GPT-6.1 Sol arrives in Codex and ChatGPT Work. The big question is the cost of finished work.

A work model, not a new name for every ChatGPT conversation

OpenAI has announced GPT-6.1 Sol as its new workhorse for repeated and long-running tasks across code, applications and documents. The September 29 update positions it near GPT-6 Astra on complex work at a lower cost. That performance comparison is the company’s claim; Lumacta has not independently benchmarked the two models.

The distinction that matters before anyone buys or switches is the product surface. The launch documentation places Sol in Codex and ChatGPT Work, not the regular Chat mode. A model name, an application and a subscription are different things. Calling this simply “the new ChatGPT” would conceal the access conditions readers actually need.

Our reading is that this launch is less about producing a spectacular isolated answer than making capable assistance practical to use repeatedly. A business does not pay only for a polished demonstration. It pays for the ordinary jobs in between: revising a spreadsheet, checking a change, finding the relevant file and repairing work that did not pass review.

Source notes: 1, 2. Analysis and hypothetical examples are identified in the text.

Who gets it—and why a model can still be missing

At launch, OpenAI lists Plus, Pro, Business, Enterprise and Edu access, subject to rollout and client or workspace settings. Free and Go are not included. Enterprise and Edu administrators must enable the model because it starts disabled there. The documented clients include the Codex desktop app and CLI, and ChatGPT Work on web and mobile.

Standard and Fast are launch options; Ultrafast is described as coming later. These are service choices, not a promise that every account shows an identical menu immediately. A missing option should first prompt a check of the plan, application and workspace policy—not a conclusion that the announcement was false.

For a team, the practical rollout is to identify which users and workflows need the model before making it the default. A mixed workspace can otherwise create a confusing comparison: one colleague may be using a different model or speed setting while both call their tool “ChatGPT”. Recording the actual selection makes feedback useful.

Source notes: 2. Analysis and hypothetical examples are identified in the text.

A large context window is capacity, not guaranteed attention

The API identifier is gpt-6.1-sol. OpenAI lists a 1,050,000-token context window and a maximum output of 128,000 tokens. It accepts text and images and produces text, rather than providing native audio or video output. The documented knowledge cutoff is April 30, 2026.

Those specifications create possibilities, but they do not prove that the model will recover every detail in a large collection. Lumacta would treat the context window as a capacity limit, not a memory guarantee. A million-token allowance does not make a missing source, a misleading instruction or a contradictory document disappear.

An important document workflow should therefore include retrieval checks. Put the decisive clause or figure in an identifiable source, ask for its location and verify the answer against the original. If the task is about today’s events, the cutoff also matters: supply current evidence or use appropriately authorised retrieval instead of assuming that a larger model already knows the latest news.

Source notes: 3. Analysis and hypothetical examples are identified in the text.

The API rate difference is clear. The productivity difference is not.

For Standard API requests with no more than 272,000 input tokens, OpenAI lists Sol at $2 per million uncached input tokens and $10 per million output tokens. Astra is $10 and $50 respectively. Cached Sol input is $0.10 per million. These are token charges, not the monthly price of a ChatGPT subscription.

Here is an explicitly hypothetical comparison: 50,000 uncached input tokens and 5,000 billed output tokens cost $0.10 + $0.05 = $0.15 with Sol, versus $0.50 + $0.25 = $0.75 with Astra. That is an 80% lower token charge for the assumed identical usage. The output allowance includes reasoning tokens; it is not a count of words visible to the reader.

This example is not a measured task cost and excludes tools, cache writes and additional attempts. Long-context pricing also matters: above 272,000 input tokens, Sol’s listed rates become $4 input and $15 output per million for the whole request. Checking the billing tier belongs before declaring a long document job cheap.

Source notes: 4, 5. Analysis and hypothetical examples are identified in the text.

Do not turn an API saving into a subscription discount

OpenAI’s plan documentation separates API billing from ChatGPT usage. Signing in with ChatGPT uses the plan’s shared Work and Codex allowances; using an API key uses API charges. Included usage also depends on task length, complexity and settings. A token-price table cannot tell a subscriber how many substantial jobs remain.

Fast has another accounting distinction: OpenAI lists 2.5 times the included subscription usage, but twice the token-credit usage on purchased credits and Enterprise pay-as-you-go. Those multipliers describe consumption, not a guaranteed speed increase. They should not be used as a shortcut for predicting completion time.

