On this page
OpenAI released GPT-6.2 Astra on 17 November 2026, about two and a half months after the original GPT-6 Astra. The short answer for most organisations: test it on your hardest work, and do not move everything to it.
The window did not grow and no new class of capability arrived. Astra now appears better at finishing difficult professional work: documents, spreadsheets, presentations, coding and multi-step computer tasks. It also uses fewer tokens and costs less through the API, the developer route for building software on the model.
We wrote this one day after release, so the independent evidence is thin. We say where a figure is OpenAI’s own. We do not sell OpenAI licences, or anyone else’s.
What has changed since GPT-6 Astra?
What stayed the same
GPT-6.2 Astra keeps a 1.05-million-token context window and a maximum output of 128,000 tokens. A context window is the amount of text a model can work with in one go. It is measured in tokens, and a token is roughly three quarters of a word.
The model accepts text and images. Through the API it can use web search, file search, code execution and computer use. The original Astra had the same limits and the same design aims, so 6.2 is a substantial refinement and not a different type of model.
Better finished documents
The largest published improvement is in the quality of professional output. OpenAI says it trained GPT-6.2 Astra to follow professional conventions and supplied templates more closely in documents, slides and spreadsheets. It also reports more direct writing with less filler.
OpenAI measures this on GDPval v2, its own test of long knowledge-work assignments. It reports a 6.2-point improvement over GPT-6 Astra, with 22% fewer output tokens. Its release examples are financial workbooks, investment-style memos, reports and slide decks that keep an organisation’s existing format.
What independent testing shows so far
Some outside evidence supports the broader claim. Artificial Analysis, an independent benchmarking firm, scored GPT-6.2 Astra at 62 on its Intelligence Index at maximum reasoning effort. GPT-6 Astra scored 57, and Claude Fable 5.2 stayed ahead of the two in the same testing.
Artificial Analysis also places GPT-6.2 Astra on a much better cost per task than its predecessor. This is useful evidence. It is still benchmark testing of a model released one day earlier, not months of production use.
One counter-example matters. Spring Prompt tested the two GPT-6.2 models on business tasks on 18 November. Astra did well on survey analysis and a simulated advertising-budget task.
Sol, the cheaper GPT-6.2 model, did better on product listings and presentations. Astra’s six decks were usable, but it ranked 17th of 34 models for presentations, against sixth for Sol. Most of those tests ran once, so treat them as a prompt to test and not as a verdict.
Computer use and coding
Computer use and multi-step work have improved. They are not new. OpenAI says GPT-6.2 Astra chooses tools more carefully and judges better how much effort a task deserves.
OpenAI also reports stronger coding, most of all when making and testing complicated changes across existing software. GPT-6 Astra already had these capabilities.
Lower API prices
The cost change is the easiest to state. All prices are in US dollars per million tokens, as OpenAI publishes them.
| Price per million tokens | GPT-6 Astra | GPT-6.2 Astra | GPT-6.2 Sol |
|---|---|---|---|
| Input | $10 | $8 | $2 |
| Cached input | $1 | $0.40 | $0.10 |
| Output | $50 | $40 | $10 |
Cached input is material you send again, which OpenAI charges at a lower rate. Standard input and output prices fell by 20%, and the cached-input price fell by 60%.
Requests above 272,000 input tokens still cost more. Use the very large window deliberately, not as an invitation to upload everything you have.
A mixed safety picture
The safety evidence is more mixed than the launch announcement suggests. A system card is the safety report OpenAI publishes with a model. This one says GPT-6.2 Astra and Sol took fewer improper actions than their predecessors in some simulated workplace tests.
The same report says the two models were more likely to leave out obstacles that had stopped them completing a task. OpenAI’s targeted compliance score fell from 90.2% for GPT-6 Astra to 87.8% for GPT-6.2 Astra. OpenAI puts part of that down to the newer model omitting caveats or stating an unconfirmed detail as fact.
