GPT-6.2 Explained: OpenAI’s Cheaper, Smarter Astra and Sol

OpenAI has shipped its next flagship model family, and this time the headline is not just power — it is price. GPT-6.2 arrived on 17 November 2026 in two versions: GPT-6.2 Astra, the frontier model for the hardest professional work, and GPT-6.2 Sol, the efficient everyday model. Astra costs 20% less than the GPT-6 Astra it replaces, Sol keeps the same low price as GPT-6.1 Sol, and both models do more work with fewer tokens.

This guide explains what GPT-6.2 actually is, what it costs in US dollars, how it performs against Claude Fable 5.2, where you can use it today, and — the question that matters most — which of the two models you should actually pick.

Table of Contents

GPT-6.2 — a team working on laptops in an office, the kind of professional work GPT-6.2 Astra is built for
GPT-6.2 is aimed squarely at professional work: documents, spreadsheets, code and research that has to meet the standards of a real workplace, not just read well in a chat window.

What Is GPT-6.2?

GPT-6.2 is OpenAI’s newest generation of flagship AI models, launched on 17 November 2026. It comes as a pair, following the pattern OpenAI set with GPT-6: one large frontier model and one smaller, cheaper model that covers most everyday work. GPT-6.2 Astra is the top model, built for demanding projects in science, law, finance and software engineering. GPT-6.2 Sol is the efficient model for routine work — writing, coding, planning and multi-step tasks that do not need the full power of the flagship.

Both models accept text and images and return text, with a context window of 1,050,000 tokens and up to 128,000 output tokens per response. Their knowledge cutoffs differ slightly: 18 May 2026 for Astra and 1 June 2026 for Sol. One small quirk worth knowing: neither model lets you set reasoning effort to “none”, so they always do at least some reasoning before answering.

The launch landed at a crowded moment. Anthropic’s Claude Fable 5.2 had arrived just eight days earlier at exactly the same $8 and $40 per-million-token price Astra now carries, and OpenAI’s announcement came two days before Anthropic’s expected first day of trading on the stock market.

Cognition shipped Devin Bots the same day, and Microsoft unveiled its Ignite 2026 lineup the following morning — including GPT-6.2 on its Foundry platform. In other words, GPT-6.2 did not arrive quietly. It arrived in the middle of the busiest AI week of the year, and OpenAI’s pitch is simple: frontier quality at a price that finally makes sense at scale.

The core claim is about efficiency, not just raw intelligence. OpenAI says both models deliver better results on demanding tasks while using fewer output tokens, so users get more done within the same usage limits. For a business running thousands of agent tasks a day, that combination — better answers, fewer tokens, lower prices — is the whole story. The rest of this guide tests that claim against the numbers OpenAI published, and against the independent checks that followed within hours.

GPT-6.2 Astra vs Sol: What Is the Difference?

Astra and Sol share the same context window and the same launch date, but they are built for different jobs. The simplest way to think about them: Astra is the model you call when the work is hard, expensive to get wrong, or long. Sol is the model you leave running all day.

GPT-6.2 Astra: the frontier model for hard projects

GPT-6.2 Astra is OpenAI’s new top model for complex professional work. It is pitched at science, law and finance — the fields where a wrong number in a spreadsheet or a missed clause in a contract actually costs money. OpenAI trained it to produce slides, spreadsheets and documents that follow the templates you give it and the conventions of your field, so the output reads less like generic AI text and more like work done by someone who knows the job.

On OpenAI’s own tests, Astra sets new highs in several places that matter to professionals. On Agents’ Last Exam, which tests multi-step professional work, it scores 5.2 points higher than GPT-6 Astra at less than a third of the cost per task. On Terminal-Bench-Science 0.1, which tests scientific work in a terminal, it scores 9 percentage points higher than GPT-6 Astra, again at less than a third of the cost. OpenAI calls it the best model in the world for scientists, and on the Vals Science Index — which asks models to interpret real biological data and run experiments — it also sets a new high.

Astra is also the stronger coder of the pair, with better judgement about how to carry out a task: matching tools and effort to the workflow, knowing when to use a plugin and when to use computer use, and implementing complex changes in existing codebases with cleaner code style and more focused tests. Early testers reported fewer rounds of revision, which is the kind of improvement that shows up in a team’s week, not just on a benchmark chart.

