Analysis

GPT-6.2 Astra and Sol: a 20% lower Astra price, where both run and what the new safeguards block

1AYM is an OpenAI Select Partner.

Published . Vendor facts checked on against each vendor’s own pages, listed at the end. Plans, prices and availability change. Neither OpenAI nor Anthropic has reviewed this analysis.

The short version

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November 2026, and GitHub added both to Copilot the same day. Astra's API price falls to $8 input and $40 output per million tokens, and Sol stays at $2 and $10.

The models and the price
Astra costs $8 input and $40 output per million tokens, down from $10 and $50 for GPT-6 Astra. Sol stays at $2 and $10.
Where it runs
Out on 17 November in the API, ChatGPT Work and Codex, and GitHub Copilot. Enterprise and Edu workspaces need an admin to enable them, and Copilot Business and Enterprise switch them on by default.
Safeguards and blockers
Tighter cyber safeguards restrict some legitimate security testing, and OpenAI found the models more often fail to mention an obstacle that blocked their work.

GPT-6.2 Astra and Sol: Astra's API price falls 20%, and Sol's stays the same

What was announced

OpenAI released two models on 17 November 2026. It pitches GPT-6.2 Astra at the hardest projects and GPT-6.2 Sol at everyday work. OpenAI says both give better results on demanding tasks while writing fewer output tokens.

A token is the small unit of text the API bills by. Astra's standard API price is $8 per million input tokens and $40 per million output tokens. Cached input, text the API has seen recently such as instructions sent with every request, costs $0.40 per million tokens.

OpenAI says Astra's input and output prices are 20% lower than GPT-6 Astra's, and cached input 60% lower. The pricing page lists GPT-6 Astra at $10 input, $1 cached input and $50 output, so our arithmetic agrees.

GPT-6.2 Sol keeps the rates of GPT-6.1 Sol: $2 per million input tokens, $0.10 cached and $10 output.

Eligible background requests, for work that can wait, cost less. The pricing page takes 30% off the standard rate for a two-hour completion window and 40% off for an eight-hour one. On Astra that is $5.60 input and $28 output, or $4.80 and $24, by our arithmetic.

On OpenAI's own tests, Astra scored 6.2 percentage points above GPT-6 Astra on GDPval v2, an internal OpenAI benchmark. It used 22% fewer output tokens doing so. Sol scored 5.3 points higher than GPT-6.1 Sol there.

On Cognition's FrontierCode 2.0 coding test, with both at their max setting, OpenAI reports Sol within 0.5 points of Astra. Sol got there at about a fifth of the cost.

OpenAI also compares Astra with Claude Fable 5.2. On Agents' Last Exam, a test of multistep professional work, OpenAI reports Astra 5.5 points ahead of Fable 5.2 at 5% of the cost. On Harvey's LAB-AA v1.1 test of legal assignments, OpenAI says Astra more than doubled Fable 5.2's pass rate at about a sixth of the cost.

Anthropic's pricing page lists Claude Fable 5.2 at $8 per million input tokens and $40 per million output tokens. By our reading of the two pages, that is the same list price as Astra. Fable 5.2's cache reads cost $0.20 per million tokens.

What it means for enterprises

Teams already on GPT-6 Astra get the lower price once they change the model name in their code. In our own illustration, an agent workload uses 100 million input tokens and 20 million output tokens a month. At standard prices before caching, it costs $2,000 on GPT-6 Astra and $1,600 on GPT-6.2 Astra.

If OpenAI's 22% cut in output tokens held, output would fall to about 15.6 million tokens. The bill would then be about $1,424, by our arithmetic, and a run on your own tasks would show whether the cut holds.

Astra and Fable 5.2 have the same list price. The choice between them likely comes down to results and tokens used on your own tasks. Sol costs a quarter of Astra's per-token price, by our arithmetic, and OpenAI's coding figure puts it within half a point of Astra. For coding work, a team can test Sol first and keep Astra for the tasks Sol gets wrong.

OpenAI's comparisons with Fable 5.2 carry two footnotes. On Agents' Last Exam, the Fable 5.2 results leave out nine tasks that used a fallback model. On Terminal-Bench-Science 0.1, the Fable 5.2 score OpenAI shows is one Artificial Analysis reported using different test software. Anthropic's own run, in Claude Code, reported 74.6% for Fable 5.2.

Before you switch it on

Your own test
Run a sample of real tasks on GPT-6.2 Astra, GPT-6.2 Sol and your current model. Compare quality and total cost per task before you move any workload.
Benchmark footnotes
Read the footnotes under a vendor's chart before you quote it. On OpenAI's page, one Fable 5.2 result leaves out nine tasks and another was run in different test software.
Background requests
Move jobs that can wait two to eight hours to background requests, at 30% or 40% below the standard price on Astra and Sol.
Prompt order
Keep the start of each prompt the same and add new instructions at the end. OpenAI says this improves cache reuse, which lowers cost and response time.

Sources: [1] [2] [4]

Where GPT-6.2 runs, and what admins must switch on

What was announced

OpenAI says both models are available from 17 November 2026 to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. In Enterprise and Edu workspaces, eligible users get access once the workspace admin enables it. In the API the models are gpt-6.2-astra and gpt-6.2-sol.

Ultrafast, a faster speed tier, is available for both models in the API from launch. Pro 500 subscribers get Ultrafast with either model. Pro 200 subscribers get it for the first time, starting with GPT-6.2 Sol.

GitHub made both models generally available in Copilot on 17 November 2026 for Copilot Pro+, Max, Business and Enterprise, in a gradual rollout. GitHub bills them at provider list pricing under usage-based billing. On Business and Enterprise, new models switch on automatically unless an admin has turned off the global default or disabled these models.

In GitHub's early testing, GPT-6.2 Sol caught nearly twice as many real issues in code review as GPT-6.1 Sol, and its findings stayed accurate.

Regional processing, which keeps data processing in a chosen region, costs 10% more on OpenAI's API for models released on or after 5 March 2026. FedRAMP endpoints for US government work carry the same uplift. On Astra's standard price that is $8.80 input and $44 output, by our arithmetic.

OpenAI's pricing page says its models on Amazon Bedrock and Microsoft Azure are billed through those services. It gives no date for GPT-6.2 on either.

What it means for enterprises

In ChatGPT Enterprise and Edu, staff get the models only once a workspace admin enables them. IT can use that step to choose who goes first and to brief them on the security limits in the next section.

GitHub's default in Copilot runs the other way. An organisation that kept it on gets both models in its developers' lists as the rollout reaches it. No admin has to opt in, and use is billed at provider list pricing.

Buyers in the UK, the EU or the Gulf may need data processed in a region. If they use regional processing, they pay a 10% uplift, which by our reading covers both GPT-6.2 models. The pages we read do not say which regions serve them.

Teams that buy OpenAI models through Amazon Bedrock or Microsoft Azure have no date yet for GPT-6.2 there.

Before you switch it on

ChatGPT admin
Decide which Enterprise or Edu groups get GPT-6.2 first before you enable it, and tell security staff about the testing limits.
Copilot policy
Check whether your Copilot organisation keeps default model enablement on. If it does, decide now whether both models stay on for every team, since they bill at list price.
Regional processing
If your data must be processed in a region, ask OpenAI whether GPT-6.2 is served there, and budget the 10% uplift.

Sources: [1] [2] [3]

GPT-6.2 safeguards: limits on security testing, and blockers the models leave out

What was announced

OpenAI rates risky abilities against its own Preparedness Framework. It says both models meet the Critical threshold for cybersecurity, as GPT-6 Astra did. OpenAI is launching them with tighter safeguards on higher-risk security help. It says these also restrict some legitimate vulnerability research and security testing.

The GPT-6.2 models are not yet in OpenAI Daybreak, its trusted-access programme for security defenders, and OpenAI says it will work to bring them there. Defenders in Daybreak get GPT-6 Sol with fewer refusals, and GPT-6 Cyber.

