Price an agentic coding session: turns, context, tool results, cache hit rate, and output tokens, across live model rates.
100% free · no signup · browser-local
Share on XAgentic coding resends context every turn. Size the session, then price it.
210,000 tokens per session. Cost per 1,000 turns: $11.70. Cached input is billed at Claude Sonnet 5.5's cache-read rate of $0.10 per 1M.
Identical session shape, different model. Output tokens are usually the cheaper half here.
| Model | Monthly | Per session |
|---|---|---|
| Claude Sonnet 5.5 | $25.74 | $0.23 |
| Claude Opus 5.5 | $51.48 | $0.47 |
| GPT-6.1 Sol | $25.74 | $0.23 |
| Gemini 3.8 Flash | $10.23 | $0.09 |
A chat is one request. An agentic session is dozens: every turn resends the context, the tool schemas, and the conversation so far. The bill grows with the session, which is why the habits that keep a chat cheap matter more here.
CLAUDE.md, project knowledge, and retrieved files travel on every request. Keep them short and stable.
MCP and built-in tools add their schemas to the context whether or not they are called.
A byte-identical prefix qualifies for the cache-read rate, which is usually a large discount.
On agentic sessions, input dominates. Cap output anyway; it compounds across turns.
An agentic session resends the whole transcript every turn: CLAUDE.md, project knowledge, retrieved files, tool schemas, and conversation history. The cost grows with the session, not with the question. A 20-turn session can cost 20 times a single prompt.
If CLAUDE.md, project knowledge, and the system prompt are byte-stable, most providers cache that prefix and bill it at the cache-read rate. A 60 to 80 percent hit rate is realistic for a disciplined setup; below 40 percent means your prefix is changing too often.
Sonnet for the daily driver: fast and far cheaper than Opus. Opus for large refactors, hard debugging, and multi-file changes. Haiku for commit messages, summaries, and other cheap tasks. Route by difficulty instead of defaulting to the biggest model.
Keep CLAUDE.md and custom instructions short and stable so they cache. Ask for diffs instead of full rewrites. Start a new session per task. Cap output. Measure a real session before you budget a month of them.
Estimates use list input and output rates from the TokenCalc AI catalog. Cached input is priced at the model's cache-read rate when available, otherwise at the full input rate. Output tokens, retries, and negotiated discounts are excluded.
Simulate daily, monthly, and annual LLM bills across every model in the catalog.
Convert any price between per-token, per-1K, per-1M, per-word, and per-request.
Start with a monthly budget and see how many tokens and requests each model allows.
We'll email you when model prices change or new tools launch. Unsubscribe anytime.