Context7's two tools: 1,131 tokens. DeepWiki's three: 395.

Updated November 10, 2026.

Context7's two tools were 1,131 tokens. DeepWiki's three were 395. I counted both on the morning of November 10, 2026, from the JSON those servers returned for tools/list, using the live prompt and tool checker.

On gpt-5, with 200 output tokens and 40 turns a day entered, Context7 was $0.0034525 a turn and $4.143 over 30 days. DeepWiki was $0.0025325 and $3.039. The $1.104 difference is the tool text. The prompt and the output box were the same on both rows.

This site's own directory server returned seven tools and 973 tokens, still under Context7's two. Tool count did not order the bill.

What went into the checker

Each server got initialize with protocolVersion set to 2025-06-18, then notifications/initialized, then tools/list. No Authorization header. The 2025-06-18 schema defines InitializeResult.instructions as a hint that a client may add to the system prompt. Revision 2026-07-28 is a later lifecycle. I requested protocolVersion 2025-06-18. All five servers completed that handshake and returned tools/list.

The tool paste was JSON.stringify of result.tools, compact, key order unchanged. The prompt on every tool row was:

You are a coding agent. Use a tool when it can answer the question. Do not invent a repository, a library id, or a documentation URL.

The request was POST /api/agent-prompt-checker with format set to mcp, modelId set to gpt-5, outputTokens set to 200, and tasksPerDay set to 40. gpt-5 is a tiktoken o200k_base count of that pasted text. The checker does not add message framing, and it does not store the paste.

GitHub's remote server at api.githubcopilot.com/mcp/ answered 401, missing Authorization. Stripe and Notion did the same. They are not in the table. The five below answered tools/list with no token.

The names were resolve-library-id and query-docs on Context7; microsoft_docs_search, microsoft_code_sample_search, and microsoft_docs_fetch on Microsoft Learn; search_cloudflare_documentation and migrate_pages_to_workers_guide on Cloudflare; ask_wiki_question, read_wiki_contents, and read_wiki_structure on DeepWiki. This site's seven are the read-only directory tools on the MCP server page.

What gpt-5 priced

The prompt was 31 tokens on every row. Input tokens are 31 plus the tool tokens. Output tokens are the 200 I entered.

Per turn is the checker's perTurn. The 30-day column is that number times 1,200, which is 40 turns times 30 days. Context7, this site, and DeepWiki matched the thirtyDay field exactly. Microsoft's field was 3.9045000000000005. Cloudflare's was 3.0569999999999995. Those are the floating-point product of the per-turn figure and 1,200. The prices above are the decimal product.

The checker page rounds a total of at least $1 to the cent. These five rows show as $4.14, $3.90, $3.91, $3.06, and $3.04.

gpt-5's standard price on OpenAI's pricing page, read November 10, 2026, is $1.25 per million input tokens and $10 per million output tokens. That is the row the checker used. Cached, batch, and fast prices are not in it.

The 200 output tokens are $2.40 of every row: 200 × $10 / 1,000,000 × 1,200. The 31-token prompt is $0.0465. Context7's 1,131 tool tokens are $1.6965. DeepWiki's 395 are $0.5925. Add those three pieces and you get $4.143 and $3.039.

Where the extra tokens sit

I deleted the description field and posted the same objects again. The checker warned that the description was missing, left ok true, and still priced the text.

The removed tokens are the key, the quotes, and the prose, not the prose alone. Context7 still moves by 543. DeepWiki moves by 51.

Context7's resolve-library-id description tells the model to call that tool before query-docs, unless the user already passed a library id of the form /org/project. It also says: "Do not call this tool more than 3 times per question." DeepWiki's ask_wiki_question description is one sentence: ask a question about a repository and get an answer grounded in its DeepWiki. The long text is an operating rule. It is also input on every later turn.

A third paste kept only name, description, and inputSchema.

Context7's other fields are title and annotations. Microsoft Learn and DeepWiki also send outputSchema. DeepWiki adds _meta. The tools page for revision 2026-07-28 defines inputSchema as the argument schema and title as a display name. A host that forwards the whole tool object pays for the output schema. A host that forwards the three call fields does not. The checker priced the paste. It did not guess which fields a provider will inject.

I did not repeat that third paste for Cloudflare or for this site.

DeepWiki's MCP check returned no issues. ask_wiki_question still types repoName as anyOf: a string, or an array of strings. A clean MCP result does not mean that union was tested against OpenAI strict mode.

