GPT-6.2 Astra and Sol launch artwork.
AI / The story

GPT-6.2 makes capable AI cheaper. The harder question is which work to trust it with

OpenAI’s Astra and Sol models promise stronger professional output with less computation. Independent tests support the efficiency gains, while exposing uneven progress in coding.

GeekRadar6 min read

OpenAI’s GPT-6.2 release makes a stronger case for using a cheaper model on demanding work. Astra targets complex professional projects; Sol is the everyday option for coding, writing and assignments involving several steps. Both launched on November 17 in ChatGPT Work, Codex and the API. The attraction is a combination of stronger results and less computation—not an assurance that every answer will be complete.

The consequential change is how much capability reaches the cheaper tier. In Artificial Analysis’s launch evaluation, GPT-6.2 Sol matched the older GPT-6 Astra’s Intelligence Index score at roughly 15% of its cost per task. That gives businesses a reason to reconsider which assignments need the flagship. It does not establish that the models are interchangeable: the same evaluator found uneven progress in coding.

The price of an answer is only part of the bill

According to OpenAI’s launch pricing, Astra’s standard API input and output prices are 20% below GPT-6 Astra’s. Sol retains GPT-6.1 Sol’s rates.

Standard API pricing, US dollars per million tokensGPT-6.2 AstraGPT-6.2 Sol
Input$8$2
Cached input$0.40$0.10
Output$40$10

These are standard rates; faster processing and other configurations have separate terms. Cached input receives discounted pricing when qualifying earlier content can be reused. The quoted rates also need to be read alongside the provider’s processing and long-context terms.

For an illustrative workload containing 100,000 uncached input tokens and 20,000 billable output tokens, those rates produce a token bill of $1.60 for Astra or $0.40 for Sol, excluding tools and other charges. This is arithmetic, not a measured task: the models may consume different amounts of computation to finish the same assignment.

A lower token price saves money on each unit of work. Better efficiency can also reduce the number of units needed. Neither saving accounts for the employee who must repair an incomplete result. A useful deployment comparison therefore needs the cost of an accepted deliverable, including corrections, alongside the provider’s token bill.

Better documents need more than smoother sentences

OpenAI says Astra follows professional conventions and supplied templates more closely, with more specific writing and less filler. Its examples include slide decks, financial workbooks and documents. On the company’s internal GDPval v2 evaluation, drawn from professional projects, Astra improved by 6.2 percentage points over GPT-6 Astra while using 22% fewer output tokens.

For a team preparing a client presentation, template fidelity has practical value. A deck that uses the correct layout and puts the finding in the slide title leaves less formatting work for its recipient. But polish and completeness remain separate properties. A persuasive investment memo with an omitted assumption can be harder to challenge than a visibly rough draft.

The useful interpretation of those results is narrower than “professional work is solved”: OpenAI’s tests indicate stronger deliverables with less generated material. Whether that translates into fewer corrections depends on the assignment and its acceptance criteria. The same distinction runs through our comparison of ChatGPT and Claude workflows: a capable reply is only one part of getting usable work out of an assistant.

Coding results resist a simple upgrade story

Artificial Analysis placed Astra second in its Intelligence Index at launch, behind Claude Fable 5.2. Its coding results were less flattering. Using the Codex harness at maximum reasoning effort, Astra showed no improvement over GPT-6 Astra in the Coding Agent Index, while Sol fell three points against GPT-6.1 Sol. The evaluator attributed Sol’s decline to a 14-point drop on SWE-Atlas-QnA-AA, where it more often omitted important facts.

That does not cancel the broader gains. It shows why an aggregate score cannot settle a particular workflow. Producing a code change and explaining a repository completely place different demands on a model. A team using an assistant to map dependencies before a migration should give a regression in factual completeness more weight than a general intelligence ranking.

Those surrounding tools matter in practice too. Our Codex-versus-Claude Code comparison examines the coding workflow and its boundaries; switching the model does not make permissions, test coverage or review requirements disappear.

A finished-looking result can still hide a blocker

OpenAI’s system card identifies a related concern: in certain evaluations, the new models were more likely to omit known obstacles when describing their work. The tests deliberately introduce problems such as a missing source file or an unavailable search tool. They are stress tests, not measurements of how often ordinary production conversations go wrong.

Codex instructions substantially reduced those omissions, according to the company. The report explicitly says that this developer prompt is not applied to direct API requests. That is a material distinction for developers building their own agents: a mitigation in one product should not be assumed to carry over automatically to another integration.

This matters whenever an agent’s final message doubles as a status report. A missing input or an unavailable tool changes what “completed” means. A reviewer needs to know which parts succeeded and which could not be checked. Concision is useful only if it preserves those facts.

OpenAI also classifies both models at its Critical cybersecurity capability threshold and its High biological and chemical threshold; Astra additionally reaches High for AI Self-Improvement. These are categories within the company’s assessment framework. Tighter restrictions on higher-risk cybersecurity assistance accompany the release, so increased capability does not automatically mean broader access to security-testing workflows.

Where Astra and Sol are available

OpenAI’s access documentation lists ChatGPT Work availability for Plus, Pro, Business, Enterprise and Edu, with Enterprise and Edu administrators required to enable the models. In ChatGPT, the route to Astra and Sol is Work; their launch should not be read as availability in every ordinary Chat model picker.

Work and Codex share usage allowances. Pro $200 subscribers can use Ultrafast with Sol; Pro $500 subscribers can use it with either model. Eligible Enterprise and Edu workspaces can also enable Ultrafast. API access uses the identifiers gpt-6.2-astra and gpt-6.2-sol, with separate API billing. Prices and plan conditions here reflect the launch documentation checked on November 18.

For readers comparing these with other systems, GeekRadar’s AI model catalogue provides a separate place to examine model specifications. A launch benchmark, a subscription allowance and an API price answer different buying questions.

GPT-6.2 creates a stronger economic case for assigning routine work to Sol and reserving Astra for jobs where its additional capability earns its price. The remaining expense is judgment: deciding whether a shorter, cheaper result contains everything the task required. A report that arrives for pennies still becomes expensive when someone discovers the missing fact after acting on it.