I Found the Performance–Cost–Speed Sweet Spot With LLMs

A developer's benchmark of 22 GPT-5.6 Sol configurations found that High reasoning effort in Fast mode offers the best performance-cost-speed trade-off, scoring 57 on the Artificial Analysis Intelligence Index at an estimated $0.43 per test task, versus Max's 61 at $0.95. The author, who previously favored Anthropic's Claude models, now defaults to GPT-5.6 Sol High after observing that Max and Xhigh add disproportionate latency and cost for marginal gains, with Sol Max spending about 146 seconds thinking before writing in Standard mode.

I have spent an unreasonable amount of time using large language models. Most of that time went to Anthropic and OpenAI, with plenty of Kimi and Qwen mixed in. I used them for research, writing, coding, and long agent runs. I finally have a default for GPT-5.6 Sol reasoning effort: High in Fast mode. That sounds like a small choice. It took me months to make it 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. Until this year, I was completely Anthropic-pilled. Claude was the model I opened first. That changed after the Fable 5 incident https://philippdubach.com/posts/krugman-fable5-europe-decline/ . Anthropic released Fable 5 https://www.anthropic.com/news/claude-fable-5-mythos-5 , suspended it days later, and later brought it back with new safeguards https://www.anthropic.com/news/redeploying-fable-5 . 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. I compared 22 configurations using data captured on August 30, 2026. Each point uses the same nine tests in the Artificial Analysis Intelligence Index https://artificialanalysis.ai/models/releases/gpt-5-6-sol . The index is a useful comparison. It is not a complete measure of intelligence. 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. These figures estimate the cost of the benchmark tasks. They are not my subscription bill. The wait grows faster than the score. In my data, 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. OpenAI describes Max https://openai.com/index/previewing-gpt-5-6-sol/ as a setting for problems that need extended reasoning. I now treat it that way. I do not leave it on by default. Time to first token is the pause before a model starts its answer. Output speed tells me how quickly it writes after that. I combined both into 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 https://philippdubach.com/posts/kimi-k3-inside-claude-code/ . Cheap tokens do not help if I need more retries or the model never closes the task. Price per million tokens is an input price. I care about the price of finished work. OpenAI dominates my normal time budget. Anthropic still has the highest scores in this comparison. Opus 5 Xhigh scores 63 https://artificialanalysis.ai/models/claude-opus-5-xhigh . Fable 5 scores 62. Sol Max reaches 61. 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 the parts I notice every day. They do not measure 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. OpenAI says Fast mode can run GPT-5.6 up to 2.5 times faster https://openai.com/api-fast-mode/ . It also costs more. The $0.43 estimate for Sol High in my table uses standard benchmark pricing, not Fast mode pricing. I pay the premium because waiting breaks my concentration. If a model takes two minutes to begin, I switch windows. When it answers, I have to load the problem back into my head. The API bill counts tokens. It does not count that interruption. I would choose differently for a background job. Nobody cares if a batch finishes a little later overnight. For interactive work, I would rather pay more and stay in the problem. This is what I use now: