A chat turn sends a short message and gets a reply. An agent turn sends the whole conversation, every tool definition and every earlier tool result, then does it again after each tool call. That difference, not the model, is why agent sessions cost more. The prices here are vendor list prices fetched on 18 October 2026.
Three things that multiply the bill
Context re-reads. The API is stateless. Each request carries the full history, and you are billed for all of it as input every time. Anthropic's pricing page defines a cache hit as reading a prompt prefix from the cache, which is the part of the request that repeats.
Tool loops. One task can mean a dozen model calls: read a file, run a test, read the error, edit, run again. Each call is a full request.
Tool overhead. Anthropic's page says that declaring its computer use toolset adds about 4,500 input tokens to a request, and its browser toolset about 6,600, before you say anything. Web search is billed at $10 per 1,000 searches on top of tokens, and search results count as input in later turns.
What the vendors publish about caching
| Vendor | Published cache rule |
|---|---|
| Anthropic | 5 minute cache write is 1.25x input, 1 hour write is 2x. Reads are 0.1x input for most models. The page lists 0.05x for Opus 5.5 and Sonnet 5.5 and 0.025x for Fable 5.1 |
| OpenAI | Cached input is discounted about 90 percent for GPT-5.x and GPT-6 models and 50 percent for GPT-4.1 and GPT-4o. For GPT-6 and GPT-5.6, cache writes are priced at 1.25x input |
| Context caching has its own price, for example $0.075 per million on Gemini 3.8 Flash against $0.75 input, plus storage per million tokens per hour on models that bill it |
One Anthropic detail matters for budgets: on Claude Haiku 5.5, a prompt over 100,000 tokens is priced at the higher rate, and the page says the prompt length counts cache reads and cache writes too. A cache hit does not keep a long prompt in the cheaper bracket.
What it does to one session
Using the calculator from our guide to estimating a session, a made up 40 turn agent session on Claude Sonnet 5.5 (15,000 token prefix, 3,000 new input and 600 output tokens per turn) comes to about $7.30 without caching and $0.92 with it. A 10 turn chat of the same shape comes to $0.17 and $0.07. These are list price estimates of invented scenarios.
How to keep an agent session cheaper
- Keep the stable part of the prompt first and unchanged, so it can be a cache hit.
- Keep turns inside the cache lifetime, or pay to write the cache again.
- Return less from tools. A 3,000 token tool result is re-read on every later turn.
- Start a fresh session for a new task instead of dragging a long context along.
If you connect Safeguard to an agent over MCP, tool calls are unlimited, and you can bring your own model, so the model bill is the one you control with these habits.
Quick check
Why does a tool loop raise cost even if each tool returns little?
Each call is a full request that re-sends the whole context, so the number of calls multiplies the input billed.
Sources
- Claude API pricing, fetched 18 October 2026
- OpenAI API pricing, fetched 18 October 2026
- Gemini API pricing, fetched 18 October 2026
