Opinion

GPT-6.1 Sol: watch the cost of finishing, not just generating

OpenAI's new Sol release makes capability per dollar the headline. The important comparison includes retries and review.

By Saud AlotaibiPublished 3 October 20263 min readEvent 29 September 2026Sources checked 3 October 2026
The takeaway

Measure the price of an accepted result, including retries, tool use, and human review.

The most interesting model launch is not always the one with the largest score. Sometimes it is the one that makes capable work practical at a different cost. That is the case OpenAI makes for GPT-6.1 Sol.

The announcement positions it as an upgrade to GPT-6 Sol that approaches Astra on several demanding tasks. Its published standard API rates are $2 per million input tokens, $0.10 for cached input, and $10 per million output tokens. The product page distinguishes availability in Work and Codex from availability in ordinary Chat. Those details were checked on 3 October; they should be rechecked before an implementation or purchase.

A cheaper answer can still be expensive

The token bill is visible. The failed attempt is less visible. If a model needs three revisions to produce an acceptable result, its apparent price advantage may shrink. If a person spends twenty minutes correcting an output, the workflow has another cost that the API invoice cannot show.

This is why I prefer a comparison built around a finished task. Define the result first, then count what it took to obtain it. For an extraction job, that may mean a valid record whose important fields match the document. For code, it may mean a change that passes the appropriate checks and survives review.

My view: the workhorse tier deserves attention

A dependable everyday model can have a larger practical effect than a flagship used occasionally. Most projects contain a mixture of straightforward steps, moderately difficult judgment, and a smaller number of exceptional cases.

A sensible routing design can keep the easy cases simple and reserve a stronger model for cases that actually need it. The route should be driven by observed failures or explicit task requirements, not a vague belief that a more expensive name must always be better.

The published results are OpenAI's evaluation, not a benchmark I ran. They give a reason to test the candidate. They do not establish that the same advantage will appear in every repository, language, document, or tool environment.

The comparison I would build

Keep a representative set of tasks and a clear acceptance rule. Record completion quality, total tokens, elapsed time, retries, and review effort. Use the same evidence and tool permissions for each candidate.

For a workflow that reuses context, track what was actually eligible for caching. Do not assume every input token receives the cached rate. For a task with tools, include the effect of tool failures and repeated calls.

When a switch is justified

A migration becomes worth considering when the model meets the quality bar and improves the complete workflow. Preserve a rollback path and record the version and settings used in the comparison.

The useful conclusion is conditional: Sol deserves a place in a task-specific evaluation. A published price and a strong launch claim are the beginning of that evaluation, not the end.

Sources & further reading