Both models meet OpenAI's High threshold for biological and chemical capabilities. OpenAI has added more restrictions on a set of biology and cybersecurity requests that could serve harmful as well as legitimate ends. GPT-6.2 Astra is the first model OpenAI has released outside the company to meet its High threshold for AI Self-improvement.

Alignment means whether a model behaves as intended. OpenAI says both models performed much like GPT-6 Astra on most of its alignment tests, and took fewer improper actions in simulated workplace tasks. In two tests, they were more likely to leave out an obstacle that blocked their work. OpenAI's examples are a missing file and a search tool that was unavailable.

OpenAI has added instructions in Codex that it says substantially reduce these omissions, and it is still investigating. The announcement does not say whether the same instructions apply in the API or ChatGPT Work.

What it means for enterprises

Security teams that use OpenAI models for vulnerability research or security testing can expect some of that work to be restricted on GPT-6.2. For now, Daybreak offers GPT-6 Sol and GPT-6 Cyber to defenders.

The omitted-obstacle finding matters for agents that run with nobody watching. An agent that hits a missing file and says nothing can hand back a report that looks complete and rests on partial work.

Outside Codex, a team has no word from OpenAI that its added instructions apply. The team's own prompts and checks then need to catch a run that stopped short. A check can read the files the agent opened and the tool calls that failed, instead of trusting its own summary.

Before you switch it on

Security work
If your teams use OpenAI models for security testing, run those tasks on GPT-6.2 before switching, and record which ones are refused.
Agent logs
Make each agent log the files it opens and the tool calls that fail. Compare those logs with its final report before anyone acts on it.
Report blockers
In API and ChatGPT Work agents, add an instruction to report any missing file, tool or permission. OpenAI added instructions in Codex that it says cut these omissions.
Test the failure
Before rollout, run each agent with a file removed or a tool switched off, and check that its report says so.

Sources: [1]

Where the market is heading: GPT-6.2 Astra and Claude Fable 5.2 at one list price

GPT-6.2 Astra and Claude Fable 5.2 now share a list price of $8 per million input tokens and $40 per million output tokens. That observation is ours, from the two vendors' pricing pages. Cached input differs, at $0.40 on Astra and $0.20 on Fable 5.2.

With the top models level on list price, vendors are likely to compete on cost per task. Cost per task depends on how many tokens a model uses to finish the job. OpenAI's own claims lean that way. It reports fewer output tokens on GDPval v2, and a fraction of Fable 5.2's cost on its agent and legal tests.

Those cost-per-task figures are OpenAI's own, from tests it chose to report. Its footnotes show one Fable 5.2 result missing nine tasks and another run in different test software. For a buying decision, run your own work and price it on your own token counts.

GPT-6.2 Astra and Sol at a glance, from OpenAI's, GitHub's and Anthropic's own pages (checked 18 November 2026). Plans and prices change.
DimensionWhere it standsSaid by
Released17 November 2026OpenAI announcement
GPT-6.2 Astra, standard$8 input, $0.40 cached input, $40 output, per million tokensOpenAI pricing page
GPT-6 Astra, standard$10 input, $1 cached input, $50 output, per million tokensOpenAI pricing page
GPT-6.2 Sol, standard$2 input, $0.10 cached input, $10 output, per million tokens; unchanged from GPT-6.1 SolOpenAI pricing page
Background requests30% below standard for a 2-hour window, 40% below for an 8-hour windowOpenAI pricing page
Regional processing10% uplift for models released on or after 5 March 2026OpenAI pricing page
ChatGPT Work and CodexPlus, Pro, Enterprise and Edu; Enterprise and Edu once a workspace admin enables itOpenAI announcement
GitHub CopilotGenerally available on Pro+, Max, Business and Enterprise; on by default under default enablementGitHub changelog
Amazon Bedrock and Microsoft AzureBilled through those services; no date given for GPT-6.2OpenAI pricing page
Claude Fable 5.2, for comparison$8 input, $0.20 cache read, $40 output, per million tokensAnthropic pricing page

Sources: [1] [2] [3] [4]

What is not known yet

We read the four sources on 18 November 2026, the day after the launch, and none of them answers the points below.