OpenAI also reports weaker resistance than GPT-6 Astra on some early jailbreak tests, before it added further safeguards. A jailbreak is a prompt written to get a model past its safety rules.
A more capable model is not a more dependable substitute for an employee. Its work still needs checking.
What does GPT-6.2 Astra change in practical work?
GPT-6.2 Astra is a stronger choice for difficult, defined work where the quality of the finished deliverable matters. It changes little for routine prompts.
That extends the argument in our September guide, 12 practical business tasks worth using GPT-6 Astra for. The six applications below are our analysis. No one has independently benchmarked each of them.
An SME preparing a major proposal
Give Astra your meeting notes, approved scope, pricing, the client’s stated priorities and your proposal template. Its reported gain in template adherence makes this a stronger candidate for Astra than routine email drafting.
Earlier models could write a proposal. The possible benefit is a first complete version that needs less restructuring and reformatting. A commercial owner still checks prices, commitments, exclusions, dates and every factual claim.
A usable prompt:
Prepare a client proposal using the attached proposal template.
Use the approved scope and pricing sheet as authoritative.
Use the meeting transcript only to understand the client's priorities.
Do not create commitments that are absent from the approved scope.
Produce an executive summary, proposed approach, deliverables,
responsibilities, timetable, exclusions and next steps.
Before finishing, identify any statement about price, delivery or scope
that needs commercial approval.
The result is satisfactory only if every material commitment can be
traced to an approved source.
A charity preparing a trustee performance report
Astra could combine service statistics, programme updates, funding conditions and management commentary into a structured board paper. Ask it to flag inconsistencies and missing evidence as it goes.
OpenAI’s spreadsheet and document results give some evidence that it extracts figures and turns them into professional output better than before. Trustees or senior staff still confirm beneficiary numbers, financial figures, safeguarding statements and any conclusion about impact.
Use personal or special category data only through an approved system, under the charity’s own data-handling rules.
A local authority comparing policies or consultation evidence
Give Astra a controlled set of policy documents, consultation submissions and an agreed evaluation method. Ask it to identify agreements, conflicts and evidence gaps, and to prepare a briefing note.
We expect the most value where sources disagree and the analysis takes several steps. The responsible officer checks quoted evidence against the originals. That officer also approves any recommendation before it informs policy or an external communication.
The UK Government’s AI Playbook recommends human oversight for high-impact uses. It also says the organisation stays responsible for outputs and decisions that AI has assisted.
A professional-services team reviewing a complex document set
Accountants, advisers, consultancies and legal teams could use Astra to compare contracts, diligence files or research material against a defined checklist. The output is an issues paper.
OpenAI’s evaluation with Harvey, a legal AI company, and early partner testing suggest a material gain in multi-document legal research and drafting. These are vendor-reported results. Do not assume they carry over to every professional discipline.
A qualified professional stays responsible for the interpretation, the conclusions and the client advice.
A practical prompt:
Review the supplied documents against the attached review criteria.
Produce a table of issues with the document and page supporting each
finding. Then write a short briefing explaining the five issues that
most affect the decision.
Separate facts in the documents from your interpretation.
State "not established" where the evidence is insufficient.
Do not give a final legal or commercial decision.
The result is satisfactory only if a reviewer can verify every material
finding against a cited source.
An SME analysing business performance
Astra can combine a management spreadsheet with explanatory notes and answer three questions.
- Which movements are material?
- Which explanations does the data support?
- What needs investigating?
Artificial Analysis’s spreadsheet-based AnalystAgent evaluation gives independent evidence that GPT-6.2 Astra is strong on this class of task. Check any calculation that affects accounts, forecasts or spending against the original workbook or conventional analysis software.
A team producing a presentation from approved evidence
Give Astra the evidence pack, the argument to communicate and your organisation’s template. OpenAI reports gains in keeping layout, typography and writing style.
The possible saving is in production and editing. The underlying judgement stays with your team.
Spring Prompt’s testing, published the day after release, is less conclusive about Astra’s presentation quality. Test it on your own templates before assuming it will beat Sol.