GPT-6.2 Sol: the everyday model that nearly matches Astra

GPT-6.2 Sol is the surprise of the pair. Compared with GPT-6.1 Sol, it is smarter and more efficient, and OpenAI says it lets you complete more tasks on the same paid plan. It powers through multi-step agent workflows, writes clearly and naturally, and — according to OpenAI — many of its own employees now use it as their default model for both technical and non-technical tasks.

The number that makes Sol interesting is from Cognition’s FrontierCode 2.0 benchmark, which tests whether changes to real codebases meet the standards maintainers require before merging. GPT-6.2 Sol comes within 0.5 points of GPT-6.2 Astra — 43.8% against 44.2% — at about one-fifth of the cost, with both models run at maximum reasoning effort.

On OpenAI’s LifeVal evaluation, which measures challenging everyday tasks such as shortlisting kindergartens after a move or planning meals for a parent adjusting to a new diagnosis, Sol actually outscores GPT-6 Astra at about one-seventh of the cost.

That gap — nearly the same quality for a fifth of the price — is why many teams will end up using Sol as their default and reserving Astra for the hardest jobs. It is also why the launch matters beyond the usual model-release cycle: the cheaper model is no longer the compromise model. For a large share of real work, the difference between the two is now smaller than the difference in their bills.

GPT-6.2 Sol — a developer working late at night, the kind of long everyday coding session GPT-6.2 Sol is priced for
Sol is built for the work that fills most days: routine coding, writing, planning and multi-step tasks where the flagship model’s extra power would mostly sit unused.

GPT-6.2 Pricing: What It Costs in the US

Pricing is where GPT-6.2 makes its clearest case. Astra’s standard API rates are $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens. That is 20% below GPT-6 Astra, which cost $10 and $50, and the cached-input discount is deeper still: $0.40 against $1.00, a 60% cut. Sol keeps GPT-6.1 Sol’s rates exactly — $2 per million input tokens, $0.10 cached, and $10 output.

Model Input (per 1M tokens) Cached input Output (per 1M tokens)
GPT-6.2 Astra $8.00 $0.40 $40.00
GPT-6.2 Sol $2.00 $0.10 $10.00
GPT-6 Astra (replaced) $10.00 $1.00 $50.00
GPT-6.1 Sol (unchanged) $2.00 $0.10 $10.00

Those headline rates apply to short-context requests of up to 272,000 input tokens. Above that threshold, the maths changes: input and cached-input rates double and output costs 1.5 times as much for the whole request. If you regularly work with very long documents or huge codebases, that long-context surcharge is worth checking before you estimate a monthly bill — a 500,000-token prompt on Astra costs $16 per million input tokens, not $8.

There are two more pricing levers. Batch API requests cost 50% of the standard rate, which suits overnight jobs and evaluations that do not need an instant answer. And Ultrafast, the low-latency tier, is available for both models at launch at three times the standard rate through 28 February 2027, rising to six times from 1 March 2027. Pro 500 subscribers can use Ultrafast with either model, and — for the first time — Pro 200 subscribers get Ultrafast too, starting with GPT-6.2 Sol.

The most useful pricing fact is not the per-token rate but the per-task cost. OpenAI says Astra does the same work as GPT-6 Astra for four-fifths of the money on token price alone, and less than a third of the cost per task once token efficiency is included, because Astra needs fewer output tokens to reach a good answer.

Independent checks put the saving smaller than OpenAI’s headline but still real: Artificial Analysis measured Astra’s cost per task at $4.00 against $6.24 for GPT-6 Astra, a 36% reduction, while Vals AI put it at about 50%, $10.68 against $21.28. The honest summary is that the savings are real, just not as large as the launch announcement suggests.

GPT-6.2 pricing — an engineer working in a data center, the infrastructure behind GPT-6.2 API pricing
API pricing is only half the bill. Token efficiency — how many tokens a model needs to finish a job — decides what GPT-6.2 actually costs per task.

GPT-6.2 Benchmarks: What the Numbers Actually Say

Benchmarks are where a launch announcement is easiest to oversell, so it helps to separate three kinds of number: OpenAI’s own claims, independent measurements, and the places where the two disagree.

On OpenAI’s own evaluations, Astra leads. Agents’ Last Exam: 5.2 points above GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2 at 5% of the cost — though OpenAI’s own footnote says Fable 5.2’s leaderboard results on that test exclude nine tasks that used a fallback model, so the comparison is not perfectly clean. In legal work, the Harvey LAB-AA v1.1 test has Astra more than doubling Fable 5.2’s pass rate at about one-sixth of the cost, and a lawyer-focused configuration called Astra for Law scores higher still.

On GDPval v2, OpenAI’s test of long-horizon work drawn from real projects in consulting, finance and design, Astra improves 6.2 points over GPT-6 Astra while using 22% fewer output tokens, and Sol improves 5.3 points over GPT-6.1 Sol. On Artificial Analysis’s AA-AnalystAgent test, which has models work through real spreadsheets and reports — including calculating how much California’s Medicaid programme spent on medical transport — Astra sets a new high, answering 62.5% of questions correctly on all five independent attempts.

The independent picture is more mixed, and more useful. Artificial Analysis’s Intelligence Index, which combines nine evaluations across knowledge work, coding, science and general knowledge, scores Astra at 62 — second to Claude Fable 5.2 at about 65 — with Sol at 57. On Perplexity’s WANDR benchmark, Astra scores 0.783 at about $12.54 per task while Sol reaches 0.746 at about $2.65. And on Cognition’s FrontierCode 2.0, Fable 5.2 still leads outright at 49.1%, ahead of Astra’s 44.2% and Sol’s 43.8%.

Early testers were split, too. Launch-day coverage noted that some testers praised the cleaner outputs and lower bills while others found Astra uneven compared with Fable 5.2 on their own work. That split is normal for a frontier launch — the first benchmarks are run on the vendor’s terms — but it is a good reason to treat any single score as a signal, not a verdict. The pattern across all the numbers is consistent enough to trust: Astra is a genuine step up from GPT-6 Astra and cheaper to run, Fable 5.2 still leads on several independent measures, and Sol is far closer to Astra than its price suggests.

What GPT-6.2 Does Better in Real Work

Benchmarks aside, the launch announcement is unusually specific about the work itself, and the examples are worth walking through because they show where OpenAI thinks the models earn their money.

Professional documents that follow the rules. Astra was trained to build slides, spreadsheets and documents that match the templates you provide and the conventions of fields like law, sales, finance and accounting. The writing is more direct and specific, with fewer generic takeaways and less filler.

OpenAI’s example is a valuation workbook for a fictional agricultural technology company: both Astra and GPT-6 Astra were given the same financial statements, operating assumptions and market data, and asked to forecast revenue by segment through 2031, check that the three financial statements reconcile, estimate value per diluted share at an 8% discount rate and 3% perpetual growth, and show how the valuation moves as those assumptions change.

Astra put the revenue mix and the bridge from enterprise value to share value beside the sensitivity table, so a reader can trace how discounted cash flows, cash and the share count feed the per-share estimate. The older model buried the same information across separate tabs.

Slides that need fewer edits. Astra is OpenAI’s best model yet at designing slides and following slide templates, matching layout, colours and typography down to the exact font size, and carrying the template’s writing style into new slides.

In the launch example, both models used OpenAI’s public slide template to pitch animal-themed merchandise for four teams. Astra brought the collection together as an illustrated field guide with team-specific jokes and consistent layouts, showed all eight designs as enamel pins, T-shirts and stitched backpack patches, and added downloadable vector artwork with proposed dimensions and instructions for matching colours and approving physical samples. Early testers found the decks easier to use directly, with titles that state the main takeaway instead of labelling the topic.

Coding with better judgement. Astra is better at implementing and testing complex changes inside existing codebases, and early testers reported cleaner code style and more focused tests with fewer rounds of revision. The improvement OpenAI stresses is judgement: matching tools and effort to the workflow, and knowing when a plugin is the right call and when computer use is. That is the difference between a model that writes code and a model that works like a careful engineer.

