Three agents, one experiment: feel the difference

The three agents

Agent Caoimhe is the frontier build: a custom agent on OpenAI GPT 5.6 Sol, priced by OpenAI at $5.00 per million input tokens and $30.00 per million output tokens. Her illustrative running cost is about RM 2,050 a month at premium usage.

Agent Nara is the budget build: an OpenClaw agent on a DeepSeek V4 Pro and Flash mix, priced by DeepSeek at $0.435 or less per million input tokens and $0.87 or less per million output tokens. Her illustrative running cost is about RM 120 a month, based on roughly 40 million input and 10 million output tokens a month at off-peak rates. Actual cost varies with usage.

Agent Melissa is the local build: a custom agent on Alibaba Qwen 3.5 122B, running entirely on two RTX 6000 Pro 96GB GPUs that BASIC owns. There is no per-token bill at all: the cost model is capex plus electricity and maintenance. Inference is private, so nothing she reads or writes leaves the building. She handles personal scheduling and calendar management across WhatsApp, email and calendar, with high context and long-running loops.

All three agents carry identical tools and identical data. Only the brain, and where it runs, differs. That is the whole point of the experiment: any difference you feel in the replies is the model, not the plumbing.

What to try

Ask anything about the Malaysian market. If you want a starting point, these questions exercise the datasets well:

• Compare median household income in Johor Bahru and Shah Alam.
• How many cafes are there in Penang?
• Which district in Selangor has the highest spending power?
• Profile the businesses within 2km of our venue.
• Where in Malaysia are pharmacies growing fastest?

The most useful move: send the same question to both agents, then compare. Look at three things. Depth: which answer goes further into the data? Caution: which agent flags what it does not know? Speed: which one comes back first? The differences are not subtle.

The rules of the sandbox

These rules are not decoration. They are a small, live example of the governance principles on the governance slide: constrain what an agent can reach, log what it does, and put a time limit on it.

• Replies take 30 to 60 seconds, and longer when the room is busy. The agents plan, query and check before answering; that takes time.
• The agents are sandboxed on isolated infrastructure, with read-only access to public Malaysian datasets.
• They cannot act on the world. No email, no spending, no reach into any company system.
• Conversations are logged for the demonstration and deleted after. Do not send personal or confidential information.
• The agents stay live for one week after the workshop, so you can keep testing from your desk.

What you are feeling

Three gaps at once. The capability gap: the frontier model tends to reason further, hedge more honestly and handle messier questions. The cost gap: the budget model does respectable work at a fraction of the price. And the sovereignty gap: the local model answers without a single token leaving hardware you control, trading per-token bills for capex. No agent is simply better. The lesson for an insights or marketing leader is that model choice is a management decision, not a technical one: match the model, and where it runs, to the job.

Sources