DeepWiki's hint names 22 tools the list does not return

The same initialize responses carried instructions. I posted each string alone, with the tool box empty and output tokens set to 0. Turns per day stayed 40. Cloudflare's result had no instructions text. An empty prompt came back as HTTP 400, "Paste a system prompt." That row has no instruction price.

DeepWiki's thirtyDay field was 1.2765000000000002. Microsoft's was 0.43499999999999994. The table is the decimal product of perTurn and 1,200. Context7 and this site matched the field.

DeepWiki's instructions were 851 tokens. Its tool list was 395. The hint names 25 tools. Twenty-two of those lines end with "private mode only." tools/list on this unauthenticated session returned the other three: read_wiki_structure, read_wiki_contents, and ask_wiki_question. One of the private names is list_wiki_repos. The opening lines say the devin_* tools act for one organization once a user OAuth token is present.

A client that copies instructions into the system prompt, which the 2025-06-18 schema says it may do, spends 851 tokens teaching names that were not in this tools/list. Microsoft's instructions describe the same three tools the list returned. This site's instructions are the 55-token read-only note, and they do not invent an eighth tool.

The OpenAI target rejects the same JSON

Auto-detect on DeepWiki returned ok: true, format "MCP tool," and no issues. The same paste with format set to openai returned ok: false and one error on each of the three tools:

This matches MCP tool. An OpenAI function tool uses type "function" and a parameters schema. The Responses API puts name and parameters on the tool. Chat Completions nests them under function.

format set to anthropic returned three errors as well:

This matches MCP tool. An Anthropic user-defined tool puts name, description, and input_schema on the tool, with no type "function" wrapper.

Context7 did the same with two errors on each of those targets. The token counts did not change. The checker's note on the failing rows says the cost includes the tool text, and a provider can reject the request before it bills the turn. The $3.039 is the price of the text if the call is accepted. It is not a receipt for a rejected request.

The break is the field name. MCP calls the argument schema inputSchema. OpenAI calls it parameters. Anthropic calls it input_schema. I did not rewrite the JSON and post it again, so this page does not report a token count for that rewrite.

The November 7 walkthrough, Why tool JSON costs more than your system prompt, used one weather tool in OpenAI's shape. The November 9 walkthrough, Agent tool JSON cost across models, used three support tools written for that page. The JSON here is tools/list from servers that answered on November 10.

The same Context7 paste on two other prices

Claude Fable 5.2 and grok-4.7 reused the 1,162 input tokens. The checker labels both as estimates. tiktoken does not bill those models, and this server does not run their tokenizers.

Fable 5.2 was $0.017296 a turn. Times 1,200, that is $20.7552 over 30 days. The thirtyDay field was 20.755200000000002. The checker's price is $8 per million input tokens and $40 per million output tokens. Anthropic's pricing page, read November 10, 2026, lists Claude Fable 5.2 at those two rates. The published Claude tool-use system prompt was not added. The checker said that count applies when the tools are Anthropic's shape, and this JSON is not.

grok-4.7 was $0.003524 a turn and $4.2288 over 30 days, on the under-200,000-token band. xAI's models page, read the same day, lists grok-4.7 at $2 and $6 per million tokens below 200,000 prompt tokens, and $4 and $12 once the prompt reaches that line. This paste is far under the line.

What to check before you attach the server

Call tools/list on the session you will actually run, and paste that array. DeepWiki's unauthenticated list is three tools. Its hint describes twenty-two more. Pricing the hint, or pricing a catalog you only see after OAuth, answers a different question.

Run the checker twice. The MCP target tells you the list matches the MCP shape the checker tests. The OpenAI or Anthropic target tells you the field names still have to change before that API will take the object. Auto-detect will report "MCP tool" and look finished.

Then read the description. Context7 spends 543 of its 1,131 tool tokens on description fields that say which tool to call first and when to stop. That rule is useful. It is also resent on the next turn.

MCPHallu is the paired note. The paper's Table 7, as copied on that page, puts correct-tool-only runs at 84.7% pass and runs that also call a confusion tool at 32.3% pass. A description that says "call this first" is one way a server tries to stay in the first group. The token count above is what that sentence costs when it rides along in the tool list. The paper's servers expose roughly 300 tools. This page does not price that catalog. It prices five servers that answered with no token, and between two and seven tools.

The prompt and tool checker is the page that produced these figures. The cost calculator is the shorter tool once you already know the token counts.