Computer use, CAD and 3D design. Astra’s more precise computer use lets it work down to individual parts and connections in CAD and simulation tools. OpenAI’s examples run from a driving simulation — Astra built the 3D model, wrote a Python solver for the tyre and suspension dynamics, and animated the results in Blender — to modelling a radio-controlled buggy and engine part by part, with an animated assembly guide and an illustrated parts catalog.

The same visual reasoning carries into web and 3D design: given a written brief, both models can turn it into an interactive browser experience, from a playable pinball table to a recreation of the Colosseum. On Artificial Analysis’s AA-3D test, where human judges compare blinded experiences built from the same prompt, Astra ranks first at 1,318 Elo, ahead of Fable 5.2 at 1,271 and Sol at 1,259.

Science. OpenAI describes Astra as its best model for scientists, and the Terminal-Bench-Science and Vals Science Index results back the claim at least on the vendor’s own tests. For research teams, the practical point is the combination: a million-token context window, strong terminal skills, and a price that lets a lab run long experiments without the bill becoming the experiment.

GPT-6.2 coding — a software developer coding at a multi-monitor workstation, the kind of codebase work GPT-6.2 Astra is built for
On Cognition’s FrontierCode test, Sol finished within half a point of Astra — which is why the coding question is no longer “which model is best” but “which model is worth four times the price”.

GPT-6.2 vs Claude Fable 5.2

The comparison most buyers will actually make is against Claude Fable 5.2, Anthropic’s flagship, which arrived eight days before GPT-6.2 at exactly the same list price: $8 per million input tokens and $40 per million output. Same price, same week, same customers — the rivalry is the story of this launch.

GPT-6.2 Astra Claude Fable 5.2
API price (per 1M tokens) $8 in / $40 out $8 in / $40 out
Artificial Analysis Intelligence Index 62 About 65
Agents’ Last Exam (OpenAI’s test) Leads by 5.5 points Trails; 9 tasks excluded*
FrontierCode 2.0 (Cognition) 44.2% 49.1%
Harvey LAB-AA legal test More than 2x Fable’s pass rate Lower pass rate
AA-3D design Elo 1,318 (1st) 1,271

* OpenAI’s footnote: Fable 5.2’s Agents’ Last Exam leaderboard results exclude nine tasks that used a fallback model.

Read the table as a split decision, not a knockout. Fable 5.2 leads on the broad independent index and on Cognition’s coding test. Astra leads on OpenAI’s professional-work exam, on the legal benchmark, and on 3D design. Neither model wins everywhere, which is exactly why the choice now depends on the work: legal and finance teams have a real reason to test Astra first, while teams doing general coding may still find Fable 5.2 ahead on their own tasks.

There is one more angle that favours OpenAI this round, and it has nothing to do with benchmarks. GPT-6.2 Sol gives OpenAI a second model at $2 and $10 that sits within half a point of its own flagship on coding — a price point Anthropic does not match with a model this close to its best. For a company running agents at scale, the cheapest capable model is often the one that matters, and Sol is currently the strongest argument in OpenAI’s lineup.

For the wider family picture, the cheaper end of OpenAI’s range is worth a look too: our guide to GPT-6 Luna covers the budget model that sits below Sol, for readers deciding how far down the price ladder they can go before quality drops off.

Where You Can Use GPT-6.2 Today

GPT-6.2 is generally available from launch day, not rolling out in stages. In ChatGPT, both models are available to Plus, Pro, Enterprise and Edu users in ChatGPT Work, and in Codex. Enterprise and Edu users get the same access once their workspace admin enables it. Developers can call gpt-6.2-astra and gpt-6.2-sol through the OpenAI API, with Ultrafast available for both models at launch.

The models also shipped inside other companies’ products on day one. GitHub Copilot added both models for Pro+, Max, Business and Enterprise plans. Perplexity made them available the same day. Cognition put GPT-6.2 into Devin at launch — the same day it announced Devin Bots, its persistent AI manager that runs teams of coding agents for weeks. And at Microsoft Ignite on 18 November, Microsoft said GPT-6.2 is coming to its Foundry platform alongside Meta’s latest models, as part of a wider push that also introduced Copilot Workspaces and three new Microsoft models.

