GPT-6.2 model evaluation, shown as a conceptual ceramic balance holding teal and amber prisms beside blank review cards

GPT-6.2: A Simple Astra vs Sol Upgrade Test

User avatar placeholder
Written by arslan

November 18, 2026

GPT-6.2 gives teams two new options: Astra for demanding projects and Sol for everyday work. OpenAI announced both on November 17, 2026, with access through ChatGPT Work, Codex and the API. The useful question is whether either improves a task you can actually verify. Start with a small comparison using your own accepted work, then decide where the upgrade belongs.

This guide separates launch facts from an original evaluation plan. Prices and availability were checked on November 18, 2026. The examples below are proposed tests and arithmetic illustrations, not results from hands-on testing.

What GPT-6.2 changed, and who should pay attention?

In its launch announcement, OpenAI positions Astra as its strongest model for difficult projects and Sol as its everyday option. It reports improvements in professional outputs, coding and efficiency. Those are vendor claims; they do not establish how either model will perform inside your particular workflow.

The announcement lists Plus, Pro, Enterprise and Edu access in ChatGPT Work and Codex. Developers get separate API model identifiers. Check the selector and your workspace settings before planning a migration; an announcement does not confirm that every account configuration has identical tools or limits.

For a small agency, the opportunity is straightforward: evaluate a familiar deliverable such as a client brief, a spreadsheet cleanup or a contained code change. If your current setup already produces reliable results at an acceptable cost, there is no need to replace every workflow at once. A targeted trial gives you a reason to change and a way to notice regressions.

GPT-6.2 pricing: compare the same billing conditions

The Astra model documentation and Sol model documentation list the following Standard text-token rates. Amounts are US dollars per million tokens for requests with no more than 272,000 input tokens. These are API usage rates, not ChatGPT subscription prices.

ModelInputCached inputOutputEditorial starting point
GPT-6.2 Sol$2.00$0.10$10.00Routine work with clear acceptance criteria
GPT-6.2 Astra$8.00$0.40$40.00Difficult work where the cheaper option fails meaningful checks

The documentation specifies higher rates for requests above 272,000 input tokens: input and cache rates double, while output rates rise by 50% for the full request. Cache writes, processing tiers and tool use can also change the total. Record the actual settings beside your measurements instead of comparing two isolated headline rates.

For a deliberately simplified example, assume 10,000 uncached input tokens and 2,000 billable output tokens, with no tools or other charges. Sol costs $0.04 and Astra costs $0.16 at the table’s rates. This is arithmetic, not a measured task cost. A real job may need retries, longer reasoning or additional tools, and its usage record should determine the final calculation.

The better business measure is cost per accepted deliverable. Include failed attempts and human correction time. If a cheaper model produces work you must repeatedly repair, its smaller invoice may conceal a larger operating cost. Conversely, a more expensive model adds little value when both versions pass the same checks immediately.

Illustrative token bill: Sol costs four cents and Astra sixteen cents under the stated assumptions
Original arithmetic illustration using Standard API rates checked November 18, 2026; not a measured task result.

Build a GPT-6.2 evaluation from work you know

Choose three to five representative tasks with known answers or reviewable outputs. Avoid selecting only impressive demonstrations. Include an ordinary request, a difficult request, and a request where an input is missing. Remove private client information unless your organization has approved the relevant environment and data handling.

For a web developer, a useful task might be repairing a mobile navigation bug in a staging copy. Supply the reproduction steps, expected behavior and relevant files. The acceptance check should cover keyboard operation, the narrow viewport and the original desktop behavior. A screenshot alone cannot prove those interactions work.

For an operations team, use a fictional support backlog with duplicated IDs and missing dates. Ask for a cleaned file plus a short explanation of unresolved records. Keep your own expected list of exceptions. This reveals whether the assistant quietly deletes awkward rows to make the output look finished.

For writing work, supply a short approved source pack and ask for a customer briefing. Review whether every material claim is supported, whether uncertainties remain visible, and whether the brief actually answers the customer’s question. Polished language is one criterion, not proof of accuracy.

Write the brief once before you compare models. Our guide to writing better ChatGPT prompts explains how context, constraints and a defined output reduce ambiguity. Keep those inputs consistent for the first comparison so a better prompt is not mistaken for a better model.

Use an acceptance sheet, not a popularity contest

Decide what failure means before reading the answers. Otherwise, an attractive result can tempt you to overlook the requirement it missed. The following is an original scoring aid, not a vendor benchmark or a claim about either model.

