As of July 2026, two stories moving through Hacker News this week point at the same operational blind spot: nobody is watching what happens before your AI agent starts working, and nobody is watching what happens when the model underneath it changes. Both cost real money, and both are things a business owner running an agent stack needs to check on a schedule, not once.
33,000 tokens before it reads your prompt
An engineering team at Systima ran a controlled comparison between two popular agentic coding tools, Claude Code and OpenCode, logging every token sent to the model before the tool ever processed the user's actual request. The result: Claude Code sent roughly 33,000 tokens of system prompt, tool definitions, and scaffolding before it read a single word the developer typed. OpenCode sent about 7,000 tokens to do the same job — a 4.7x difference in pure overhead. The post landed hard: 691 points and 381 comments on Hacker News in under two days, mostly from engineers who had noticed their usage meter climbing and never traced it back to the framework itself.
That overhead is not a one-time cost. It gets paid on every call, every session, every day the tool runs. At metered API pricing, a 4.7x tax on baseline token spend compounds fast across a team using the tool dozens of times a day. We covered this same pattern in the hidden tax inside your AI agent stack — the tools you wire together to run an agent are not neutral. Some are simply more expensive to operate for identical output, and the difference is invisible unless someone is logging tokens at the request level.
2.2x faster, 27% cheaper — from a model swap alone
The second story is a production case study from the team at Ploy, who migrated a live agent from an older model to GPT-5.6. Same agent, same task, same tools — they only swapped the model underneath it. The result: 2.2x faster response times and a 27% drop in inference cost. No architecture change, no rewrite. The variable that moved was model choice, changed at the right moment.
Put the two stories together and the lesson for anyone running or building an agent stack is straightforward: the framework you pick and the model you point it at are both live cost variables, not fixed ones. A framework can carry a built-in tax you never see on an invoice line — it just shows up as a higher token count per task. A model provider can ship a faster, cheaper option and your agent keeps paying the old rate until someone actually runs the migration.
What this costs an SMB running its own stack
If you are a local business that had a developer or an agency wire up a chatbot, a booking agent, or an outreach bot in the last year, ask two questions this week: what is the token overhead per request before your agent does anything useful, and when was the underlying model last benchmarked against what is current? Neither answer is static. Frameworks update their scaffolding, providers ship new model versions every few months, and the gap between what you are running and what is now available is where the 27% and the 4.7x numbers above come from. Nobody bills you for the gap directly — it just shows up as a bigger invoice and a slower response than a competitor running the same category of agent.
This is the part of an agent build that does not stop once the demo works. It needs someone checking token overhead and model pricing on a recurring basis, the same way you would audit a phone bill or a vendor contract. For a solo operator or a two-person shop, that is an easy thing to let slide for six months — and six months is enough time for a 27% cost gap to open up unnoticed.
What this means if you're weighing AI marketing or an agent build: the sticker price of standing up an agent is the easy part; the ongoing cost of the framework and model choices underneath it is where the real spend lives, and it needs the same recurring attention you would give any other operating expense. If you want a read on where your business actually stands when customers ask ChatGPT or Perplexity who to call, start with our free AI Visibility Report.