Our recommendation is to compare a real week of accepted work, keeping the plan and speed selection visible. A user who values a faster response may reasonably choose a different trade-off from a team running many unattended jobs. Neither choice changes the need to verify the result before it becomes a public statement or an external action.

Source notes: 6. Analysis and hypothetical examples are identified in the text.

The useful denominator is one accepted result

OpenAI’s model-selection guidance recommends comparing Sol and Astra on the same tasks. Lumacta’s proposed version of that comparison starts with a written acceptance test: what must the result contain, which checks must pass and what would make it unusable? Decide this before looking at either answer.

Consider a hypothetical website repair. The goal is not “produce a patch”; it is “fix the reported problem without changing unrelated behaviour”. Both models should receive the same starting files, permissions and test instructions. Count all attempts, failed tests and human repair time, not only the final response that looks convincing.

For a document, the standard could include traceable quotations, correct totals and no unsupported claims. Reviewers should inspect the underlying references and calculations. Fluent prose may make review pleasant, but it is not a substitute for passing those checks. A cheaper draft that needs extensive correction can be more expensive to finish.

Source notes: 7. Analysis and hypothetical examples are identified in the text.

A fivefold rate gap does not settle the purchasing decision

The arithmetic above creates a useful thought experiment, not a performance result. With that fixed token mix, five Sol attempts would have the same token charge as one Astra attempt. If only one of the five is usable, the apparent saving has disappeared before anyone counts the cost of reviewing the rejected work.

Conversely, if Sol reliably passes a bounded task on the first attempt, paying Astra’s higher rate for that same job may bring little practical benefit. We do not have evidence to assign either success rate. The correct conclusion is that the price difference creates room for evaluation, not that it identifies a universal winner.

A complete comparison would add human labour, tool charges and the consequence of an error. A low-stakes internal draft and a production migration need not have the same approval process. Model selection becomes more useful when the organisation classifies the task and its consequences rather than choosing one setting for everything.

Source notes: 4, 7. Analysis and hypothetical examples are identified in the text.

Tool support is not permission to act

OpenAI documents tool-enabled work through the Responses API; Chat Completions is tool-less for this model. It also lists medium as the default reasoning effort and does not support none or minimal. These integration details matter when evaluating an existing application, rather than merely comparing names in a dropdown.

Lumacta’s engineering advice is to start with the smallest permissions needed for a bounded test. Reading a repository, modifying local files and deploying those files are separate capabilities. A tool-capable model should not inherit approval to publish, purchase or contact someone merely because it can prepare the necessary material.

The same principle applies to business documents. Access to a folder does not establish permission to upload its contents elsewhere. Keep the boundary explicit, preserve an audit trail and require a human decision before consequential external changes. A lower inference bill is useful; it is not a reason to weaken the controls around the work.

Source notes: 3, 5. Analysis and hypothetical examples are identified in the text.

“Near Astra” needs an operational definition

Our scientific perspective is that a performance claim becomes informative when someone can specify how it might fail. “Near Astra” could mean similar correctness, completion rate, time or usefulness on a particular task set. Those are different outcomes. Without the test conditions, the phrase is a positioning claim rather than a result a reader can reproduce.

A fair local evaluation would use representative jobs, repeated runs and the same acceptance criteria. Where practical, assess outputs without telling the reviewer which model produced them. Record failures as well as successes, and keep the tools, input material and reasoning settings consistent. These are proposed conditions; Lumacta has not run this experiment.

The strongest case for Sol will be evidence that it preserves useful outcomes while reducing total cost across ordinary work, not a single impressive answer. Astra may still justify its premium on a difficult task; Luna may be sufficient for a narrow one. The question is which option clears the same quality bar with the least overall burden.

Source notes: 1, 7. Analysis and hypothetical examples are identified in the text.

A meaningful launch—provided the evaluation follows the price sheet

Lumacta’s editorial judgment is that Sol is worth a serious, bounded comparison for readers already doing substantial work with AI. Its launch combines a clear lower API rate than Astra with a company claim of comparable usefulness on complex work. The first is documented arithmetic; the second still needs testing in the setting that matters to the reader.

Start with a task whose outcome you can check, retain the previous workflow as a baseline and track what it takes to finish. That approach can reveal a genuine saving without pretending a release announcement is an independent review. The important milestone is not the first answer. It is the first result you can confidently use.

Source notes: 1, 4, 7. Analysis and hypothetical examples are identified in the text.