Skills & plugins token costs
Short answer: We track 32 skills, plugins, MCP servers, coding agents, frameworks and tools in 8 categories. 14 of them have a token basis: one typical task burns a median of 35K input and 2,500 output tokens and costs $0.0339 to $0.2880 on the mid-tier reference model (Grok 4.7) — a 8× range. The other 18 are resources such as gateways, directories and leaderboards, with no per-task token bill. Browse by category below; every entry links its official site, GitHub or docs.
Browse by category
Every skill, plugin, MCP server, coding agent, framework, open-source project and tool in our data, grouped by product kind. Category order follows src/tools.json; entries with a token basis show the cost of one typical task on our mid-tier reference model, resource entries link their official site or repository.
AI coding agents
6 entries in this category. Catalogued entries with a cost page open the full token breakdown; the rest open the official site, GitHub or docs.
Claude Code
Anthropic's terminal coding agent for multi-step software tasks.
Official site · GitHub · Docs
OpenAI Codex CLI
OpenAI's open-source terminal coding agent, billed via ChatGPT credits.
Official site · GitHub · Docs
GitHub Copilot coding agent
GitHub's asynchronous coding agent, billed with AI Credits from token use.
Browser Use
Open-source browser agent, priced by model tokens plus browser time.
Official site · GitHub · Docs
OpenAI deep research (o3-deep-research)
OpenAI's long-running research model, billed per input and output token.
Skills
4 entries in this category. Catalogued entries with a cost page open the full token breakdown; the rest open the official site, GitHub or docs.
Claude Skills
Claude Code skills that load only when invoked and cap at 5k tokens.
Google Agent Skills
Google-maintained Agent Skills for Google products and technologies.
Plugins
2 entries in this category. Catalogued entries with a cost page open the full token breakdown; the rest open the official site, GitHub or docs.
ChatGPT custom GPT / plugin
ChatGPT custom GPTs and plugins, rebuilt on skills and connected apps.
Claude Code plugins
Claude Code's installable extension format beyond skills or MCP servers.
MCP servers
2 entries in this category. Catalogued entries with a cost page open the full token breakdown; the rest open the official site, GitHub or docs.
MCP server
Model Context Protocol servers, whose tool definitions load into context.
Official site · GitHub · Docs
MCP Registry
Official registry for discovering and publishing MCP servers.
Official site · GitHub · Docs
Open-source projects
8 entries in this category. Catalogued entries with a cost page open the full token breakdown; the rest open the official site, GitHub or docs.
Aider
Open-source pair-programming CLI with published benchmark token counts.
Official site · GitHub · Docs
OpenHands
Open-source AI coding agent platform that runs locally with any model.
Official site · GitHub · Docs
Cline
Open-source VS Code coding agent with terminal and browser control.
Official site · GitHub · Docs
Continue
Open-source coding agent for VS Code and JetBrains IDEs that runs on any model.
Official site · GitHub · Docs
LibreChat
Open-source chat platform that unifies AI conversations in one interface.
Official site · GitHub · Docs
Ollama
Run open models locally or in the cloud with a simple CLI and API.
Official site · GitHub · Docs
vLLM
High-throughput, memory-efficient open-source LLM inference server.
Official site · GitHub · Docs
Vercel Agent Browser
Browser automation CLI for AI agents, maintained by Vercel Labs.
Sites & tools
6 entries in this category. Catalogued entries with a cost page open the full token breakdown; the rest open the official site, GitHub or docs.
Helicone
LLM observability and gateway layer for monitoring requests and spend.
Official site · GitHub · Docs
PromptLayer
Prompt management, evals and production tracing for LLM applications.
LiteLLM
Open-source AI gateway that tracks and caps LLM spend across providers.
Official site · GitHub · Docs
Workflow automation
2 entries in this category. Catalogued entries with a cost page open the full token breakdown; the rest open the official site, GitHub or docs.
n8n AI Agent
Workflow automation whose AI Agent nodes resend tool results every step.
Official site · GitHub · Docs
Frameworks & platforms
2 entries in this category. Catalogued entries with a cost page open the full token breakdown; the rest open the official site, GitHub or docs.
LangChain agent
Open-source agent framework that rebuilds prompt context on every step.
Official site · GitHub · Docs
Reference models
The three price points used across this page. Cheapest, mid and premium are picked by the cost of one typical task, not by brand, so the comparison does not assume which model your skill or plugin actually calls.
| Tier | Reference model | Input /1M | Cached /1M | Output /1M | Per typical task |
|---|---|---|---|---|---|
| Cheapest current model cheapest | GPT-6 Luna OpenAI | $0.1 | $0.01 | $0.5 | $0.00302 |
| Mid-tier reference model mid | Grok 4.7 xAI | $2 | $0.5 | $6 | $0.0561 |
| Premium reference model premium | GPT-6 Cyber OpenAI | $20 | $2 | $100 | $0.6035 |
Cost per task, paid skills & plugins
The table below is the typical-task scenario on the mid-tier model. Change a field to recalculate every row in your browser; the formula is the same one used by the server-rendered table.
