Agentic Rate Card
Updated
August 14, 2026
August 14, 2026
| What you want to accomplish | Time | Power | Tokens in/out | OpenAI / Codex | Anthropic / Claude | Kimi | GLM |
|---|---|---|---|---|---|---|---|
| Ask a question, draft, or rewriteExplain an idea, improve a paragraph, or make a quick first draft.Chat | 10 sec–2min | 2 / 5 | 1–10K / 0.2–2K | GPT-5.6 Luna<$0.001–$0.005 | Claude Haiku 4.5$0.002–$0.02 | Kimi K2.5$0.001–$0.02 | GLM-4.5-Air$0.001–$0.001 |
| Do one Slack actionAsk for a channel summary, find details in a thread, or draft a reply.Connected chat | 1–5min | 2 / 5 | 10–100K / 1–8K | GPT-5.6 Luna$0.003–$0.03 | Claude Haiku 4.5$0.02–$0.14 | Kimi K2.5$0.01–$0.08 | GLM-4.5-Air$0.001–$0.02 |
| Summarize a document or meetingTurn one transcript, deck, or report into themes, decisions, and next steps.Chat + file | 3–15min | 3 / 5 | 20–150K / 2–12K | GPT-5.6 Terra$0.06–$0.44 | Claude Sonnet 5$0.06–$0.42 | Kimi K2.5$0.03–$0.35 | GLM-4.5$0.001–$0.09 |
| Search Slack or Figma and synthesizeExplore many channels, frames, or comments and turn the evidence into themes.Connected agent | 10–45min | 4 / 5 | 0.1–1M / 5–50K | GPT-5.6 Terra$0.09–$0.90 | Claude Sonnet 5$0.08–$0.80 | Kimi K2.5$0.10–$1.10 | GLM-4.5$0.03–$0.28 |
| Research, analyze data, or create a shareable documentGather sources, compare evidence, develop a point of view, and polish a DOCX, HTML file, spreadsheet, slide deck, or PDF.Research agent | 15–90min | 4 / 5 | 0.5–3M / 20–150K | GPT-5.6 Terra$0.39–$2.70 | Claude Sonnet 5$0.35–$2.40 | Kimi K2 Thinking$0.30–$3.50 | GLM-4.5$0.07–$0.88 |
| Make small edits to a web appChange styles, adjust a component, fix a contained bug, or add one new page.Coding agent | 10–60min | 4 / 5 | 0.5–3M / 5–40K | Codex · GPT-5.6 Terra$0.21–$1.38 | Claude Code · Sonnet 5$0.20–$1.30 | Kimi K2 Thinking$0.25–$3.50 | GLM-4.5$0.06–$0.88 |
| Iterate heavily on the design of an appTake many screenshots, compare visual details, and go back and forth until it feels right.Visual coding agent | 1–4hr | 4 / 5 | 2–10M / 20–120K | Codex · GPT-5.6 Terra$0.84–$4.44 | Claude Code · Sonnet 5$0.80–$4.20 | Kimi K2 Thinking$1–$12 | GLM-4.5$0.25–$3 |
| Diagnose a difficult software problem or review codeTrace behavior across a codebase, reproduce the issue, test theories, and verify a fix.Reasoning agent | 30 min–3hr | 5 / 5 | 2–15M / 20–150K | Codex · GPT-5.6 Sol$2.10–$15.75 | Claude Code · Opus 5$2–$15 | Kimi K2 Thinking$1.50–$18 | GLM-4.5$0.38–$4.5 |
| Deep, decision-ready knowledge workWork across many sources, challenge assumptions, synthesize a position, and refine it.Long-running agent | 2–6hr | 5 / 5 | 3–20M / 50–300K | Codex · GPT-5.6 Sol$3.75–$24 | Claude Code · Opus 5$3.50–$22.50 | Kimi K2 Thinking$3–$28 | GLM-4.5$0.75–$7 |
| A heavy day of software developmentImplement several features, debug, run tests, review the whole system, and revise repeatedly.Coding agent | 4–10hr | 5 / 5 | 8–40M / 0.1–0.6M | Codex · Terra → Sol$9–$48 | Claude Code · Sonnet → Opus$8.50–$45 | Kimi K2 Thinking$8–$65 | GLM-4.5$2–$16.25 |
| Build a modest first version of an app from zeroPlan the structure, create the interface, connect data, test the flows, and make it shareable.Build agent | 8–24hr | 5 / 5 | 15–80M / 0.2–1.2M | Codex · Terra → Sol$17–$96 | Claude Code · Sonnet → Opus$16–$90 | Kimi K2 Thinking$15–$130 | GLM-4.5$3.75–$32.5 |
| Build and test an AI video pipelineResearch video models, compare renders, wire the pipeline, package model weights, deploy GPU workers, and monitor cloud tests.Model + infra stack | 1–3days | 5 / 5 | 50–250M / 0.3–10M | Codex · Terra + Sol + video models$50–$500 + GPU | Claude Code · Sonnet + Opus + video models$45–$450 + GPU | Kimi K2 Thinking$45–$480 + GPU | GLM-4.5 + video models$11.25–$120 + GPU |
| Extreme: overnight team of 4–8 AI agentsSplit a large goal into parallel research, design, build, testing, and review workstreams.Multi-agent | 8–16hr | 5 / 5 | 40–250M / 0.5–4M | Codex · Terra + Sol team$45–$308 | Claude Code · Sonnet + Opus$43–$288 | Kimi K2 Thinking$40–$330 | GLM-4.5 team$10–$82.5 |
| Extreme: agent swarm across working treesRun 8–20 coding agents in parallel branches or worktrees, with continuous tests, reviews, merges, and retries.Agent swarm | 4–8hr | 5 / 5 | 60–250M / 0.3–20M | Codex · Terra + Sol swarm$200–$800 | Claude Code · Sonnet + Opus swarm$170–$690 | Kimi K2 Thinking$160–$720 | GLM-4.5 swarm$40–$180 |
2/5 fast + economical4/5 strong synthesis5/5 maximum reasoning
Input includes repeated and cached reading; output includes what the model writes or reasons through. K = thousand tokens; M = million. An agent can use tools and complete a workstream.
Validated locally: Agentic Codex and Claude Code turns processed roughly 3–5M median input tokens versus 0.4–0.9M for no-tool turns. A swarm across working trees can reach hundreds of millions of processed tokens.
Cost assumption: Chat rows use standard list prices. Agentic rows assume repeated context is mostly cached—about 15% of standard input cost—while output is full price. The provider columns show a model stack, not one model. Kimi / GLM ranges are rough API planning estimates; plans and regional pricing can differ.
Agentic workflow calculator
Estimated input1.3MEstimated output0.08M
Cost by model
API-equivalent estimate for this workflow. Cached input and GPU costs are not included.
Scale: $0–$100Bars grow as scope grows
GPT-5.6 Sol$9.08
Claude Opus 5$9.08
Grok 4.5$3.14
DeepSeek V4 Pro$0.65
Kimi K2 Thinking$2.18
GLM-4.5$0.17
One outcome establishes the base. Modifiers add the work that makes a workflow larger: project context, external sources, browser loops, visual iteration, verification, parallelism, and infrastructure.
Before you use this rate card
This is a ballpark estimate.
Use it to compare approaches and set a budget range—not as a quote, invoice, or promise of delivery time.
Actual cost can move with model choice, cached context, long files or codebases, screenshots, browser/tool loops, retries, parallel agents, and cloud or GPU usage.
Terminology
Input tokensEverything the model reads—your prompt, a pasted code file, a Slack thread, screenshots, or test output. Example: reopening a 200K-token repo adds that context again.
Output tokensEverything the model writes or reasons through. Example: a short Slack reply may be 300 tokens; a generated feature and test plan may be 20K.
Cached inputContext the provider has already seen and can reuse at a lower price. Example: an agent rereading the same repository map on its fifth test loop.
Cache writeThe first pass that stores reusable context. Example: the initial upload of a codebase or long meeting transcript before later reads become cheaper.
Tool callA concrete action outside the chat. Examples: search Figma, read Slack, edit a file, run tests, take a screenshot, or deploy to Vercel.
Working treeAn isolated checkout for one agent. Example: eight agents each develop in a separate worktree, then reviews and merges reconcile the changes.
Context windowHow much conversation and project material one model call can hold. Example: a large repo may need summaries or multiple passes when it does not fit at once.