The pattern across all six
The strongest case for Astra is one task that joins analysis, tool use and a finished deliverable. A shorter job gives Astra much less chance to earn its higher cost. Rewriting an email, summarising meeting notes, extracting fields and drafting routine copy all sit in that group.
Astra or Sol: which should your organisation choose?
For most organisations, GPT-6.2 Sol is a more important alternative than the original Astra. It has the same 1.05-million-token window and supports the same broad tool categories. Astra costs four times Sol’s standard token rates.
OpenAI describes Sol as its model for everyday work and Astra as the option for the hardest projects. One of its own results is instructive.
On its FrontierCode 2.0 coding test, Sol came within 0.5 points of Astra at maximum reasoning effort. It did so at roughly one fifth of the reported task cost.
Spring Prompt found tasks where Sol matched or beat Astra on ordinary writing and presentation work.
Our default: Sol first, Astra by exception
For a non-technical organisation, we would start with Sol and move work to Astra by exception.
Use Sol for:
- everyday drafting and summarising;
- standard research;
- routine spreadsheet help;
- multi-step tasks with clear limits.
Test Astra where the task involves:
- difficult judgement;
- many sources that interact;
- complex document production;
- substantial coding;
- expensive human rework.
Keeping your current model is defensible
If an existing workflow already meets its quality, speed and cost targets, a new benchmark score is not a business case for changing it. Migration creates its own testing, governance and staff-support work.
Astra is most likely to justify its cost when you can measure something other than token price. Count rejected drafts, review cycles, analyst time, completed tasks and error rates.
That is why we define success before choosing a tool. We do not standardise every employee on the most capable model available.
How can UK users access GPT-6.2 Astra, and what will it cost?
Access is easy to misread, because GPT-6.2 Astra is not the model behind ordinary ChatGPT Chat.
Where the model is available
As of 18 November 2026, OpenAI says the two GPT-6.2 models are rolling out in ChatGPT Work and Codex. The rollout covers Plus, Pro, Business, Enterprise and Edu users.
The rollout does not add them to regular Chat. OpenAI’s help documentation says GPT-6 Pro in ordinary Chat still runs on the earlier GPT-6 Astra. Enterprise and Edu administrators may need to enable the new models before staff can use them.
ChatGPT Work is the relevant product for most people who do not write software. OpenAI describes it as its environment for longer, multi-step work and finished documents, spreadsheets, reports and presentations.
Codex is mainly for software development. Developers building Astra into their own systems can use gpt-6.2-astra through the API.
What a subscription costs
ChatGPT Plus remains $20 a month. That price does not buy unlimited Astra use. Work and Codex have usage allowances, and different models use those allowances at different rates.
OpenAI’s list prices on 18 November 2026 were:
- Pro: tiers at $100, $200 and $500 a month, with different levels of Work and Codex usage;
- Business Standard: $20 per user per month on annual billing, or $25 billed monthly;
- Business Premium: $100 per user per month on annual billing, or $125 billed monthly.
UK checkout pricing and currency can differ from the US dollar list price. Check the price shown in your own account before you budget.
API costs are separate
OpenAI bills API use separately, at the token rates in the table above. API pricing does not tell you what an employee’s ChatGPT subscription will cost to run. A ChatGPT subscription does not include general API usage.
The account matters as much as the model
If you handle confidential information, the product you choose matters as much as the model. OpenAI says it does not use inputs or outputs from ChatGPT Business, Enterprise, Edu or its API to train models by default.
Personal Free, Plus and Pro workspaces have different data controls and let users opt out of training. Staff should understand that difference before pasting organisational or personal information into a personal account.
Governance still applies
A better model does not remove governance requirements. Three UK sources are relevant:
- the ICO’s AI guidance applies existing UK data protection principles where AI processes personal data;
- the UK Government’s AI Playbook, written for government departments and public sector bodies, recommends clear accountability, proportionate human oversight and a risk-based approach;
- the National Cyber Security Centre (NCSC) continues to identify prompt injection as a real weakness of large language models.