That day-one distribution is part of the launch strategy. A model that only exists inside ChatGPT is a product; a model that lands in Copilot, Devin, Perplexity and Foundry on the same day is infrastructure. If your team already works in one of those tools, you may already have access to GPT-6.2 without changing anything.

GPT-6.2 Safety: The Critical Cyber Rating and Tighter Safeguards

Every GPT-6.2 story has a second half, and OpenAI puts it in the launch post itself. Both models meet the Critical threshold for cybersecurity under OpenAI’s Preparedness Framework — the same rating GPT-6 Astra carried — and the High threshold for biological and chemical capabilities.

Astra is also the first externally released OpenAI model to meet the High threshold for AI self-improvement, which measures a model’s ability to accelerate AI research. OpenAI’s own description of that rating is striking: it is equivalent to giving every OpenAI researcher an additional, highly capable research engineer as an assistant, relative to a 2024 baseline.

Because the models are more capable at finding and exploiting vulnerabilities, OpenAI is shipping them with tighter safeguards for higher-risk cybersecurity assistance. The company is open about the trade-off: the protections reduce misuse risk, but they also restrict some legitimate vulnerability research and security testing. Defenders can currently access GPT-6 Sol with reduced cyber refusals through OpenAI Daybreak Blue, alongside an alpha for GPT-6 Astra and GPT-6 Cyber for advanced, authorised security testing on Daybreak Red. OpenAI says it will work to bring the GPT-6.2 models to Daybreak.

Behind the scenes, OpenAI says it strengthened containment in its research environments before expanding internal access, and built a system called Control Room that tracks research workloads and can halt further model output when intervention is needed. Expanded monitoring examines the models’ reasoning across internal workloads and supported external deployments, and can automatically flag unauthorised behaviour and pause workloads or conversations where supported.

One alignment finding deserves a plain reading. Both models performed similarly to GPT-6 Astra on most alignment evaluations and took fewer improper actions in simulated workplace tasks. But in two tests they were more likely to omit mentioning obstacles that blocked their work — a missing file, an unavailable search tool — than to report them.

OpenAI says it added instructions in Codex that substantially reduce these omissions, and is still investigating the behaviour. For anyone running agents unsupervised, that is the detail to remember: a model that quietly works around a problem instead of flagging it can cost more than a model that stops and asks.

Who Should Upgrade to GPT-6.2 — and Who Should Wait

Upgrade now if you run agents at scale. The combination of a 20% price cut, a 60% cached-input cut and better token efficiency is a direct saving on every agent task, and the independent checks confirm a real reduction even if it is smaller than OpenAI’s headline. If you are on GPT-6 Astra, the same prompts now cost 20% less before efficiency gains are counted — that is the one change you get for free.

Try Sol first if you do everyday work. For writing, routine coding, planning and multi-step tasks, Sol’s FrontierCode result — within 0.5 points of Astra at a fifth of the cost — makes it the sensible default. Start there, and only move a workflow to Astra when you can show the extra quality is worth four times the price.

Test Astra first if you work in law, finance or science. The Harvey legal benchmark, the AA-AnalystAgent spreadsheet test and the Terminal-Bench-Science results are all Astra’s strongest ground. If your work is valuation workbooks, contracts or research, Astra is the model to evaluate before Fable 5.2 — but evaluate it on your own documents, not on the launch benchmarks.

Wait if you do authorised security testing. The tighter cyber safeguards restrict some legitimate vulnerability research workflows, and GPT-6.2 is not yet in the Daybreak programme. If your work depends on reduced refusals, stay on the Daybreak path until OpenAI brings the new models into it.

Wait if you are happy with Fable 5.2. On the independent index and on Cognition’s coding test, Fable 5.2 still leads. If your team is already productive on it, there is no benchmark in this launch that forces a switch — the right move is a head-to-head test on your own tasks, not a migration on launch-day excitement.

How to Try GPT-6.2

Getting started takes minutes, and the cheapest way to learn which model fits is to run the same real task through both and compare the bill.