| Tool | Category | Evidence | Cheapest | Mid-tier reference | Premium reference | Monthly at own volume |
|---|---|---|---|---|---|---|
| Aider | Coding agent | Measured | $0.00221 | $0.0339 | $0.4428 | $4.07 |
| ChatGPT custom GPT / plugin | Skill / plugin | Estimated | $0.00197 | $0.0347 | $0.3942 | $5.20 |
| Zapier Agents | Workflow automation | Estimated | $0.00228 | $0.0400 | $0.4560 | $8.00 |
| Claude Skills | Skill / plugin | Documented | $0.00251 | $0.0444 | $0.5025 | $13.31 |
| Dify | LLM app platform | Estimated | $0.00271 | $0.0489 | $0.5425 | $48.88 |
| LangChain agent | Agent framework | Estimated | $0.00293 | $0.0558 | $0.5850 | $55.75 |
| MCP server | Skill / plugin | Measured | $0.00302 | $0.0561 | $0.6035 | $22.45 |
| Browser Use | Browser & research agent | Documented | $0.00319 | $0.0585 | $0.6380 | $17.55 |
| n8n AI Agent | Workflow automation | Measured | $0.00373 | $0.0712 | $0.7450 | $35.62 |
| OpenAI Codex CLI | Coding agent | Documented | $0.00382 | $0.0722 | $0.7650 | $10.84 |
| Cursor | Coding agent | Estimated | $0.00405 | $0.0760 | $0.8100 | $15.20 |
| Claude Code | Coding agent | Documented | $0.00422 | $0.0810 | $0.8440 | $16.20 |
| GitHub Copilot coding agent | Coding agent | Estimated | $0.00618 | $0.1180 | $1.24 | $11.80 |
| OpenAI deep research (o3-deep-research) | Browser & research agent | Documented | $0.0177 | $0.2880 | $3.54 | $11.52 |
Per task costs one typical task exactly as defined in the method; monthly multiplies by each tool's own tasks-per-month number, so the column compares different volumes, not the same workload. Prices checked 2026-11-18.
Cost per task = (uncached input × input rate + cached input × cached rate + output × output rate) ÷ 1,000,000, with each tool's documented token counts and cached share.
Light, typical and heavy work
| Scenario | Tools | Input tokens | Output tokens | Median cost (mid) | Range (mid) |
|---|---|---|---|---|---|
| Light task one short request with a small context and one or two tool calls | 14 | 5K – 30K | 500 – 4K | $0.0166 | $0.0114 – $0.0750 |
| Typical task one coding or agent task with several tool calls and a medium context | 14 | 12K – 120K | 1,800 – 20K | $0.0573 | $0.0339 – $0.2880 |
| Heavy session a multi-turn agent session with a large repo or long history; input counts all tokens re-sent across turns | 14 | 40K – 300K | 4K – 50K | $0.2014 | $0.0945 – $0.5500 |
Why agents burn so many tokens
- Input grows, output stays small. Across our sample the input:output token ratio runs from 4.0× to 18× (median 14.1×). Tool definitions, system prompts and the conversation history are re-sent every turn, so the input side is where agent bills are made.
- Output is a small share of tokens but not of the bill. On the mid reference model, output accounts for 21% to 53% of a typical task's cost while it is only a fraction of the token count, because output is priced higher than input.
- Caching already changes the input rate. The cached share in our data ranges from 30% to 70% of input tokens; on the mid model an already-cached prefix costs $0.5 per 1M against $2 uncached, saving 19%–44% of the task price where it applies.
- Volume compounds whatever one task costs. The tools in our sample expect 40 to 1,000 tasks a month. At the median task cost that is $2.29 to $57.31 a month on the mid model, before any subscription fee.
- Tool choice is the biggest lever. The same typical task costs $0.0339 on the cheapest tool in our sample and $0.2880 on the most expensive one — a 8× spread at identical model prices.
None of this is a flaw in a particular tool: an agent needs the tool schema, the system prompt and the history to act, and a long session re-sends all three on every step. The practical question is which part of that structure you can cache, trim or batch, not whether the tool "uses too many tokens".
How to cut skill & plugin token cost
Each suggestion comes from the tool's own entry in our data or from the way its token counts are measured; the tool that supplied it is named so you can check the source.
- Add only the files that need editing instead of the whole repo. — Aider
- Use --cache-prompts to reuse the system prompt and repo map prefix across turns. — Aider
- Avoid high reasoning effort for routine edits; reasoning tokens are billed as output. — Aider
- Keep the chat short and start a new session when moving to unrelated files. — Aider
- Keep GPT/plugin instructions compact and move long reference text into fewer knowledge files. — ChatGPT custom GPT / plugin
- Start a fresh chat when the topic changes so old history is not resent. — ChatGPT custom GPT / plugin
- Attach only the documents needed for the current task. — ChatGPT custom GPT / plugin
- Disable capabilities the GPT does not use, since each tool adds instructions and possible calls. — ChatGPT custom GPT / plugin
Method and sources
The cost formula
Cost per task = (uncached input × input rate + cached input × cached rate + output × output rate) ÷ 1,000,000. Input and cached share come from src/tools.json; rates come from src/prices.json, which copies each provider's official pricing page. The same formula runs in /toolcost.js when you recalculate above and in the site calculator.