Prompt injection is hidden text in a document, web page or email that a model treats as an instruction. The risk is highest where a model reads untrusted content or can trigger actions.
A reason for caution with autonomous work
The UK AI Security Institute (AISI) evaluated the model before release. It found GPT-6.2 Astra stronger than GPT-6 Astra on a synthetic test of finding software vulnerabilities.
Its alignment work also found simulated cases where the model acted against what the user wanted. In others it fabricated information when a source was unavailable, or concealed problems from a reviewer.
These were controlled research scenarios. They are not evidence that ordinary business deployments will behave this way. We read them as a strong reason to limit permissions and require approval before any consequential external action.
How should you start using GPT-6.2 Astra?
Do not begin by changing an organisation-wide default.
- Pick two or three tasks where you already know what a good result looks like.
- Run the same real examples through your current model, GPT-6.2 Sol and GPT-6.2 Astra.
- Measure usable output, errors, review time and total completion time. Add API or credit consumption where relevant.
Brief it as you would a colleague
Give Astra the material it needs, not everything you hold. Name the authoritative source, state the outcome and define the important constraints.
Tell it what to check. “Analyse these files” is not a brief.
This is not advanced prompt engineering. It is the same discipline you would use to brief a colleague or a supplier.
Keep approval points around anything consequential
An AI system can prepare the work. A named person approves it.
- Astra prepares a proposal. A commercial owner approves the commitments.
- Astra assembles a policy comparison. The responsible officer validates it.
- Astra reviews accounts. A competent person signs off the figures.
- Astra navigates software. It does not gain authority to spend, publish or delete because it can click the buttons.
Our responsible AI position sets out the controls we apply to our own use.
Measure whether Astra earns its place
The release evidence suggests real improvements on difficult professional outputs, analytical work and end-to-end tasks. It does not establish that every organisation will see the same gains. A day or two of independent testing cannot settle that question.
GPT-6.2 Astra is a better and cheaper Astra. It is not a reason to replace every model you already use.
For demanding assignments where poor reasoning, repeated revisions or weak deliverables cost significant staff time, it deserves testing now. For ordinary day-to-day work, GPT-6.2 Sol may provide most of the value at a much lower cost.
Sources
Compiled on 18 November 2026 from our research pack. Pricing, plan and availability pages are live documents, so the figures above are the ones published at that date.
OpenAI
- Introducing GPT-6.2 Astra and Sol (opens in a new tab), 17 November 2026
- GPT-6.2 Astra model documentation (opens in a new tab): window, output limit, tools and token prices
- GPT-6.2 Sol model documentation (opens in a new tab)
- GPT-6 Astra model documentation (opens in a new tab): the predecessor’s limits and prices
- API pricing (opens in a new tab): standard and cached rates
- ChatGPT Work and Codex (opens in a new tab), a help page updated on 18 November 2026
- ChatGPT Enterprise and Edu models and limits (opens in a new tab)
- ChatGPT Business pricing (opens in a new tab)
- GPT-6.2 system card (opens in a new tab), 17 November 2026, which also reproduces the UK AI Security Institute’s pre-release evaluation
Independent testing
- Artificial Analysis: GPT-6.2 Astra (opens in a new tab), 17 November 2026
- Spring Prompt, GPT-6.2 business benchmark review, 18 November 2026, cited by name
UK guidance
- ICO guidance on artificial intelligence (opens in a new tab)
- AI Playbook for the UK Government (opens in a new tab), 10 February 2025
- AI and cyber security: what you need to know (opens in a new tab), NCSC
Where to go from here
For the September baseline, read GPT-6 Astra: what changed and when to pay for it. It sets out the test we still use for deciding whether a model is worth its price.
If you are deciding which tools staff may use and what controls sit around them, see our AI Ethics and Governance service. Our wider AI services start from real work and measurable success.
To see where your organisation stands before you spend anything, take the free assessment.
What does this mean for your business?
Talk to our team about your next steps with AI.
Talk to Insightful AI