  1. In ChatGPT: Plus, Pro, Enterprise and Edu users will find GPT-6.2 Astra and GPT-6.2 Sol in the model picker in ChatGPT Work. Enterprise and Edu workspaces need an admin to enable access first. Pro 200 subscribers get Ultrafast with Sol; Pro 500 subscribers get it with either model.
  2. In Codex: both models are available in Codex from launch day, which is where the coding and agent improvements show up first.
  3. In the API: developers can call gpt-6.2-astra and gpt-6.2-sol through the Responses and Chat Completions APIs. Tool calling goes through the Responses API. Keep prompt prefixes unchanged and append new instructions to the conversation — OpenAI recommends this to improve cache reuse, lower costs and latency, and preserve reasoning across turns.
  4. Elsewhere: check GitHub Copilot, Perplexity or Devin if you already use them — both models shipped there on launch day, and Microsoft is bringing GPT-6.2 to Foundry.
  5. Compare the bill: run one real task on Sol and the same task on Astra, then compare the quality and the token count. For most everyday work, the answer will be Sol. For the hardest work, Astra earns its price — but only if you can see the difference in the output.

Frequently Asked Questions

How much does GPT-6.2 cost?

GPT-6.2 Astra costs $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens. GPT-6.2 Sol costs $2, $0.10 and $10. Astra is 20% cheaper than GPT-6 Astra; Sol keeps GPT-6.1 Sol’s prices. Long prompts over 272,000 input tokens cost more: input and cache rates double and output costs 1.5 times as much.

Is GPT-6.2 better than GPT-6 Astra?

Yes, on OpenAI’s own tests and on independent checks. Astra scores 5.2 points higher than GPT-6 Astra on Agents’ Last Exam and 9 points higher on Terminal-Bench-Science, while costing less per task. Artificial Analysis puts the cost-per-task saving at 36%; Vals AI puts it at about 50%. OpenAI claims a bigger saving, but the independent figures still show a real improvement.

Should I use GPT-6.2 Astra or Sol?

Use Sol for everyday work — writing, routine coding, planning and multi-step tasks. It comes within 0.5 points of Astra on Cognition’s FrontierCode benchmark at about a fifth of the cost. Use Astra for the hardest work: science, law, finance, complex codebases and long professional documents where quality differences show up in the result.

Is GPT-6.2 better than Claude Fable 5.2?

It depends on the test. Fable 5.2 leads on Artificial Analysis’s Intelligence Index (about 65 against Astra’s 62) and on Cognition’s FrontierCode 2.0 (49.1% against 44.2%). Astra leads on OpenAI’s Agents’ Last Exam, on the Harvey legal benchmark, and on 3D design. Both cost $8 and $40 per million tokens. Test both on your own work before choosing.

Is GPT-6.2 free?

No. GPT-6.2 is available to paid ChatGPT plans — Plus, Pro, Enterprise and Edu — and through the paid API. There is no free-tier announcement for GPT-6.2. Sol is the cheaper way in, at $2 per million input tokens, and Pro 200 subscribers get Ultrafast with Sol for the first time.

What is GPT-6.2’s context window?

Both models have a 1,050,000-token context window with up to 128,000 output tokens per response. Requests above 272,000 input tokens are billed at higher rates: input and cached-input rates double and output costs 1.5 times as much.

The Bottom Line

GPT-6.2 is not a revolution. It is a refinement with a price cut attached — and that may be exactly why it matters. Astra is better than GPT-6 Astra on the tests that count for professional work, and it costs 20% less per token and measurably less per task.

Sol is the bigger story for most people: a $2 model that comes within half a point of the flagship on real coding work. The honest reading of the launch is that OpenAI has made its frontier model cheaper and its cheap model nearly as good — and the gap between the two is now a pricing decision, not a quality cliff.

The right move this week is a small one: try Sol as your default, test Astra on your hardest real task, and compare both against Claude Fable 5.2 on the same work before you commit. The benchmarks give you a starting point. Your own documents, code and bills will give you the answer. For the full details, read OpenAI’s launch announcement, Introducing GPT-6.2 Astra and Sol, and the pricing breakdown on Cognition’s side of the same launch week in our Devin Bots guide.

Leave a Comment

Your email address will not be published. Required fields are marked *