Does company context cut AI token cost? 100 tasks, measured.
Yes. Across 100 tasks in 7 categories, Claude with Coworker's OM2 company graph used 9x less token cost than Claude with its native connectors, and 51x less with model routing on top. It finished tasks 64% faster, and reviewers preferred its answers in 84.5% of blind comparisons. We can run the same report on your company's data for free.
Last updated . Benchmark by Coworker AI; method below.
Benchmark run
Run 3 of 3
100 tasks · 7 categories · measuring cost, speed and quality
Task
Baseline
Coworker
Won
Summarize Q3 pipeline
12.4s
24¢
4.4s
2.7¢
Find accounts at churn risk
22.8s
47¢
7.9s
5.2¢
Draft the renewal email
9.4s
18¢
3.6s
2.0¢
Who owns ACME-431?
no answer
2.1s
1.4¢
Reconcile last month's invoices
31.0s
62¢
10.6s
6.9¢
Which PRs touched billing?
15.6s
31¢
5.7s
3.4¢
Summarize the board deck
18.2s
39¢
6.4s
4.3¢
Top five support escalations
20.4s
42¢
7.2s
4.7¢
Status of the SOC 2 renewal
13.8s
27¢
5.0s
3.0¢
Who asked for SSO?
26.5s
54¢
9.3s
6.0¢
Open platform roles
11.2s
21¢
4.1s
2.3¢
100 of 100 complete
Scored blind, verified by a Coworker engineer
Time and token cost per task, with the blind pairwise winner. Baseline is Claude with its own native connectors. Coworker is the same Claude, through the Coworker MCP.
Lower cost. Less waiting. Better answers.
Cost drops with every lever.
Cost is total input and output tokens per session. The first three bars are context optimization alone, no change to your model or workflow. The last adds Coworker's Optimized Routing on top.
Cost: tokens per session
2.9x cheaper
OM2 on Claude
vs. native connectors
66% cheaper
4.1x cheaper
+ Coworker LearningRoughly 40% of conversational prompts already invoke Coworker Learning successfully in production today, and for automation-style agents that figure is above 90%. Learning helps most where agents would otherwise rebuild the same queries every time, so the gains are largest on context-heavy workloads.
all 7 categories
75.5% cheaper
9.17x cheaper
+ Learning, retrieval
context-heavy tasks
89.1% cheaper
Coworker routes to the most efficient model
51x cheaper
+ Optimized Routing
stacked on top
98% cheaper
Answers land 64% faster.
Measured end to end, from question to finished answer. The same context optimization that cuts the bill also cuts the round trips, so the work comes back sooner.
Speed: time to completion
20% faster
OM2 on Claude
vs. native connectors
45.6% faster
+ Coworker LearningRoughly 40% of conversational prompts already invoke Coworker Learning successfully in production today, and for automation-style agents that figure is above 90%. Learning helps most where agents would otherwise rebuild the same queries every time, so the gains are largest on context-heavy workloads.
all 7 categories
64% faster
+ Learning, retrieval
context-heavy tasks
Preferred more than five to one.
Reviewers compared answers blind, in randomized order, against a human-authored reference. They picked OM2 in 84.5% of pairwise comparisons, run alongside every cost and speed test.
Quality: blind human preference
15.5%
Claude + native connectors
Preferred
84.5%
Claude + OM2 MCP
Preferred
Preference stayed at this level across every configuration in the study, even as cost dropped and speed climbed.
How we ran it
100 tasks across 7 categories, on a Claude harness with real connected tools. Each task was repeated until statistically significant, then replicated across every category: business operations, engineering, general business, marketing, people, product, sales. The connected sources were Google Drive, Gmail, Google Calendar, HubSpot, GitHub, Jira, Slack, Notion, Stripe.
Two things from you. We do the rest.
Our app runs the benchmark on your own systems, we cover the token costs, and you keep the numbers whichever way they land.
You
Coworker
Connect · Build your graph · Run · Verify · Tune · Repeat
You
Your part
Two steps
- Sign an NDA and tell us which systems to cover.
- Name one or two testers and have them install the Coworker benchmarking app.
Nothing to build, no scripts to write, and nothing for your engineers to maintain afterwards.
Our part
Everything else
- Connect your sources and build OM2 on your company's data.
- Run the full question set twice, once against your existing Claude setup and once through the Coworker MCP.
- Score every answer blind and verify the results ourselves.
- Tune your graph and run it again until the number settles.
We validate, adjust your graph and run it again. The last run is your number.
The ROI calculation was straightforward. We just needed to save two to two and a half hours per month per person to break even. We exceeded that in the first few weeks.
Marcin Safranow
VP IT Operations, Huuuge Games
FAQ
Frequently asked questions
Count every input and output token a session uses to finish one task, then compare setups on the same task list. Here that was 100 tasks across 7 categories on a Claude harness, with cost as total tokens per session, speed as end-to-end time, and quality as a blind pairwise preference against a human-written reference answer.
In this benchmark, giving Claude a structured company graph (Coworker's OM2) instead of its native connectors cut token cost 9x (89.1% lower) with the same model and workflow. Claude stops re-reading raw documents and pulls the facts it needs, so each task uses fewer tokens.
Not in this benchmark. With OM2 plus routing across open and closed models, total token cost fell 51x (98% lower), and reviewers still preferred the OM2 answers in 84.5% of blind comparisons against Claude with native connectors.
We ran 100 tasks, repeated until statistically significant, across 7 categories (business operations, engineering, general business, marketing, people, product, sales) on a Claude harness against 8 connected sources: Google Drive, Gmail, Google Calendar, HubSpot, GitHub, Jira, Slack, and Notion, plus Stripe. Cost is total input and output tokens per session, speed is end-to-end time to completion, and quality is a blind human pairwise preference against a human-authored reference answer, with order randomized.
Coworker Learning is the part of OM2 that improves retrieval the more your team uses it, instead of rebuilding the same queries from scratch each time. It is what takes the baseline 66% cost reduction to 75.5%, and up to 89.1% on retrieval-heavy tasks.
No. The 9x (9.17x, or 89.1% lower token cost) and 64% faster figures come from context optimization alone: no change to your model, no change to your workflow. Add Coworker's Optimized Routing across open and closed models on top of that, and total token cost drops 98%, or 51x cheaper overall.
On roughly 5% of questions, the native-connector baseline could not answer at all. Those were excluded from the cost and speed comparisons so the baseline was not unfairly deflated, but they were kept in the quality comparison, since failing to answer is itself a quality signal.
See what OM2 does for your team
The numbers above come from a standard Claude harness. We will run the same benchmark on your company's data and hand you a Context Impact Report. We cover the token costs.