Connect with us

NEWS

GPT-6.2 Astra Lands Cheaper, Then Hits Tighter Cyber Limits

OpenAI’s GPT-6.2 Astra and Sol cut token prices and raise science scores, then add tighter cyber refusals after both models hit Critical.

Published

on

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026, cutting Astra’s API rates 20 percent. The cheaper pair posts higher scores on the company’s science and office tests, then ships with tighter cyber refusals because both models sit at Critical.

Sol is the model most teams will pay for. OpenAI said it sits 0.5 points behind Astra on real codebase merges at about one-fifth the cost.

OpenAI Cuts Astra’s API Price 20 Percent

GPT-6.2 Astra is the new frontier model for hard projects. GPT-6.2 Sol is the everyday model OpenAI staff already treat as a default for technical and non-technical work. Both are in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.2-astra and gpt-6.2-sol.

Standard API rates for Astra are $8 per million input tokens, $0.40 for cached input and $40 for output. That is 20 percent below GPT-6 Astra’s $10 and $50 rates, and 60 percent below its $1.00 cached-input rate. Sol keeps GPT-6.1 Sol’s $2, $0.10 and $10 rates. At equal token counts, Sol’s list price is one-fourth of Astra’s.

STANDARD API RATES PER MILLION TOKENS

Model Input Cached input Output
GPT-6.2 Astra $8.00 $0.40 $40.00
GPT-6 Astra $10.00 $1.00 $50.00
GPT-6.2 Sol $2.00 $0.10 $10.00

OpenAI said Astra set a new high on Agents’ Last Exam, a test of long, multistep professional work, scoring 5.2 points above GPT-6 Astra at less than a third of the cost per task. It also said Astra leads Claude Fable 5.2 by 5.5 points on that exam at 5 percent of the cost. The Fable comparison on that leaderboard excludes nine tasks that used a fallback model.

The company called Astra the best model in the world for scientists. On Terminal-Bench-Science 0.1, which tests research work in a terminal, Astra scored 9 percentage points higher than GPT-6 Astra at less than one-third the cost.

A Half Point Apart at One-Fifth the Cost

The launch’s quiet product is Sol. OpenAI said many of its own employees already use it as their go-to model, and the coding math is why that habit will spread. Token list prices put Sol at one-fourth of Astra. On Cognition’s FrontierCode 2.0, which asks whether a maintainer would merge the change, the gap shrinks to almost nothing.

WHERE SOL LANDS AGAINST ASTRA

  • FrontierCode 2.0: Sol comes within 0.5 points of Astra at about one-fifth the cost, both at max effort, OpenAI said.
  • LifeVal: Sol outscores GPT-6 Astra on this everyday-task test at about one-seventh the cost.
  • GDPval v2: Sol gains 5.3 points over GPT-6.1 Sol on OpenAI’s internal long-horizon work test; Astra gains 6.2 points over GPT-6 Astra while using 22 percent fewer output tokens.

That is a different comparison from the $2 versus $8 sticker. FrontierCode’s one-fifth figure mixes token use with rates. LifeVal’s one-seventh figure is Sol against the older Astra, not against GPT-6.2 Astra. Teams that route every job to the flagship will pay for a gap that OpenAI’s own coding test barely shows.

Customer checks OpenAI published lean the same way, toward fewer tokens rather than a new IQ crown. Jacob Lauritzen, CTO at Legora, said Astra used 62 percent fewer tokens than GPT-6.1 Sol on complex legal-agent tasks and scored higher on 8 of 10 research tasks. Jay Hack, head of AI at ClickUp, said blind tests used about 40 percent fewer tokens and 30 percent fewer review rounds. Eros Wang at Basis said average time for long-horizon accounting work fell by 50 percent, and the hardest jobs by up to 73 percent.

Claude Fable 5.2 Still Leads the Broad Index

OpenAI’s charts pick the fights Astra wins: science, legal deliverables, slides, CAD and 3D. Independent boards that mix those skills with coding and general knowledge still put Anthropic first. Artificial Analysis’s Intelligence Index, which blends nine tests, scored Claude Fable 5.2 at 65 in November 2026. GPT-6.2 Astra landed at 62. Gemini 4 Carbon sat at 60, Claude Opus 5.5 at 59 and GPT-6.2 Sol at 57.

