Why I moved my Codex default to GPT-6.1 Sol High
I had just moved my Codex main session to GPT-6 Sol at medium. Then I looked more closely at the GPT-6.1 Sol results across reasoning effort levels and changed the default again: the main session is now GPT-6.1 Sol at high. That is a configuration choice informed by a few public benchmark curves, not a claim that high is best for every person, task, or budget.
I started with the coding curve. These are the five published GPT-6.1 Sol points on DeepSWE:
| Reasoning effort | Score | Average task cost |
|---|---|---|
| Low | 64.4% | $0.17 |
| Medium | 73.0% | $0.42 |
| High | 75.2% | $0.65 |
| Xhigh | 71.9% | $0.79 |
| Max | 71.9% | $1.57 |
Medium to high buys 2.2 percentage points for roughly 55% more average task cost in this evaluation. For a main session that plans coding work and reviews the results, that measured trade felt reasonable as a default to try. I have not run these evaluations myself.
The gain depends on the benchmark
I checked two other curves before deciding. High also improves over medium there; OSWorld's best listed point is max. I want a default that reflects my coding work, while leaving room to make a different choice for a specific task.
| Benchmark | Medium score | Medium cost | High score | High cost |
|---|---|---|---|---|
| DeepSWE | 73.0% | $0.42 | 75.2% | $0.65 |
| GDP.pdf | 30.0% | $0.34 | 32.0% | $0.35 |
| OSWorld 2.0 offline | 66.8% | $0.77 | 69.6% | $0.96 |
Scores and average task costs are transcribed from the GPT-6.1 Sol release's coding, professional work, and computer-use benchmark charts. The model documentation lists medium as the default and supports low, medium, high, xhigh, and max effort. These evaluation costs are not Codex subscription credits, a prediction of my bill, or a measure of actual task latency.
Aligning the setting and the policy
I noticed a small mismatch while making the change. My main configuration already said GPT-6.1 Sol at high after I selected the model and effort in the picker. My global AGENTS.md still described GPT-6 Sol at medium as the main-session default. A person or agent reading the policy could therefore reason from a different setup than the one Codex would actually launch.
I aligned the two. The configuration now records model="gpt-6.1-sol" and model_reasoning_effort="high", while the global instructions describe that same main default. The configuration selects the model and effort; AGENTS.md describes how I want work routed. The latter is guidance for people and agents, not an automatic optimizer that measures a task and swaps models on its own.
| Role | Model and effort | Typical work |
|---|---|---|
| Main session | GPT-6.1 Sol, high | Planning, routing, synthesis, and general work. |
| Generic implementation agent | GPT-6.1 Sol, medium or high | Multi-file implementation, selected by task difficulty. |
| deep-reasoner | GPT-6 Astra, high | Hard design, unclear debugging, algorithms, or high-risk review. |
| fast-worker | GPT-6 Luna, high | Bounded implementation, tests, and cleanup; low or medium for mechanical stages. |
The routing rules matter as much as the default. Generic implementation subagents now use GPT-6.1 Sol at medium for ordinary multi-file work, or high when the task calls for more care. Astra at high stays reserved for difficult design decisions, bugs whose cause is unclear, and high-risk reviews. Luna at high handles tightly scoped implementation, tests, and cleanup; a mechanical delegated stage can be explicitly set to low or medium. An obvious typo or rename does not need an agent at all.
I keep xhigh exceptional for a hard task where I want one model to spend more time reasoning. Max is not part of this setup. I also do not delegate tiny tasks just because there is a worker role available. The practical change is to align the main default with my selected setting and keep the model and effort explicit for each delegated job.
A default to learn from
High won this choice because it fits the main-session work I expect and led the DeepSWE curve I checked. I will watch how it feels on my own coding tasks and lower effort for a particular delegation when its scope is mechanical or tightly bounded.