On this page(18)
- What GPT-6.2 Astra and Sol Are and What Changed From GPT-6 Astra and GPT-6.1 Sol
- What Do GPT-6.2 Astra and Sol Cost in the API and in ChatGPT?
- Benchmarks Where GPT-6.2 Astra Leads: Agentic, Science and Legal Work
- Benchmarks Where Claude Fable 5.2 and Gemini 4 Carbon Still Beat GPT-6.2 Astra
- GPT-6.2 Sol vs GPT-6.2 Astra: Which Model Fits Which Workload
- How GPT-6.2 Astra Handles Spreadsheets, Memos and Slide Decks
- Computer Use, CAD Modeling and 3D Web Design With GPT-6.2 Astra
- Coding Quality and Token Efficiency Reported by Early Testers
- Cybersecurity Capabilities and the Tighter Safeguards That Restrict Them
- Safety and Alignment Findings: Blocker Omissions and the AI Self-Improvement Threshold
- GPT-6.2 Astra and Sol FAQ
- Is GPT-6.2 Astra better than Claude Fable 5.2?
- Is GPT-6.2 available on ChatGPT Plus, and why can't I see it yet?
- What is the GPT-6.2 context window and knowledge cutoff?
- What is Ultrafast and who gets it?
- Can GPT-6.2 Sol replace Astra for coding?
- Does GPT-6.2 support fine-tuning?
- Where to Start With GPT-6.2 Astra and Sol
GPT-6.2 Astra and Sol are OpenAI's new frontier and everyday models, launched November 17, 2026 at $8/$40 and $2/$10 per million input/output tokens. Astra tops OpenAI's agentic, science and legal charts, while Claude Fable 5.2 still leads on coding and the Artificial Analysis Intelligence Index at 2.4 to 6.6 times the cost per task. Below: the price list, benchmark results on both sides, an Astra-or-Sol split by workload, and the safety findings that matter if you run agents. Every figure is current as of November 17, 2026.
- Agentic work: GPT-6.2 Astra scored 78.3% on Agents' Last Exam v1.1 for $2.89 per task in OpenAI's November 2026 launch data, against 72.8% and $60.05 for Claude Fable 5.2.
- Intelligence index: Claude Fable 5.2 leads GPT-6.2 Astra 64.6 to 62.1 on the Artificial Analysis Intelligence Index v5.0.1 in November 2026, at $26.54 per task against $4.00.
- Coding: GPT-6.2 Sol scored 43.8% on Cognition's FrontierCode 2.0 for $1.29 per task in November 2026, 0.4 points behind Astra at $5.97.
- Agent honesty: with no developer prompt, GPT-6.2 Astra failed to disclose a blocking obstacle 22.28% of the time in OpenAI's November 2026 testing, about double GPT-6 Astra's 11.09%.
- Ultrafast pricing: the speed tier costs 3x standard API rates through February 28, 2027 and 6x from March 1, 2027, per OpenAI's GPT-6.2 Sol model page as of November 2026.
What GPT-6.2 Astra and Sol Are and What Changed From GPT-6 Astra and GPT-6.1 Sol
GPT-6.2 Astra and Sol are two OpenAI models released on November 17, 2026: Astra is the frontier model for the hardest projects, and Sol is the cheaper, faster model for everyday and agentic work. Astra succeeds GPT-6 Astra. Sol succeeds GPT-6.1 Sol.
"Both models deliver better results on demanding tasks with fewer output tokens, so you can get more done within your usage limits."
OpenAI, in its announcement of GPT-6.2 Astra and Sol on November 17, 2026
The two models share almost every spec. Per OpenAI's developer documentation as of November 17, 2026, only the model id and the knowledge cutoff differ.
| Spec | GPT-6.2 Astra | GPT-6.2 Sol |
|---|---|---|
| Role | Frontier model for the hardest projects | Efficient model for everyday and agentic work |
| API model id | gpt-6.2-astra | gpt-6.2-sol |
| Context window | 1,050,000 tokens | 1,050,000 tokens |
| Max output | 128K tokens | 128K tokens |
| Knowledge cutoff | May 18, 2026 | June 1, 2026 |
Astra's window splits into 922K tokens of input and 128K of output. One quirk on the Sol side: its API rejects any request that sets reasoning effort to none, so you can't run it as a no-reasoning model.
What changed from the previous generation comes down to four things, all as described by OpenAI:
- Token efficiency: both models reach better results with fewer output tokens, which stretches ChatGPT usage limits and lowers API cost per task.
- Document quality: Astra's slides, spreadsheets and documents follow supplied templates and field conventions more closely, and OpenAI says they "read less like AI".
- Code and frontends: Astra writes cleaner code and builds better-looking websites, games and 3D work.
- Sol's upgrade: OpenAI calls Sol smarter and more efficient than GPT-6.1 Sol, and says many of its own employees use it as their default model.
