” One vendor’s tests, not a property proven for every model. For the conversational version, see loops that degrade.
Attached material is charged in full, every turn
A document you attached once is in context for the rest of the session. Attaching five reference files to answer one question about one of them pays for all five, repeatedly.
Agents multiply everything
An agent reads to decide what to read next. Ten steps means ten context loads, each larger than the last as the trace accumulates. The cost is not ten times a single call — it is more.Reasoning, September 2026 — arithmetic, not a measured bill: if each step re-reads the trace so far and every step adds about the same amount, ten steps read 1 + 2 + … + 10 = 55 step-sized chunks, not 10.
Multi-agent multiplies again, per handoff. Which is the economic argument on top of the reliability one in multi-agent.
Retries are silent
A call retried three times because its answer was rejected bills for all three answers. Loops that reprompt on dissatisfaction bill every pass. The runs you abandoned still cost what they cost.
The cheapest fix is usually the model
Frontier models are priced for frontier tasks. In this site’s view, much of what people run through them is not.
- Classification, extraction, formatting, routing — a small model handles these at a fraction of the cost and often faster. On Anthropic’s own list, Claude Haiku 4.5 costs a quarter of Claude Opus 5.5 per token and is rated the faster of the two.Anthropic, Pricing, price table read at source 23 Sep 2026: Claude Haiku 4.5 “$1 / MTok” base input and “$5 / MTok” output; Claude Opus 5.5 “$4 / MTok” and “$20 / MTok”. Anthropic, Models overview, read at source 23 Sep 2026: comparative latency “Fastest” for Claude Haiku 4.5, “Moderate” for Claude Opus 5.5. Whether the small model is good enough for your task is the test below, not something a price list settles.
- Reasoning, synthesis, judgement, long-context work — worth the frontier price.
- The test: would a competent junior do this correctly with the instructions you gave? If yes, a small model probably will too.
Many cost problems are model-selection problems wearing a prompt-optimisation costume.
What actually reduces spend
- Shorter context, not shorter prompts. In a long thread, the prompt you type is a small share of what gets re-read.
- Fresh conversations. By the arithmetic above, often the highest-leverage habit.
- Cap agent steps. An uncapped agent is an uncapped bill.
- Cache what does not change. Where a provider supports it, a stable system brief is far cheaper re-used than re-sent.Anthropic, Pricing, read at source 23 Sep 2026: “A cache hit costs 10% of the standard input price” (less on some newer models), while a 5-minute cache write costs “1.25x base input price”, so caching pays only when the same block is read again.
- Cap the retries. Two attempts, then stop and report — see the agent brief.
- Ask it to check rather than regenerate. A check usually writes less than a rewrite, and output tokens cost more than input on the price list above. A formal result also puts a model’s error rate when generating at a floor of roughly twice its error rate when judging whether an output is valid; that floor says the gap exists, not how big it is.Kalai, Nachum, Vempala & Zhang, Why Language Models Hallucinate, 4 Sep 2025, §1.1, read at source 23 Sep 2026: “(generative error rate) ≳ 2 ⋅ (IIV misclassification rate)”. A formal lower bound, not an empirical measurement.
If you are billing a client for this
The advertised per-minute rate is a floor, not a price. Retell publishes $0.07/min and Vapi $0.05/min — but those are base platform rates. Once speech-to-text, an LLM, text-to-speech and telephony are stacked on, the real cost lands at $0.13–$0.31 a minute.CloudTalk, Retell AI pricing, read at source 17 Sep 2026: “the advertised $0.07/min covers the voice infrastructure layer only”, with real production costs “landing most teams at $0.13–$0.31/min once a working agent is configured.” · Retell AI, Vapi AI review, read at source 17 Sep 2026: “Vapi’s hosting cost starts at $0.05 per minute, but it's actually a fraction of the total deployment cost.” Autocalls, Aircall, Trillet, Kommunicate and WhiteLabelAI analyses, Mar–Aug 2026, were also cited when this was written and were not re-read.
Reported agency economics: cost to run an agent $0.10–$0.20/min, clients billed at a $297–997 monthly retainer with 50–70% margins, or $300–1,500/month plus a $500–3,000 one-off setup.Not read at source: Trillet, Jan 2026 · Kommunicate, Aug 2026 · Suyash Raj, May 2026 — reported figures, not re-read; they describe what agencies say they spend, not a published rate.
A retainer priced off the advertised rate is a retainer priced off a number that does not exist. The gap between $0.05 and $0.31 is roughly six-fold, and it is entirely made of things the pricing page does not list.Reasoning: arithmetic on the two published figures above, 17 Sep 2026 — $0.31 divided by $0.05 is 6.2.
A retainer priced without the meter is a retainer that gets thinner the more the client uses it.