GPT-6.1 Reasoning Effort Test: Quality, Speed, and Cost Compared

GPT-6.1 Sol supports five reasoning.effort levels: low, medium, high, xhigh, and max. The default is medium. The model does not support none or minimal.

These levels are not “make the answer longer” settings. Higher effort generally allows the model to spend more work on planning, comparing alternatives, checking mistakes, and handling difficult sequences. It can also increase latency and token use. OpenAI’s guidance is to use higher levels only when evaluation shows that the quality improvement justifies the added cost and delay.

This article presents a reproducible test method and explains the typical role of each level. It does not present estimated numbers as measurements when no raw execution log is available.

What each reasoning level is for

| Effort | Good starting point for | Common misuse | | low | Drafts, routine analysis, execution-oriented coding, lower-risk tool work | High-risk questions that require deep tradeoffs | | medium | Most production tasks, research, agentic coding, and professional documents | Escalating to high without first measuring the baseline | | high | Complex debugging, deep planning, and high-value analysis | Simple extraction and formatting | | xhigh | Long-running research, security review, and difficult agent tasks | Assuming slower automatically means more accurate | | max | A small set of the most complex, highest-value tasks | Setting it as the global default |

OpenAI describes medium as a balanced point for quality, reliability, latency, and cost across most workloads. That makes it the right baseline: compare both downward and upward from there.

Test 1: Structured extraction

Prepare 50 short documents with inconsistent formatting and require output that matches a fixed JSON Schema. Score field accuracy, schema validity, and cost.

The key question is whether low already passes reliably. If it does, additional reasoning may add little business value. Structured extraction often benefits more from a precise schema, examples, and a validator than from maximum reasoning.

Test 2: Complex code repair

Select 20 real defects with reproducible failing tests. The model must identify the cause, make a minimal change, and run the relevant tests. Score test success, regressions, unnecessary changes, and review time.

This category is more likely to benefit from high because the model must reason across files, compare possible fixes, and check side effects. Still, measure whether medium reaches the same result. If success is equal, extra reasoning only adds cost and delay.

Test 3: Long-document decision analysis

Supply several policies, contracts, or research documents and require a conclusion, counterevidence, and itemized citations. Score factual accuracy, citation match, omissions, and unsupported inference.

Higher effort may help with conflict and ambiguity, but source selection and evidence requirements can matter more than the setting. Filling a million-token context with irrelevant documents does not automatically improve the analysis.

Test 4: Story and shot planning

Use one story passage and request scene objectives, character states, shot order, continuity constraints, and required assets. Have editors or directors evaluate the outputs without seeing the reasoning level.

Do not score creativity as “more detail.” Evaluate whether the structure serves the story, the shots are executable, continuity is correct, and revision time is reduced. Higher reasoning may produce more alternatives without improving the final choice.

If the plan later moves into ElserStudio for shot generation and editing, add production measures: How many changes does the shot list require? Did it omit a character, location, or prop? Can the generated material form a coherent sequence? This connects model evaluation with a real creative workflow.

How to measure speed and cost

Log the model, effort level, reasoning mode, input tokens, cached input, output tokens, total duration, time to first token, tool calls, retries, and final acceptance result for every run.

Do not report average latency alone. Production systems should also examine P50 and P95. Do not report only token cost per attempt; calculate cost per successful task.

Cost per successful task = cost of all attempts / number of accepted tasks

For work that requires human review, add editor, engineer, or subject-matter-expert revision time.

Standard and pro are not additional effort levels

GPT-6.1 Sol also supports standard and pro reasoning modes. Standard is the default. Pro performs more model work and is intended for difficult tasks that can tolerate greater latency and token use.

Reasoning mode and reasoning effort are independent. You can compare standard + medium with pro + medium, or compare medium and high within one mode. Change only one variable at a time if you want an interpretable experiment.

A practical rollout policy

1. Run the full evaluation set at medium to establish a baseline.
2. Run low-risk tasks at low; if quality does not materially fall, make it the default for that route.
3. Test high only for valuable tasks that fail at medium.
4. Enable xhigh or max only when blind evaluation and cost data both show a meaningful benefit.
5. Compare standard and pro separately on the hardest tasks; do not enable pro globally.

The bottom line

There is no universally best GPT-6.1 reasoning effort. Low favors efficiency, medium is the production baseline, high belongs on difficult high-value work, and xhigh or max must earn their place through evaluation.

A mature system should route effort based on task type, cost of failure, validation results, and latency budget rather than asking users to guess. More reasoning is not automatically a better business outcome. The best setting is the least expensive one that reliably passes acceptance.