Migrating a production AI agent to GPT-5.6: 2.2x faster, 27% cheaper

What happened when Ploy moved its AI agent to GPT-5.6?

Ploy, a company that builds websites with an AI agent, moved its production agent from Claude Opus 4.8 to GPT-5.6 Sol. Builds ran 2.2x faster and cost 27% less, and Ploy reported no drop in quality. Mean cost per build fell from $3.06 to $2.22, and time per build fell from about 8 minutes to 3 minutes 42 seconds.

Output tokens fell 48%, from 33K to 17.1K per build. The headline is the new model. The more useful story is what Ploy had to change around it.

Did the savings come from simply switching models?

No. The savings came from engineering work around the model. Ploy's prompt cache hit rate started near 0% and rose to 83.7% once it used workspace-scoped cache keys with proper breakpoints. That change alone cut uncached tokens by 28%.

Tool design mattered too. GPT-5.6 filled all 25 parameters in the code tool, including unused ones, which caused 52-64% empty file reads at first. Making optional fields "required but nullable" cut tool calls by roughly 30%. About a third of early failures came from the evaluation harness, which was built around the old model's habits.

Why does this matter for a UK small business?

It matters because you are probably buying AI work, not building it, and the price you are quoted depends on how well the system is designed. A cheaper model does not guarantee a cheaper result. Caching, tool-call shape and retries decide the bill.

Adoption is moving fast. Simply Business found 47% of UK small businesses now use AI, up from 22% a year earlier, and a further 13% plan to start within 6 to 12 months. Yet only 19% feel "very confident" using it day to day. That gap is where badly scoped, expensive projects happen.

Should I switch my AI tools to the newest model?

Not automatically. Our view at Braynex Services is that a model swap is a change to a live system, not a settings tweak. Ploy's own result shows the first run of a new model can fail in ways the old one never did, so test before you commit.

Treat any migration like a small project:

1. Pick 20 to 30 real tasks, such as enquiries, booking requests or product descriptions, and run them on both models.
2. Record cost, time, and how often a human has to fix the output.
3. Check whether failures come from the model or from your own checks and prompts.
4. Roll out to a small share of traffic first, with the old setup ready to restore.

What should I ask a supplier about AI running costs?

Ask how they keep costs down, not which model they use. A supplier who cannot explain caching, retries and tool design is quoting from the price list, not from experience. Ploy's 28% cut in uncached tokens came from design choices, not a cheaper rate.

• Do you reuse repeated instructions through prompt caching, and what hit rate do you see?
• How many retries does a typical task need, and what do failures cost?
• What happens to my data, and can I leave with it?
• If the model changes or is withdrawn, what breaks?

The data question matters because the top barriers for UK owners are security and privacy (44%), unclear use cases (39%) and accuracy concerns (36%).

Where should a small business actually start with AI?

Start where money leaks and the task is repetitive. Simply Business reports the most common uses are content creation (63%) and admin automation (46%). We would put missed calls, booking follow-ups and enquiry replies ahead of content, because they tie directly to revenue.

Fix the foundations first. An AI assistant on top of a rented platform inherits its limits. A nail salon we worked with moved off Fresha, which was costing about £1,800 a month in commission, to its own booking system at £35 a month. Once you own the system and the data, automation has something solid to work with.

What is Braynex Services' view?

Model prices will keep falling and models will keep changing. That makes your own infrastructure, meaning your website, bookings and customer data, the lasting asset. Models are swappable parts. Owning the system around them is what lets you take the 27% saving when it arrives.

It also affects visibility. Being recommended by ChatGPT, Gemini and Google AI Overviews is becoming the new local SEO, and those tools favour businesses with clear, owned, well-structured information.

If you want to know where your business is losing money and where automation would genuinely pay, book a free audit at braynexservices.com. Braynex Services will review your site, bookings and follow-ups, and give you a plain list of what to fix first.