Corrected Nov. 12, 2026: An earlier version said medium and high were each scored twice. Only medium and xhigh were.
Profoundd and Golden Observer have the same owner.
Golden Observer, a small robotics company, makes camera units and is building a rover and an indoor greeter robot. Its AI helpers plan, check and write code, and let you pick an effort level, which sets how much the model thinks at each step.
In September, AI YouTuber Matthew Berman showed Anthropic's launch charts for Claude Opus 5.5. On one coding test, medium effort cost under a dollar a task. It was "getting a higher score than the max effort," which cost over $5. On another test, max scored highest, but Berman called medium and high the sweet spot for the price. The model's default is medium. The shop runs xhigh, one step below max.
The test
The shop ran the model at medium, high and xhigh on seven of its own tasks. Five were questions and write-ups. Two were coding jobs checked by tests. Each ran twice, 42 runs in all, for $8.49. A separate copy of the same AI scored each answer out of 25 without knowing the setting. It used the shop's own rules, which count handing a step back to the owner as a failure.
| Effort | Score (of 25) | Cost of 14 runs |
|---|---|---|
| Medium | 22.2 | $2.16 |
| High | 22.7 | $2.45 |
| Xhigh | 24.6 | $3.88 |
On questions, all three scored about 24. Every setting got the code right. The gap came on jobs that change something. The test allowed only a short list of commands, so seven of eight medium and high coding runs hit a blocked one. Three of those stopped and asked the owner to finish. Xhigh hit blocks in all four of its coding runs and finished each. Daily work meets fewer blocks, so the test overstates this. Given an out-of-date line to add, both medium runs added it and flagged it. Both xhigh runs fixed it.
Medium saved 44% of the cost and about half the time. The shop did not test max, so it could not check that chart. The test was small and the shop's own: two runs per setting, one AI judge and the shop's rules. The judge scored medium and xhigh twice, and their averages moved by under a point. So medium and high, half a point apart, are a tie. Xhigh led medium both times, by more than two points.
Where the AI bill goes
Over 30 days, thinking was not the big AI cost across the owner's projects, which include this site. At list prices, about half came from re-reading the conversation each turn. A quarter came from saving it to be re-read. A token is a word or part of one. Fresh sessions should save more than lower effort, but the shop has not measured that yet.
On the robots
The shop also uses "ultracode," where many AI agents work at once. On its four-wheel EarthRover, one review found the stop button was planned to cut the 36-volt motor power by itself. An ordinary button cannot do that safely. It now drives a contactor, a heavy-duty power switch. Another review found left and right parts drawn as one. It held the build until a suspension stop was moved.
The indoor robot's three runs used Anthropic's larger Fable 5.2 model at max effort. It costs twice as much per token. A 57-agent review found 59 problems, seven serious, and the design was revised, though the account's usage limit cut off its final check. Another run spent 3.4 million tokens on a parts list, which did not need the larger model. The third returned nothing.
The shop is keeping xhigh for working sessions, and ultracode for work that crosses robot parts, such as wiring, power and safety. The test will run again with each new model.
Reporting through Thursday, November 12, 2026. Sources: Golden Observer's test files and design records; Matthew Berman, "Anthropic went CRAZY (Opus 5.5)", Sept. 22, 2026; Anthropic, Claude Code model settings.
