GPT-6.1 Sol High vs. Medium: Reasoning, Speed, Usage and How to Choose
GPT-6.1 Sol High and Medium are two reasoning-effort settings for the same model, not separate models. Medium balances speed and depth, while High applies more reasoning to difficult work; neither label guarantees that High is always more accurate or slower by a fixed multiple.
What changes between GPT-6.1 Sol High and Medium?
Reasoning effort guides how deeply the model works through a request. The official GPT-6.1 Sol model page lists low, medium, high, xhigh and max under the same model ID, gpt-6.1-sol. Medium is the default when effort is omitted in the application programming interface (API). An app’s selected default can differ by account and client, so check it separately.
Think of asking the same worker to devote more effort to reviewing a problem. The setting does not add a separate knowledge database or switch to another model. For the model’s broader capabilities and comparisons with other models, see the introduction to GPT-6.1 Sol.
Effort is also different from answer length. You can ask High for a short conclusion and Medium for a long explanation. Visible length is not a measure of internal reasoning. Keep effort, speed settings such as Fast, and the API’s standard or pro reasoning mode separate when comparing results.
High and Medium compared at a glance
This table describes selection tradeoffs, not benchmark results. The official reasoning guide explains that reasoning also adapts to task complexity. A setting therefore does not fix the exact processing time or token count.
| Criterion | Medium | High |
|---|---|---|
| Underlying model | GPT-6.1 Sol | The same GPT-6.1 Sol |
| Reasoning emphasis | Balance of speed and depth | Deeper work on difficult problems |
| Response time | Relatively faster starting point | May involve additional waiting |
| Usage | Balanced starting point | May use more reasoning tokens |
| Candidate tasks | Clear drafts, organization and edits | Complex bugs, conflicts and planning |
| Not guaranteed | Identical speed or accuracy every time | Always better results or a fixed multiplier |
Judge the finished work. A fast answer can still create a long task if a person must repair many omissions. Conversely, a longer High run may add little value if the result barely improves. Include human review and correction time alongside generation time.
When Medium is a useful starting point
Start with Medium when requirements and acceptance criteria are clear and you can promptly inspect the result. Examples include an announcement that preserves supplied facts, extracting decisions, owners and dates from meeting notes, or a tightly scoped wording edit. These are recommendations based on task structure, not performance guarantees.
For code, a change with a known location and well-defined expected behavior can also be tried at Medium first. Supply relevant files, error messages and reproduction conditions, and restrict the scope to reduce guessing. Inspect the changed files and actual tests rather than relying on a confident completion report.
Iterative work need not place every decision in one response. You might draft at Medium and use High for a remaining difficult issue. For a controlled comparison, however, use separate sessions so that the first answer does not supply clues to the second attempt.
When to consider High
High is worth evaluating when multiple steps interact or requirements conflict. Examples include tracing a bug across several files, planning a migration with compatibility and recovery constraints, or reconciling conflicting source documents. Intermediate decisions can strongly affect the final outcome in these tasks.
Even within blogging, polishing a paragraph differs from comparing several policies and explaining their exceptions. The latter requires checking applicability and contradictions, making additional reasoning worth testing. Supplying the sources and their dates is more valuable than increasing effort while leaving the evidence absent.
Give High a specific objective. Instead of “make it better,” ask it to inspect dependencies, failure conditions and recovery steps, and identify unresolved questions. Define how to verify the result so you can tell whether extra deliberation actually improved the work.
Does High always produce a better answer?
No. On short, clear tasks, the results may be similar, and individual answers can vary even at the same setting. Do not confuse preferred style or greater length with improved accuracy. Check required facts, omissions, format and constraints.
More reasoning does not automatically repair missing information or a false premise. An invented product feature or outdated price can still lead to a detailed but incorrect explanation. Verify current information against primary material or authorized search tools, and ask the model to identify what it cannot establish.
