Finding the Performance–Cost–Speed Sweet Spot With LLMs
Contents
I have spent an unreasonable amount of time using large language models for both work and personal projects. Most of that time went to Anthropic and OpenAI, with some Kimi and Qwen mixed in. I used them for research, writing, coding, and long agent runs.
I finally feel like I have found the sweet spot for most tasks: GPT-5.6 Sol reasoning effort High in Fast mode.
It took me some time to settle on that because I kept assuming that the strongest setting must be the best one. It is not. Max can be smarter and still make me less productive.
I kept choosing the highest reasoning effort
Until this year, I was completely Claude-pilled. Claude was the model I opened first. That changed after the Fable 5 incident. Anthropic released Fable 5, suspended it days later, and later brought it back with new safeguards.
I continued to use Anthropic after that. Opus 4.8 was good. Fable 5 was often very good, but too verbose for my taste. Then came Opus 5. Yes, it is good. I still do not like it. It is somewhat slow and, gosh, it is hard to read.
I moved more of my workload to GPT-5.6 Sol. Naturally, I selected the biggest settings. I used Fast mode and reached for Codex Ultra on large jobs. Ultra can divide one coding task among several sub-agents and run them in parallel. At first, that sounded ideal. In practice, the agents sometimes expanded the plan, checked each other’s work, and found new reasons not to finish.
Sol can also get stuck without Ultra. I watched it inspect a problem, revise the plan, inspect again, and then produce another plan instead of the change I asked for. The model was busy. I was not getting an answer.
The GPT-5.6 Sol reasoning effort curve bends at High
I compared 22 configurations using the Artificial Analysis Intelligence Index. Sol shows a clear pattern:
| Effort | Index | Estimated cost per test task |
|---|---|---|
| Low | 51 | $0.18 |
| Medium | 56 | $0.29 |
| High | 57 | $0.43 |
| Xhigh | 59 | $0.63 |
| Max | 61 | $0.95 |
Medium is the obvious bargain. High adds only one point, but it gives the model more room on difficult tasks without making every interaction feel slow.
After High, the trade gets worse. Xhigh costs about 47% more for two points. Max costs more than twice as much as High for four points.
Sol Max spends about 146 seconds thinking before it starts writing in Standard mode. Moving from Xhigh to Max adds about 52.6 seconds of waiting for two points.
Working time matters more than token price
Combining time to first token with output speed gives the total working time: the initial pause plus the time needed to generate the full response.
On the full test set, Sol takes about 88 hours at High and 150 at Xhigh. That is 1.7 times the working time for two points. Max rises to roughly 270 hours.
The same problem appeared when I tried Kimi K3 inside Claude Code. Cheap tokens do not help if I need more retries or the model never closes the task.
GPT-5.6 Sol vs Claude Opus 5 at the top
Anthropic has the same problem at its upper end. Opus 5 reaches about 63 at Xhigh. Max stays near 63 after adding roughly 87 hours across the test set. More effort does not guarantee a better result.
Benchmarks also miss whether a response is pleasant to read. They do not know whether an agent recognizes that it is stuck. A model can score higher and still waste my time.