Reference models: cheapest, mid and premium
We rank current models by the cost of one typical agent task, keep the cheapest and the most expensive, and pick the mid model closest to their geometric midpoint while avoiding the two end providers when the data allows it. That makes the three columns a cross-vendor price range rather than one lineup or three copies of the same rate. The three are named in the table so you can check them: GPT-6 Luna, Grok 4.7, GPT-6 Cyber.
Evidence labels
Every tool row carries an evidence label: 6 estimated, 5 documented, 3 measured. "Documented" means the tokens come from the tool's own documentation; "measured" means a published measurement; "estimated" means the basis field states the assumption. An estimate is never presented as a measurement.
Monthly and volume
Monthly figures multiply the per-task cost by the tool's own tasks-per-month field. The default on this page is 200 tasks per month, the median of the tasks-per-month fields in src/tools.json; change it in the form above. A heavy session may count as several tasks, so compare tools on the same task count when you decide.
Missing data is not estimated
If a tool has no token basis, it does not appear in the ranking instead of getting a placeholder number. Ranges on this page are computed from the tools that do have data and are labelled as such.
Token estimates for AI tools, coding agents and plugins. Evidence legend: 'documented' means the vendor publishes the token, credit or cost figures the estimate is anchored on; 'measured' means an independent or vendor-published measurement supplies token counts; 'estimated' means we modelled token use from the documented architecture and state the assumptions in basis. Input = total input tokens billed across every request in a task (so tool loops and long sessions multiply the context), not the size of one prompt. cached_share is the fraction of input tokens served from prompt cache at the vendor's cache-read rate. All figures are token-equivalents at published list prices; subscription plans can include the same usage in a flat fee. Sources were read on 2026-11-04. v3 (2026-11-05) adds the kind taxonomy, an English tagline, official/github/docs links and a cost_model flag to every entry. Entries marked cost_model=resource are sites, directories, plugins or open-source projects with no published per-task token figure, so they carry no token estimate and link to their official page instead.
Related
- Claude Code token cost
- OpenAI Codex CLI token cost
- Cursor token cost
- n8n AI Agent token cost
- Browser Use token cost
- MCP server token cost
- Best LLM for coding in 2026
- Find a model for your use case
- Token cost calculator
- Compare models on price
- Every model price by brand
Questions
Which skills, plugins and agents are on this page?
Every entry in our tool data, grouped into 8 categories: AI coding agents (6), Skills (4), Plugins (2), MCP servers (2), Open-source projects (8), Sites & tools (6), Workflow automation (2), Frameworks & platforms (2). 14 entries have a measured, documented or estimated token basis and show a per-task cost; 18 are resources such as gateways, directories and leaderboards, which carry no per-task token bill and link their official site instead.
How many tokens does an AI agent use per task?
A typical task across the 14 tools in our sample uses a median of about 35K input and 2,500 output tokens. The range is wide because a coding agent resends a repository context that a short support bot never sees, so check the tool's own row rather than an industry average.
What does one agent task cost?
On the mid-tier reference model (Grok 4.7) one typical task costs $0.0339 to $0.2880 across the tools we track. On the cheapest and premium reference models the same tasks span $0.00221 and $3.54. The tool matters more than the model tier because token counts differ by an order of magnitude.
Are these token counts measured or estimated?
Both are labelled. Of the tools with token data: 6 estimated, 5 documented, 3 measured. An "estimated" row states the assumption in its basis field, and we publish the basis instead of a single unexplained number.
Why do agents use so many more input tokens than output tokens?
Every step re-sends the system prompt, the tool schema and the conversation history, so input grows with the number of steps while the answer stays short. The published guidance for these tools is to reuse a stable prefix so the cached rate applies, and to trim or summarise history instead of resending it in full.
How can I cut the token bill for a skill, plugin or agent?
Cache the stable prefix, trim and summarise history, use a cheaper model for simple steps, batch asynchronous work and cap retries. The suggestions above are attached to the tool that supplied them, and the <a href="/guides/prompt-caching-cost-savings">caching guide</a> explains the rate maths.
How current are these numbers?
Tool token data was last reviewed on 2026-11-05 and model prices were checked on 2026-11-18. Each tool page lists the sources and dates behind its row; when a newer measurement appears the row is updated and its evidence label and date move with it.
Next step → Find a model for your use case
Price data last checked 2026-11-18 against official sources · how we verify
Price-change alerts
One short email when a provider changes a price or a promo ends. No spam, unsubscribe any time.