Your MCP tool definitions cost 64% more than your token counter says

Someone has finally put a repeatable measurement behind a line item most teams estimate by feel: how much of an agent's context window the attached MCP servers consume before the first user message ^[5]. The finding that matters is not the size of the bill but the accuracy of the meter, because on one fixed schema string Claude's tokenizer counts a median 64.1 percent higher than tiktoken's o200k_base ^[8].

Start with the prior art, because it is thinner than the citation count suggests. The author checked the six vendor-published numbers that get cited and reports that exactly one, StackOne's from 2026-03-31, is a real measurement study, covering GitHub and Atlassian and working through Cloudflare's code-mode case ^[2]. Of the other five, one uses hypothetical unnamed servers, three contain no server-specific token study at all, and one could not be located at any domain ^[3]. That leaves the published state of the art as a single March snapshot of a surface that changes every few weeks ^[4].

The replacement, called loadline, covers 14 servers under methodology 0.2.0, with one run dated 2026-08-18, monthly runs from here, and every row reproducible from the repo with your own credentials ^[5]. All three counters are pointed at the same canonical serialization from methodology section 1.5: same bytes, three counters ^[6]. Across the 12 rows that produced counts, Claude's premium over o200k runs from 47.0 percent on github to 70.5 percent on kubernetes, with 8 of the 12 between 60 and 66 ^[7]. Gemini stays within 11 percent of o200k on every row, which rules out a bigger-models-count-more effect and points at Claude's tokenizer on schema text specifically ^[9]. Every MCP cost study the author has seen counts with tiktoken, the only one of the three counters that runs offline and free ^[10]. If your loop runs on Claude, those studies are describing a context load roughly 60 percent smaller than the one you pay for ^[11].

The github row makes the gap concrete: 47 tools at 59,084 tokens under o200k_base and 86,843 under claude-opus-5 ^[12], a 27,759-token difference ^[13] that is 43 percent of a 200k window spent before the first user message ^[14].

Two qualifiers travel with that number. The 47 tools are what the server exposed to the classic personal access token this run used, and the spec lets a surface vary with the authorization presented, so the real GitHub surface is larger; loadline publishes auth scope per row for that reason ^[15]. And nobody on a tool-search client pays the naive figure: Claude Code has defaulted to progressive disclosure since January 2026, which on this surface models out to roughly 4,300 to 6,800 tokens, a 500-token stub plus 3 to 5 tools at the measured 1,257-token average ^[16] - nine to fourteen times less than the naive load ^[17]. The author flags "modeled" as load-bearing there, and it is.

Cheap surfaces exist. Cloudflare's aggregate endpoint answers tools/list with three tools, docs at 362 tokens, search at 572, execute at 660, totalling 1,596 under o200k, 37 times smaller than github's naive load, with hygiene holding at B (85.19) ^[18]; the qualifier is that Cloudflare also ships 16 product-scoped endpoints this run did not enumerate ^[19]. Google's Chrome DevTools MCP exposes 52 tools at 7,984 tokens, 153 per tool, against Playwright's 167, filesystem's 192 and linear's 335, with hygiene at B (80.28) ^[20]. Github's 1,257 tokens per tool is about eight times that ^[21].

Separately, fetch and postgres both failed to launch from a clean install at 15:38 UTC on 2026-08-18 ^[22], because both declare unbounded dependencies on the MCP Python SDK and a resolve landing on mcp 2.0.0 breaks them in two different places: a renamed McpError symbol for fetch, a missing mcp.server.fastmcp module for postgres ^[23].

Watch whether the 64 percent premium survives the next monthly run, and whether anyone measures progressive disclosure instead of modelling it.