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.
The session ledger, interactive: flip caching off
| # | What the agent did | Context in | Out | Cost |
|---|---|---|---|---|
| 1 | Read the repo, ran the tests | 27,261 | 307 | $0.3561 |
| 2 | Read the buggy file | 28,202 | 129 | $0.0455 |
| 3 | Fixed both bugs (two edits) | 28,808 | 1,022 | $0.0869 |
| 4 | Re-ran the tests: all pass | 30,148 | 157 | $0.0534 |
| 5 | Wrote the summary | 30,500 | 244 | $0.0467 |
| One bug fix, five calls | 144,929 | 1,859 | $0.5887 |
Metered on Claude Fable 5 rates: $10/M input, $50/M output, cache writes at 1.25× input, cache reads at 0.1×. The "without caching" column re-prices the identical token movement with every call re-sending full context at the input rate. Totals also include the session's 10 uncached fresh input tokens. Full narrative in the Fable 5 breakdown.
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.
Price your agent's loop
Defaults are our measured session's averages (5 calls, ~29k context/call, 79% of context served from cache). Cache reads at 10% of input rate; the 1.25× cache-write premium is left out for simplicity (it was ~13% of our measured bill). Rates verified 2026-07-28.
Building on Pareto
Preliminary Pareto 26.10 scores and cost per task on four benchmarks are on the model card. These results may change before final publication, and no composite score is published.
Pareto speaks both API formats agents are built on, so moving a framework over is a base URL, a key and the model name pareto, not a rewrite. There is a setup guide for each framework: the Claude Agent SDK (Anthropic format, through ANTHROPIC_BASE_URL), the OpenAI Agents SDK, LangChain, LlamaIndex, CrewAI, Mastra, Pydantic AI, DSPy, the Vercel AI SDK and n8n. Coding agents, SDKs and chat apps are in the full list.
For whole workflows rather than single calls, the use-case recipes run an agent end to end on Pareto: fixing a failing CI test with Codex or Droid, batch code migrations across repos, structured extraction from text and an always-on inbox-triage assistant. Before you move a production agent, run a side-by-side on your own prompts.
Context diet vs model downgrade: which actually saves more?
Baseline: our measured session shape (5 steps, 29k context/step, 79% cached, 400 output/step) on Claude Opus 5 at $5/$25. Failed cheap runs are re-billed as a full baseline run on the frontier model, which is generous to Option B: review time is free here and is not in real life. Illustrative math with editable assumptions; the measured anchor is the 78x read ratio.
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. The $0.5887 session above is the Fable 5 bill; Pareto's published rate card is $0.80 input and $3.20 output per million tokens, 8% of Fable 5's input rate and 6.4% of its output rate, so the same 144,929 tokens read and 1,859 written would bill about $0.12 on Pareto even with no cache discount at all. At that rate $100 in credits covers more than 800 sessions like it. Config for your framework is in the setup guides above; anything they miss is in the FAQ.
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?
Yes, with a hand-built model map, and it becomes a config file you maintain forever. A blended model drops the map: every step calls Pareto, which selects frontier and open models for each request and blends their outputs into one answer; our benchmarks include the agentic suites so you can check the quality claim first.
Which agent frameworks work with Pareto?
Frameworks that accept a custom OpenAI-compatible or Anthropic base URL can usually call Pareto with the model name pareto. We publish setup guides for the Claude Agent SDK, the OpenAI Agents SDK, LangChain, LlamaIndex, CrewAI, Mastra, Pydantic AI, DSPy, the Vercel AI SDK and n8n.
Compare any two models
List rates and dated competitor measurements: prices and measured bills. Preliminary Pareto 26.10 benchmark costs per task from October 1, 2026 are on the model card. Verbosity: Edition 2.