Cost Analysis · MCP Economics

The token math of homegrown analytics.

Every direct-MCP question pays for discovery from scratch: loading docs, exploring schemas, writing SQL, retrying. Route the same question through Supper and that discovery already happened once, in a governed semantic model — so your agent's context stays manageable even during long analyses.

§ 00

The cost of knowing what's up

What a small team spends just on tokens.

$51,576/ yr

In Claude token costs alone — for a 25-person team asking 2,000 questions a month through direct MCP connections on Sonnet 5. That's before the engineers who keep the schema docs from going stale. Route the same volume through Supper and it's included in your plan.

§ 01

How tokens are used

DIY "skill soup" vs Supper's universal context

Homegrown · LLM + data source MCPs

Discovery from scratch, every call.

→~30k tokens of MD files and skill docs loaded per question
→~10 tool-call turns: schema lookups, SQL attempts, retries
→The growing thread is re-sent on every turn
→>600k cumulative input tokens per question

With Supper

One governed call, small context.

✓One governed call: question in, validated answer out
✓Semantic model, business logic, and SQL generation live in Supper
✓Claude's context stays small: ~8k input / 1.5k output
✓Same approved definitions for every user and every agent

§ 02

Cost per question

The per-question bill.

Standard published rates, effective Summer 2026

Model	Claude + direct MCPs	Via Supper MCP
Claude Sonnet 5 ($3 / $15)	$2.15	Included
Claude Opus 4.8 ($5 / $25)	$3.58	Included

Through Aug 31, 2026, Sonnet 5 runs an introductory $2 / $10 rate — about 33% cheaper than shown above. Questions routed through Supper come off the Claude invoice entirely; they're covered by your Supper compute-based plan, no seat fees.

The part that doesn't show up as a token

Anthropic's own data team got Claude to 21% accuracy on analytics questions out of the box — even after it had read every doc, query, and dashboard available. The homegrown bill is tokens plus the engineers who keep that structure honest. Supper carries that instead.

§ 03

DIY cost vs. Supper, by team size

Token spend comparison for different teams vs using Supper.

Claude Sonnet 5, standard rate · schema complexity and question volume scale with team size

Tier	Est. questions / mo	Est. token volume / mo	Claude + direct MCPs	Supper plan	Supper vs. DIY	Annual savings
Early Stage 10–20 people, a few data-curious teams	2,000	1.25B tokens	$4,290$51,480/yr	$2,000$24,000/yr	< 50%	$27,480/yr53% lower
Growth Stage 50–100 people, self-serve analytics across teams	2,500	1.96B tokens	$6,938$83,250/yr	$3,000$36,000/yr	43%	$47,250/yr57% lower
Late Stage 300+ people, company-wide + automated reporting	4,500	4.23B tokens	$14,850$178,200/yr	$5,000$60,000/yr	34%	$118,200/yr66% lower

Assumptions: question volume and per-question token cost both scale with tier — larger orgs run wider, more sprawling schemas, so cost per question rises from ~$2.15 (Early Stage) to ~$3.30 (Late Stage), not just volume. Claude Sonnet 5 standard rate ($3/$15 per MTok), effective Sept 1, 2026; current introductory rate is ~33% cheaper through Aug 31. Supper plan pricing as provided. Estimates only — real usage varies by actual schema size and query patterns.

§ 04

ADDITIONAL LEGACY BI WAREHOUSE SOLUTIONS

The hidden token tax of third party MCPs beyond their license fees.

BigQuery, Looker, and Snowflake — without the Supper governance layer

Warehouse MCPs bring an enormous token tax when used in conjunction with agents and LLMs.

BigQuery MCP

Raw SQL, no semantic layer

~13

turns

~850k

input tokens

→Five bare tools — list datasets, tables, table info, run SQL — with no join logic supplied
→2–3 wide get_table_info pulls on 100+ column tables before the right keys surface
→Join hallucination: failed queries on wrong or ambiguous columns

Looker MCP

LookML layer, but still wide

turns

~520k

input tokens

→Joins and metrics are pre-defined — the cost shifts from columns to fields
→Orienting on the right model and explore burns calls before any data returns
→A rich explore can expose hundreds of dimensions and measures at once

Snowflake MCP

Raw SQL, no semantic view

~15

turns

~1.05M

input tokens

→Without a semantic view, the agent can reach every database and schema the role can read
→Enterprise accounts often span many databases inherited from different systems
→Full DDL pulls on wide tables, plus repeated SQL compilation errors

Source	Claude Sonnet 5	Claude Opus 4.8	Via Supper
BigQuery MCP raw SQL, no semantic layer	$3.08	$5.13	Included
Looker MCP LookML explores	$1.95	$3.25	Included
Snowflake MCP raw SQL, no semantic view	$3.78	$6.30	Included

Claude Sonnet 5 and Opus 4.8 standard rates. Every source lands as “Included” on a Supper plan — the raw-SQL paths and the wide semantic layers alike — because the discovery work happens once, in a governed model.

§ 05

Grounding the estimate

Documented across production deployments

10–100x

more tokens for agentic, tool-calling workflows vs. single-turn chat — documented across production deployments

56K–137K

tokens per query in published multi-tool, multi-source agent benchmarks — before retries

95% → 65%

how fast Anthropic's own analytics accuracy decayed in one month once the docs stopped being maintained

Context doesn't stay flat, either: production agents are commonly observed growing from a few hundred tokens per request at launch to 10,000+ within weeks as schemas drift and nobody prunes the thread. That's the mechanism behind the 600k-token estimate above — not a hypothetical, a pattern.

§ 06

Methodology

Assumptions & sources.

Assumptions: token prices are Anthropic's published standard rates, effective Sept 1, 2026 (see note above on current introductory pricing). Homegrown path modeled as ~10 agentic turns per question with growing context; Supper path as 2 light turns. Estimates only, real usage varies — happy to model your actual numbers. Token costs for questions answered through Supper are included in your Supper plan (compute-based, no seat fees).

Sources: platform.claude.com/docs/en/about-claude/pricing · anthropic.com/news/claude-sonnet-5 · claude.com/blog, "How Anthropic enables self-service data analytics with Claude" · Google Cloud BigQuery & Looker-managed MCP docs · Snowflake-Labs MCP server docs · published agentic-retrieval & text-to-SQL token benchmarks (2026)

Next Step

We'll run the math on your numbers.

Send us your team size and question volume and we'll model the homegrown token bill against a Supper plan, side by side, with your own definitions in the loop.