TL;DR: “Up to 90% lower token costs” compares the tokens billed to complete the same task two ways: raw connector context (transcripts, records, and tool payloads flowing through the model’s context window) versus Noded’s pre-aggregated Context Graph context. The reduction comes from three mechanisms — deduplication, summarize-once-reuse-many, and relevance filtering — and it is workload-dependent: context-heavy tasks see the most, light tasks see less.
Our homepage says teams see up to 90% lower token costs and a 40% improvement in response accuracy. In our token-spend guide we tell you not to trust numbers like that without the math. So here’s the math.
What we compare
The unit of measurement is tokens billed per completed task — input plus output, including every intermediate turn — for the same task performed two ways:
- Baseline: the model works from raw source context — full transcripts, CRM record exports, ticket threads, and per-tool payloads passed into the context window, the way a direct connector or multi-MCP setup delivers them.
- Noded: the model works from Context Graph context — the pre-aggregated account story Noded maintains continuously, with the task-relevant slice retrieved.
Tasks are the ones customer teams actually run all day: meeting-prep briefs, account summaries, renewal and QBR preparation, follow-up drafting, and health/risk questions across an account’s full history.
Why the difference is so large: three mechanisms
1. Deduplication. The same customer facts live in many systems — the renewal date is in the CRM, three email threads, two call transcripts, and a Slack channel. Raw context pays for every copy; the graph stores one.
2. Summarize once, reuse many. A two-hour call is roughly 50,000 tokens of transcript. In a raw-context workflow, it gets re-read — and re-billed — every time it’s relevant, and Anthropic’s engineering team has documented it flowing through context multiple times in a single multi-tool workflow. Noded distills it once, at batch prices, when it happens; every later question reads the distilled version.
3. Relevance filtering. A renewal brief doesn’t need the account’s full history — it needs commitments, open issues, sentiment, and contract facts. Retrieving the relevant slice instead of the corpus is where most per-query savings come from, and it’s also what the accuracy literature (Lost in the Middle, Context Rot) says you should do anyway: irrelevant context doesn’t just cost money, it degrades answers.
This pattern isn’t ours alone — Anthropic’s published example of filtering data before the model saw a 150,000-token workflow drop to 2,000 tokens (98.7%). Our “up to 90%” is deliberately more conservative than the best published case, because production workloads are messier than worked examples.
A worked example
“Prep me for tomorrow’s renewal call with Acme” against an account with four recorded calls, ~200 emails, 12 support tickets, and a year of CRM history:
| Raw connector context | Context Graph context | |
|---|---|---|
| Call history | 4 transcripts ≈ 180K tokens (or aggressive truncation, losing minute-43 signals) | Distilled call records with commitments and sentiment ≈ 3K tokens |
| Email + tickets | Threads and payloads ≈ 60K tokens | Open-issues and relationship slice ≈ 2K tokens |
| CRM | Record exports ≈ 8K tokens | Contract facts already joined ≈ 0.5K tokens |
| Tool definitions (multi-server MCP) | ≈ 10–55K tokens per session | One server ≈ 2K tokens |
| Order of magnitude | ≈ 260K+ tokens, before follow-up turns re-bill it | ≈ 8K tokens |
Illustrative composition, not a benchmark row — transcript and thread sizes vary; the 50K-tokens-per-two-hour-call figure and multi-server definition overhead (~1,000 tokens/tool, ~55K for a five-server setup) are from Anthropic’s engineering posts and the MCP spec discussions respectively.
On the accuracy number
The +40% response-accuracy improvement is measured by evaluating answers to account questions against ground truth in the source systems — did the model get the renewal date, the open commitments, the escalation status right — comparing raw-context and graph-context runs. The mechanism is the one the research predicts: less irrelevant context means less lost-in-the-middle failure. Accuracy math is harder to generalize than token math, and we’re more cautious with it — if you want to see it on your own data, we’ll run the comparison with you.
Limits of these numbers — read this part
- “Up to” means up to. The 90% end comes from context-heavy workloads (meeting prep, account briefs, cross-history questions). Short tasks over small contexts save far less.
- Savings scale with mess. If your data is already curated, the delta shrinks. Most customer stacks are not curated.
- Noded’s own processing isn’t free. The graph is built and maintained with models too — at batch prices, once per event, rather than per query. The savings figures are net of end-user query costs; you’re moving work from your per-prompt bill into our infrastructure, which is priced in Outcome Generations.
- Vendor prices move. Token prices changed twice in the month we published this. Ratios are more durable than dollar figures.
FAQ
Does every team see 90% savings?
No — that’s the upper end, on context-heavy workloads. The floor depends on how much raw data your current workflows push through the model.
Why does pre-aggregation reduce tokens?
Deduplication, summarize-once-reuse-many, and relevance filtering — ETL work done once as infrastructure instead of per-prompt at frontier prices.
Can we verify this ourselves?
Yes — the comparison is reproducible on your own accounts: measure tokens billed for the same brief with raw exports versus with Noded connected. Book a session and we’ll run it live.