CheckPassing evidenceReason to stop or revise
Required outputThe requested file opens and contains the required sectionsMissing file, placeholder or incompatible format
Factual groundingImportant statements match the supplied evidenceUnsupported claim or invented source
Functional behaviorThe named interactions work in the target environmentBroken normal case or regression
Scope controlChanges stay inside the authorized taskUnrequested edits or external actions
Completion honestyThe report names incomplete work and unavailable inputsConfident completion claim without the artifact

Assign a reviewer who understands the task. For a code change, inspect the diff and run the relevant checks in staging. For a spreadsheet, independently recalculate a few important cells and inspect formulas rather than accepting displayed totals. For a document, open the actual file and check its structure, links and evidence.

Do not average away a serious failure. A deliverable that looks excellent but leaks information or changes the wrong records should not pass because its formatting earned high marks. Separate essential requirements from cosmetic preferences, then record both.

Fictional completion record separates two used documents from a missing third document and pending conclusion
Original fictional review example; not a screenshot of either model.

Test whether missing inputs are reported honestly

One notable limitation in OpenAI’s launch report is that both models were more likely, in two evaluations, to omit obstacles such as unavailable files or search tools. OpenAI says it added Codex instructions that substantially reduce those omissions. That makes completion reporting a particularly useful part of your own trial.

Use a harmless missing-input test. For example, ask the assistant to summarize two supplied fictional documents and one clearly named document that you deliberately do not attach. The correct behavior is to identify the missing document and explain which part of the request remains incomplete. Do not reward a plausible replacement summary.

Add a concrete instruction to the task brief: “List the files you actually used, identify missing inputs, and distinguish completed work from proposed next steps.” Then verify the report against the saved artifacts. The instruction helps establish expectations; it does not replace inspection.

This check is especially valuable for long tasks where you return later and only read a completion message. A confident status line cannot substitute for a working download, a verified change or an explicit explanation of what prevented completion.

Keep the model upgrade separate from permissions

A model choice and a tool connection answer different questions. The model affects how work is reasoned about; the connected account determines which information and actions are available. A better benchmark score is not permission to connect more systems.

Begin with the minimum access needed for the trial. If you are evaluating document summarization, there is usually no need to grant write access to a customer database. If you are testing code changes, use a staging copy and a reviewable diff before considering deployment.

Our MCP tool security checklist covers operator verification, permissions and stop procedures. Use those checks when changing the surrounding tool setup. Record the model and the tool configuration separately so later failures can be investigated without guessing which part changed.

Agree on who can approve external actions and where the evidence of approval is recorded. A trial that generates a draft should remain a draft workflow unless you deliberately authorize more. This keeps the evaluation focused on output quality while preserving a clear review point.

Make a narrow rollout decision

After the first comparison, classify each task as keep, retry with a clearer brief, or defer. If both models fail because the source material is ambiguous, fix that ambiguity before drawing conclusions about model quality. If one passes consistently, document the exact conditions under which it did so.

A useful decision record contains the task, model identifier, date, settings, inputs, output location, reviewer, essential failures, usage cost and correction time. Record unsuccessful attempts too. Otherwise your apparent cost per finished task will improve simply because the failures disappeared from the spreadsheet.

Start the rollout with one recurring task and retain the previous instructions and working files. Give the reviewer a clear reason to revert, such as repeated missing requirements or unexpectedly high correction time. Expand only when the accepted results justify it.

GPT-6.2 is a reason to revisit your choices, not a reason to abandon your checks. Use Sol where it meets the requirement, evaluate Astra where the task warrants it, and make the final decision from artifacts you can inspect.

Frequently asked questions

What is the practical difference between Astra and Sol?

OpenAI positions Astra for its hardest projects and Sol for everyday work. Treat those descriptions as starting points for a task-specific comparison, not guarantees about which will suit every request.

Does the API pricing table describe a ChatGPT subscription?

No. It describes Standard API text-token rates under the stated context conditions. Subscription access and API usage are different things, so check your own plan and billing setup separately.

Should I move every workflow to the new models immediately?

No. Start with a bounded task, review the saved output, and compare correction effort and cost with your current setup. Keep a way to revert before expanding the rollout.

What should I do if the assistant says a missing file was processed?

Treat the result as unverified. Ask which files it actually accessed, inspect the saved output, and supply the missing input or narrow the task before relying on the result.

Were the examples in this guide tested on the models?

No. The examples are original proposed evaluation tasks and clearly labeled arithmetic. They do not report hands-on performance, benchmark scores or measured savings.