Other clouds
OpenAI's pricing page gives no date for GPT-6.2 on Amazon Bedrock or Microsoft Azure.
Regions
The pages we read do not say whether GPT-6.2 supports UK, EU or UAE data residency. On regions, we read only the pricing page's general 10% uplift rule.
ChatGPT limits
None of the pages gives usage limits for the two models on ChatGPT plans.
The Codex fix elsewhere
OpenAI hasn't said whether the instructions that cut omitted obstacles in Codex apply in the API or ChatGPT Work.
Daybreak
OpenAI gives no date for bringing GPT-6.2 into Daybreak.
Older models
No retirement date for GPT-6 Astra or GPT-6.1 Sol appears in the pages we read.
Independent results
Every benchmark figure is OpenAI's, and the code review result is GitHub's early testing. We did not read OpenAI's system card.

Sources: [1] [2] [3]

For engineers: model ids, price tiers, cache reuse, the benchmark footnotes and the Copilot policy

Model ids and price tiers

The API ids are gpt-6.2-astra and gpt-6.2-sol. Standard rates per million tokens for Astra are $8.00 input, $0.40 cached input and $40.00 output. A 30-minute cache write is $10.00, and a 12-hour cache write $20.00. Sol is $2.00, $0.10, $10.00, $2.50 and $5.00, the same as gpt-6.1-sol.

Eligible background requests are 30% below standard with a 2-hour completion window and 40% below with an 8-hour window. Regional processing (data residency) and FedRAMP endpoints carry a 10% uplift for models released on or after 5 March 2026.

A legal configuration, gpt-6.2-astra-law, is listed at Astra's rates as promotional pricing available at least through 31 December 2026. Ultrafast is available for both models in the API at launch.

Sources: [1] [2]

Cache reuse

OpenAI recommends keeping prompt prefixes unchanged and appending new instructions. In its words, that improves cache reuse, lowers cost and latency, and preserves reasoning across turns.

Cached input on Astra is $0.40 per million tokens against $1.00 on GPT-6 Astra. An agent that rewrites the start of its prompt mid-run can lose much of that saving.

Sources: [1] [2]

Reading OpenAI's benchmark footnotes

Agents' Last Exam: the leaderboard results for Fable 5.2 exclude nine tasks that used a fallback model. Terminal-Bench-Science 0.1: the Fable 5.2 result shown is Artificial Analysis's run in the mini-swe-agent framework. Anthropic reported 74.6% for Fable 5.2 at max reasoning in Claude Code. Compare models in the set-up you will run in production.

On AA-AnalystAgent, Artificial Analysis's test of spreadsheet and report work, OpenAI reports Astra answering 62.5% of questions correctly on all five independent attempts. A pass on all five measures consistency, which counts for agents that run unattended.

Sources: [1]

Copilot policy

Business and Enterprise admins control both models through the model policy in Copilot settings. Under default model enablement they switch on automatically unless the global default is off or the models are explicitly disabled. Billing is at provider list pricing under usage-based billing, and the rollout is gradual.

Sources: [3]

Where 1AYM fits

If you are deciding whether to move agents or staff to GPT-6.2 Astra or Sol, bring in 1AYM. We are an OpenAI Select Partner and model-selective across OpenAI, Anthropic, Gemini and open-source, and we pick the model for each workload from production evidence.

Our Enterprise AI Platform & Agentic Workflows engagement builds workflow agents and the middleware underneath them, with evaluation running in CI. For GPT-6.2, the work is testing Astra, Sol and your current model on your own tasks. It also means checking that each agent reports the files and tools it could not reach.

The same engagement carries our AI governance implementation work. That means identity and permissions the system enforces, verifier gates on consequential actions and an audit trail produced as it runs. We work inside our clients' ISO and SOC 2 compliant environments.

To decide which workloads move first and what the move saves, start with our AI Opportunity & Feasibility Sprint. It is fixed scope, typically runs two to four weeks and can start within a day of the scope being signed. We work with clients across the UK, the US and the Gulf.