High does not automatically expand tool permissions or guarantee that a command ran or a file was saved. Check tests for code, figures and sources for documents, and saved records for app work. When failures repeat, distinguish missing evidence, access problems and unclear requirements before changing effort.
How should speed, usage and cost be compared?
Tokens are units of information processed by the model. Distinguish input, visible output and internal reasoning tokens. API reasoning tokens are billed as output even though they are not displayed as ordinary answer text; their count is available in usage under output_tokens_details.reasoning_tokens. A short visible response is therefore not necessarily a cheap request.
Under the same model and processing conditions, High is not a separate model with its own higher token rate. Request cost depends on consumed input and output plus applicable cache, tool and processing-mode terms. More reasoning tokens can increase the bill, but there is no universal “twice as expensive” rule. Match actual usage with the pricing conditions in effect when calculating costs.
Do not convert Work or Codex subscription usage or credits directly into API dollar charges. Check the account’s own terms and usage display. The official Work and Codex model guide also notes differences by plan, client, workspace and rollout. Waiting time can vary with input length, tool execution and service conditions even at an unchanged effort.
How to compare High and Medium on the same task
Choose representative tasks and evaluation rules before repeatedly making potentially billable API calls. The procedure below is a way to conduct your own comparison, not a report of measured performance.
- Prepare a task and acceptance criteria. List source facts, required content, prohibited changes and output format. Assess facts, omissions and voice for writing; reproduction, change scope and existing behavior for code.
- Copy the originals and match conditions. Use the same materials, prompt, tools and permissions. For code, start from separate copies of the same project state. Keep the experiment away from production files.
- Run with Medium explicitly selected. In Work or Codex, check the model and reasoning control beneath the composer and open Advanced if needed. In the command-line interface (CLI), use /model; in the API, set reasoning.effort to medium for the same model ID. Do not assume an unavailable option exists.
- Change only effort to High. Use a separate session and fresh original copy with the identical request. In the API, change effort to high while preserving everything else. Changing speed or pro mode simultaneously prevents an effort-only comparison.
- Score results and total time. Record submission-to-completion time, factual errors, omissions, constraint failures and human review or repair time. Run tests and inspect diffs for code. Record actual API usage where applicable.
- Repeat representative tasks and restore settings. Avoid conclusions from one lucky answer. Choose the effort that meets each task’s criteria, and restore changed settings or experiment copies as appropriate.
For writing, try this fictional brief: “The event starts at 10:00 and ends at 12:00. It is free, but booking is required. No venue was supplied.” Give both settings the same request: “Write a two-paragraph beginner-friendly announcement. Keep the times, include both free entry and required booking, and mark the venue as unspecified rather than inventing it.” The acceptance criteria are those four requirements. A plausible but invented venue is an error, however polished the prose.
For a code comparison, choose a small project you can safely run. Reproduce a bug in an existing sorting function, then request that duplicate items retain their order, empty input works and the public function name stays unchanged. Supply the relevant complete files and reproduction inputs. Check previously passing tests plus duplicate and empty-input cases. Define expected output first because actual ordering rules depend on the project.
The official model-selection guidance likewise recommends fixed inputs, predefined criteria and accounting for review and repair. That supports a decision based on reduced rework rather than a longer answer.
Which setting should you choose?
If Medium meets requirements with little correction, keeping it is reasonable. When it repeatedly misses constraints or struggles with complex relationships despite adequate evidence, compare High on the same tasks. Weigh fewer errors against added waiting and usage.
If both settings omit the same information, improve the sources and instructions first. Specify audience, deliverable, constraints and verification, and prohibit guesses when evidence is missing. More reasoning cannot supply an inaccessible tool or nonexistent file.
You can use Medium for quick drafts and High for difficult review instead of fixing one effort for every job. Recheck familiar tasks when models or services change. Let verified results and total work cost guide the choice, rather than the apparent rank of the setting’s name.
