01Cost
60×
lower model cost
Per 1,000 retrieval answers
Claude Opus 5.5$29.20
gpt-oss-20b$0.49
Published prices. Switch only after your quality checks pass.
modalis consulting
We help teams reduce AI costs, get faster answers and control where their data goes. You keep the code.
Experience behind the work.
Delivered within
Trained professionals from
What a smaller model
can change.
02Speed
2.5×
faster generation
Time to write 500 words
Frontier API7.4 s
Mac Studio5.6 s
Rented H1002.9 s
One request. Reading your prompt takes extra time.
03Data control
Yours
Choose where it runs.
Your data. Your model.
- Your cloud
- Rented GPUs
- Your Macs
Local jobs stay inside. You approve any external-model calls.
Sources
September 2026 pricing example: Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens; gpt-oss-20b on AWS Bedrock at $0.07 and $0.30. The calculation uses about 5,300 tokens in and 400 out. gpt-oss reasons before it answers, so real output can run longer. Generation times convert published single-request rates to 500 words: frontier APIs at 90 tokens/second from Artificial Analysis medians; a Mac Studio M5 Max at 120 with a 7B model at 4-bit from llama.cpp; an H100 at 228 running gpt-oss-20b from an independent vLLM benchmark. These are different models and setups, not a controlled quality comparison.
How an engagement works.
From your current AI setup to one your team can run. These figures show an example.
01 Audit
Find where the money goes.
We map every model call your products and teams make: what it costs, how fast it is, how good it is, and what data it sends out.
You getA savings plan, based on your current costs.
$35,000 a month, mapped call by call
02 Agree
Agree the requirements.
Quality, cost, speed and data rules, written down and signed before we change anything.
You getA signed delivery charter.
Delivery charterSupport assistant
- Quality
- Today’s answers on 500 real tickets
- Cost
- Under one cent a ticket
- Speed
- Under two seconds
- Data
- Stays in your cloud
Signed before work starts
Agreed03 Build
Send each job to the right model.
Most calls go to a small model on your own hardware or cloud. The hard ones still reach a frontier model. Each call carries only the context it needs.
You getA setup running in your cloud.
Every requestRouted by job
82%Small modelIn your cloud
18%Frontier modelFor harder jobs
Context per call (tokens)75% less
Before12,000After3,000
04 Measure
Compare before you switch.
We run both setups on the same real tasks and compare cost, speed and answer quality against the requirements we agreed.
You getA measured comparison.
Same 500 ticketsBefore → after
- Cost / 1,000 tickets
- $41$9
- Answer time
- 3.8 s1.1 s
- Answer quality
- 91%92%
Quality checked against the agreed test set.
05 Hand over
The system is yours.
Models, code, tests and runbooks, and a team trained to run and extend them.
You getReady for your team to run.
- 01Models & weightsIn your cloud account
- 02Code & servingIn your repository
- 03Tests & runbooksTo check and maintain it
- 04Team trainingTo run and extend it
Estimate your savings
What could your AI cost?
check it with us A rough estimate from your numbers, not a quote. It assumes a smaller model costs about a tenth as much per call. The audit measures your real ones.
Fees and delivery
Know what you’re paying for.
Before work starts, we agree what to build, how to check it and what it costs. For selected early engagements, the build fee is due when those checks pass.
- What we check
- Answer quality, cost, speed and your data requirements, using agreed tasks.
- What you pay
- The proposal lists our fees and any cloud or model costs.
- Any fee tied to savings
- Agreed separately, with a limit and a defined period to measure the results.
The people you meet build the system.
Kukesh Kodess
Mike Fuller
our own product
Pick up a project where you left off.
modalis is our Mac app. It keeps a private, cited wiki of your files and AI sessions. Resume any project in a minute. Its models are trained on the app’s own tasks, never on your files.
whatsourced project summariesrunslocally, on your Macreleasetester build
Before you book
What buyers ask first.
01Will a smaller model be good enough?
It depends on the job. We test it on your real examples against the quality requirements we agree first. Work that needs a larger model keeps one.
02Do we need our own hardware?
No. Models can run in your existing cloud account, on hardware you own, or both. We recommend what fits your data rules and your volume.
03What happens on the 25-minute call?
The call is free. We look at the AI you run or plan to run, what it costs, and where data rules or speed hold you back. Then we say what may be worth a closer look.
04What does it cost?
The discovery call is free. We scope any deeper work before it starts. The proposal lists each fee and outside cost, and how we measure the result.
05Who owns the code, models, and data?
You do. We deploy in your cloud and your repositories and hand over the code, models, tests and runbooks. You keep control after the engagement ends.
06How do you handle our data?
We work inside your cloud and use your access controls. Some work can run on hardware you own. We do not use your data to train models for anyone else.
07We already use ChatGPT or Claude. Why do we need you?
Keep them where they are the best choice. We find the calls that don’t need them, move those to cheaper or private models, and reduce how much text each call needs.
08What if quality drops?
Every change has a written test on real examples, agreed before we build. A change ships only when it meets those requirements.
Next step
Find out what your AI could cost.
Free discovery call
Twenty-five minutes on your AI spend.
A free first conversation about the AI you run or plan to run: what it costs, how fast it is, and what data it sends out.
what we discuss
- 01your current AI use and monthly spend
- 02where cost, speed, or data rules hold you back
- 03what could move to a smaller or local model