Blog
For agent builders
Agents do not pay per answer, they pay per everything-they-read. We published a real session ledger so this page can argue from receipts: 78x more tokens read than written. Here is what that means for your architecture and your bill.
Agentic workloads bill differently from chat: our measured Claude Code session read 144,929 context tokens to write 1,859, a 78x ratio, and cost $0.5887 for one bug fix. For agent builders the levers are context discipline, caching (2.6x measured), and per-request model selection, in that order.
The ledger, since nobody else publishes one
Five API calls: read the repo and run tests, read the buggy file, make two edits, re-run tests, summarize. Output tokens were 1.3% of the total moved; everything else was the agent reading. Caching turned what would have been a $1.54 session into $0.59. The full per-call table is in the Fable 5 breakdown, and it generalizes: whatever model your agent runs on, context handling dominates the bill.
Three architecture consequences
Cache-stable prompts are a design requirement, not an optimization: system prompts and tool definitions that never churn keep reads at a tenth of the input rate. Context diets beat cheaper models: an agent that rereads half as much saves more than one that switches down a tier and starts failing. Per-step model selection matters most in agents: a session mixes trivial steps (run the tests) with hard ones (find the bug), and paying one flat frontier rate for both is exactly the waste a blended model eliminates per request.
Building on Pareto
One OpenAI-compatible endpoint; every request inside your agent loop runs through several LLMs and comes back as one synthesized answer, billed at cost. The quality evidence is public and Opus-anchored: seven benchmarks, same harness, 1–45¢ on the dollar per task, including agentic suites (SWE-Bench Verified, Terminal-Bench) with +0 latency on agentic work. Streaming and tool calling are supported; the latency trade on hardest-reasoning steps (up to 3x) is documented rather than discovered.
Prove it on your agent
Point a dev branch of your agent at Pareto, run your existing eval suite, and compare metered cost per completed task against your current setup. $100 in credits is roughly 170 sessions like our published one. If your agent framework needs config help, the FAQ covers the common setups.
Questions we get from this seat
Why is my agent so expensive to run?
Context reads. Our measured session moved 78x more tokens reading than writing; agents reread files, tool outputs, and history constantly. Cut rereads and cache stable context before touching the model choice.
Does prompt caching work mid-session for agents?
Yes, and it is where caching shines: session context accumulates and gets reread every call. Our measured session's cache reads billed at a tenth of input rate, saving 2.6x overall.
Can I pin specific agent steps to specific models?
With hand-rolled routing, yes, and it becomes a config file you maintain forever. A blended model does the same thing per request without the table; our benchmarks include the agentic suites so you can check the quality claim first.