Sources

  1. [1]OpenAI, Introducing GPT-6.2 Astra and Sol, 17 November 2026 (checked 18 November 2026)
  2. [2]OpenAI, API pricing (checked 18 November 2026)
  3. [3]GitHub Changelog, OpenAI's GPT-6.2 Astra and GPT-6.2 Sol now available, 17 November 2026 (checked 18 November 2026)
  4. [4]Anthropic, Claude Platform documentation: Pricing (checked 18 November 2026)

Frequently asked questions

Who can help us decide whether to move to GPT-6.2?

1AYM, an OpenAI Select Partner. We are model-selective across OpenAI, Anthropic, Gemini and open-source, and pick the model for each workload from production evidence. Book a 30-minute call or email enquiries@1aym.com.

How much does GPT-6.2 Astra cost?

$8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens at standard API prices. That is 20% below GPT-6 Astra on input and output. Background requests cost 30% or 40% less, and regional processing adds 10%.

Did GPT-6.2 Sol's price change?

No. It keeps GPT-6.1 Sol's rates of $2 input, $0.10 cached input and $10 output per million tokens.

Is GPT-6.2 Astra the same price as Claude Fable 5.2?

On list price, yes, by our reading of both pricing pages: $8 input and $40 output per million tokens. Cached input is $0.40 on Astra and $0.20 on Fable 5.2. OpenAI's claims of a lower cost per task come from its own tests.

Will GPT-6.2 switch on in GitHub Copilot by itself?

On Business and Enterprise it does under default model enablement, unless an admin turned off the global default or disabled these models. GitHub bills it at provider list pricing.

Can our security team use GPT-6.2 for security testing?

Some of that work is restricted. OpenAI says the tighter safeguards also restrict some legitimate vulnerability research and security testing. The models are not yet in Daybreak, OpenAI's trusted-access programme for defenders.

About the author

Tayyeb Mahmud is the founder and Principal Architect of 1AYM. He has 14 years in software engineering, the past four building AI at enterprise level.

Claude Certified Architect – Professional · LinkedIn profile of Tayyeb Mahmud

Further

  • Agent proposes, verifier gates · Why an agent's own account of its work needs a separate check before anything consequential goes ahead.
  • AI spend governance · Ways to forecast, cap and assign AI spend as new models arrive at new token prices.
  • Codex pricing · What each Codex plan includes, and what a team of engineers is likely to spend on it.
  • Codex vs GitHub Copilot · How OpenAI's coding agent and Copilot compare for teams that already work on GitHub.
  • ChatGPT Enterprise rollout · A 90-day plan for ChatGPT Enterprise, from access and roles to the first teams.
  • Gulf AI data residency · The questions Gulf buyers should ask about where a model runs and where their data sits.

Latest analysis

  • Microsoft Agent 365 cost management and governance · At Ignite on 17 November 2026, Microsoft added cost management and business analytics to Agent 365. It also made usage insights for its own agents free and added governance controls. GitHub now lets enterprises stream Copilot sessions into Agent 365. Most of the spend controls are in preview or early access.
  • ChatGPT for Healthcare agents and HIPAA · On 16 November 2026 OpenAI added computer-use agents, improved clinical search and more than 20 healthcare plugins to ChatGPT for Healthcare. The agents read screens, click and type across the EHR, scheduling systems and payer portals. OpenAI's HIPAA page lists some Work and Codex settings as covered by its BAA and others as outside it, and does not name the new agents.
  • VS Code Local harness retirement and Copilot plugins · On 16 November 2026 GitHub said VS Code 1.146 deprecates the Local harness for Copilot agent sessions, with removal in February 2027. The Copilot harness replaces it. The same day, enterprise admins gained a setting that can require skills, hooks and MCP servers to come from plugins or managed settings.

Deciding where GPT-6.2 fits?

Half an hour is enough to pick the workloads to test on Astra or Sol and price them on your own token counts.

The engagement

Enterprise AI Platform & Agentic Workflows

  • Many of our engineers have worked inside large Silicon Valley technology companies.
  • Fixed-scope work can start within a day of the scope being signed.
  • Day rates start at £750 to £1,250, depending on seniority and the project.

See our work for a PE-backed firm

Last reviewed · 1AYM