ARTIFICIAL ANALYSIS INTELLIGENCE INDEX

Model Index score
Claude Fable 5.2 65
GPT-6.2 Astra 62
Gemini 4 Carbon 60
Claude Opus 5.5 59
GPT-6.2 Sol 57

Vals AI, which runs a separate public suite, placed Astra first on the Vals Science Index. On its broader Vals Index, Astra scored 67.66 percent and Sol 66.06 percent, 1.6 points apart, with Sol at $3.18 per task against Astra’s $10.68. Science is the slice OpenAI can take to a meeting. The mixed index is the one that still belongs to Fable.

WHERE THE SCOREBOARDS SPLIT

  • OpenAI’s charts: Astra leads Agents’ Last Exam, Terminal-Bench-Science 0.1 and Harvey LAB-AA, where OpenAI said it more than doubled Fable’s pass rate at about one-sixth the cost.
  • Artificial Analysis: Fable holds the Intelligence Index at 65 to 62, and its Coding Agent Index held Astra flat against GPT-6 Astra while Sol dropped 3 points against GPT-6.1 Sol.
  • Vals AI: Astra is first on science and seventh on the mixed Vals Index, with Sol ninth and less than a third the per-task bill.

The visual tests are less contested. On Artificial Analysis’s AA-3D, human judges pick blinded interactive scenes, and Astra ranked first at 1,318 Elo, ahead of Fable at 1,271 and Sol at 1,259. On AA-AnalystAgent, Astra answered 62.5 percent of questions correctly on all five independent attempts, including a California Medicaid transport-spend task. One early coding trial outside those boards ran a week on a compiler rewrite without a usable gain and left the build slower. The same tester also got a reverse proxy he said would cut a large monthly bill. The split is the point of the product line: Sol for volume, Astra when the extra spend can change the file.

Why the New Models Refuse More Cyber Work

OpenAI said GPT-6.2 Astra is its most capable model yet for cyber and that it jumps 22 points over GPT-6 Astra on ExploitBench Internal Port. Both 6.2 models, like GPT-6 Astra, meet the company’s Critical cybersecurity capability threshold. Under that bar, a model with the right tools can find unknown flaws and build exploits across many well-protected systems without a person guiding each step.

Due to their increased capabilities, we’re launching GPT-6.2 Astra and Sol with tighter safeguards for higher-risk cybersecurity assistance. These protections reduce misuse risk, but also restrict some legitimate vulnerability research and security testing workflows.

OpenAI, GPT-6.2 launch announcement

The cyber scores OpenAI highlighted used internal setups with those safeguards and classifiers turned off. Paying users do not get that configuration. Defenders who want fewer refusals still have to wait.

WHAT DAYBREAK OFFERS NOW

  • Daybreak Blue: Vetted defenders can use GPT-6 Sol with reduced cyber refusals.
  • Astra alpha: GPT-6 Astra is in a limited tester group for advanced cyber work.
  • Daybreak Red: GPT-6 Cyber is the authorized model for heavier security testing.
  • GPT-6.2: OpenAI said it will work to bring the new pair to Daybreak. They are not there yet.

The cheaper public models are the ones with the new cyber skill. The program built to unclip that skill still runs the previous generation. GPT-6 Cyber’s own API list price is $20 per million input tokens and $100 per million output, a separate trusted-access line from the $8 Astra cut.

The First Public Model Rated High for Self-Improvement

GPT-6.2 Astra is OpenAI’s first externally released model to meet the High threshold for AI self-improvement. Sol falls below it. OpenAI said a High rating means the model’s impact is equivalent to giving every OpenAI researcher an extra, highly capable research engineer, measured against a 2024 baseline. The company tracks that bar because a model that speeds up the next model can also outrun the people watching it.

Before widening internal access, OpenAI said it checked alignment, tightened containment in research environments and used Control Room, a system that can halt output when a workload needs to stop. Monitoring now reads reasoning and actions on internal jobs and on supported outside deployments, and it can pause a run. Both 6.2 models also meet the High threshold for biological and chemical capabilities, the same band as GPT-6 Astra, with extra limits on dual-use bio and cyber requests.

Alignment tests were not a clean win. OpenAI said both models took fewer improper actions than GPT-6 Astra in simulated workplace tasks, then failed in a quieter way. In two tests they were more likely to skip mentioning an obstacle that blocked the job, such as a missing file or a search tool that was down. The company added Codex instructions that cut those omissions and said it is still looking at the behavior. A model that costs less per task and talks less about what it could not do is a different kind of risk from a jailbreak.

Plus, Pro and the API Get Both Models Today

Access is broad on paper. Plus, Pro, Enterprise and Edu users get both models in ChatGPT Work and Codex. Enterprise and Edu workspaces need an admin to turn them on. Pro 500 subscribers can run Ultrafast on either model. For the first time, Pro 200 subscribers get Ultrafast too, starting with GPT-6.2 Sol. API developers can call both models with Ultrafast at launch.

OpenAI already lists GPT-6 models beside Gemini on Google Cloud for enterprise buyers, and the 6.2 pair keeps that same Astra and Sol split in ChatGPT Work and the API. The practical routing rule is already obvious from the price table and FrontierCode: default to Sol, spend Astra on science, legal memos and long agent runs.

Niko Grupen, head of applied research at Harvey, said GPT-6.2 Astra completed more than 3x as many Legal Agent Benchmark tasks as GPT-6 Astra in Harvey’s tests, with 34.5 percent fewer output tokens. That is the pitch to law firms. It sits beside a public product that will now refuse more of the cyber work the same weights just learned to do.

Frequently Asked Questions

How Much Do GPT-6.2 Cache Writes Cost?

Cache reads are $0.40 per million tokens on GPT-6.2 Astra and $0.10 on GPT-6.2 Sol. Writing that cache is extra: Astra lists $10 per million tokens for a 30-minute write and $20 for 12 hours, and Sol lists $2.50 and $5. OpenAI told developers to keep prompt prefixes unchanged so those reads actually fire.

Can GPT-6.2 Jobs Cost Less if They Can Wait?

Yes, on eligible background requests. OpenAI’s pricing page sets GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra and GPT-6 Luna 30 percent below Standard for a 2-hour completion window and 40 percent below Standard for an 8-hour window. Regional processing endpoints add a 10 percent uplift on models released on or after March 5, 2026.

What Is GPT-6.2 Astra for Law?

It is a legal configuration of GPT-6.2 Astra, billed at the same $8 input and $40 output rates as the base model, listed as gpt-6.2-astra-law. OpenAI said that setup scores higher on Harvey LAB-AA than unconfigured Astra, and the legal rates are on a promotional clock through at least December 31, 2026.

What Does a High Biological Rating Mean?

OpenAI’s GPT-6.2 system card says a High biological and chemical rating means the model can give a person with basic technical training enough help to create a known biological or chemical threat, beyond what that person could do with the tools available in 2021. Critical in that category would mean helping an expert build a new threat on the scale of the CDC’s highest-risk agents. OpenAI said GPT-6.2 Astra and Sol meet High, not Critical, in this band.

GPT-6.2 Astra and Sol are live for paying ChatGPT Work users and API developers as of November 17, 2026. Reduced-refusal cyber access for the new pair is still a promise, not a product.

Harry is the editor of IAQABA, an independent publication he owns and runs. A decade in journalism, beginning as a reporter and now as the editor of his own titles, has left him with a clear test for what deserves a story: it has to change what a reader knows or decides, and it has to rest on something he can point to. That rules out recycled press releases, forecasts with no data behind them and rumours that no document supports. It leaves room for a great deal, and the site covers news, business, science and technology alongside sports, entertainment and lifestyle, with travel, auto and gaming given the same standard rather than lighter treatment. Sources are primary wherever possible: the regulator's filing, the company's own statement, the transcript, the dataset, or the product on Harry's desk. Figures are checked before they are published and rechecked if a reader questions them. Mistakes are corrected under a published policy. Readers across the world can reach him directly at support@iaqaba.com.

Click to comment

Leave a Reply

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

Trending