The document-quality push answers a real complaint. Early testers of GPT-6 Astra in September 2026 said its writing and creative output had slipped, according to Decrypt's November 17, 2026 report, and OpenAI now says 6.2 fixes that. It's a company claim for now.
The release pace is fast. GPT-6 Astra shipped on September 3, 2026, which puts GPT-6.2 Astra about eleven weeks behind it. Secondary reports place GPT-6 Sol and Luna around September 22, 2026 and GPT-6.1 Sol around September 29, 2026, so treat those two dates as approximate.
What Do GPT-6.2 Astra and Sol Cost in the API and in ChatGPT?
GPT-6.2 Astra costs $8 per million input tokens, $0.40 cached and $40 output in the API, while GPT-6.2 Sol costs $2, $0.10 and $10, as of November 17, 2026. In ChatGPT, both models are available to Plus, Pro, Enterprise and Edu users inside ChatGPT Work and Codex.
The rates below come from OpenAI's API pricing page as of November 17, 2026. They are per 1M tokens, for requests of up to 272K input tokens.
| Model | Input | Cached input | Output | Batch (in / out) | Fast mode (in / out) |
|---|---|---|---|---|---|
| GPT-6.2 Astra | $8.00 | $0.40 | $40.00 | $4 / $20 | $16 / $80 |
| GPT-6.2 Sol | $2.00 | $0.10 | $10.00 | $1 / $5 | $4 / $20 |
| GPT-6 Astra (prior) | $10.00 | $1.00 | $50.00 | $5 / $25 | $20 / $100 |
| GPT-6.1 Sol (prior) | $2.00 | $0.10 | $10.00 | $1 / $5 | $4 / $20 |
Astra's input and output prices are 20% below GPT-6 Astra's, and cached input is 60% lower. Sol's prices are unchanged from GPT-6.1 Sol. Anthropic's Claude Fable 5.2 also lists at $8 input and $40 output, per VentureBeat's coverage of its November 9, 2026 launch, so the two flagships tie on list price.

Three line items in OpenAI's 2026 pricing change the bill more than the headline rates suggest:
- Long prompts: above 272K input tokens, Astra rises to $16 input and $60 output, and Sol to $4 and $15. The surcharge applies to the full request, so a 300K-token Astra prompt pays $4.80 for input where the standard rate would have charged $2.40.
- Ultrafast: it costs 3x standard through February 28, 2027, then 6x from March 1, 2027. Any budget that includes Ultrafast doubles on that line in under four months.
- Background and regional: background requests run 30% below standard for the 2-hour window and 40% below for the 8-hour window. Regional and FedRAMP endpoints add 10%.
Caching is the cheapest saving on offer. OpenAI recommends keeping prompt prefixes unchanged and appending new instructions to the conversation, which improves cache reuse. Cached Astra input at $0.40 is 95% cheaper than fresh input at $8.
Access in ChatGPT depends on plan, as of November 17, 2026:
- Plus, Pro, Enterprise and Edu: both models in ChatGPT Work and Codex. They are absent from regular Chat.
- Enterprise and Edu workspaces: access starts once the workspace admin enables it.
- Pro 500: Ultrafast with either model.
- Pro 200: Ultrafast for the first time, starting with Sol.
The rollout is staged. OpenAI's community thread says availability depends on plan, workspace permissions and rollout progress. Several users in that thread, and Decrypt's reporter on a Plus plan, couldn't see the models on launch day.
Benchmarks Where GPT-6.2 Astra Leads: Agentic, Science and Legal Work
GPT-6.2 Astra posts the top score on five agentic, science and legal charts in OpenAI's November 17, 2026 launch data: Agents' Last Exam v1.1, Terminal-Bench-Science 0.1, Vals Science Index v1, Harvey LAB-AA v1.1 and the internal GDPval v2. OpenAI published every figure here, including the runs for Anthropic's and Google's models.
| Benchmark (November 2026) | GPT-6.2 Astra | GPT-6 Astra | Claude Fable 5.2 |
|---|---|---|---|
| Agents' Last Exam v1.1 | 78.3% at Xhigh, $2.89 | 73.1% at Max, $8.99 | 72.8% at Max, $60.05 |
| Terminal-Bench-Science 0.1 | 77.1% at Max, $7.15 | 68.1% at Max, $23.80 | 73.3% at Max, $45.21 |
| Vals Science Index v1 | 68.4% at Max, $4.17 | 65.1% at Max, $8.03 | 63.7% at High, $8.79 |
| Harvey LAB-AA v1.1 | 15.0% at Max, $6.35 | 8.6% at Max, $13.67 | 7.2% at Max, $39.84 |
| GDPval v2 (internal) | 50.4% at Max, 35,812 output tokens | 44.2% at Max, 45,732 output tokens | 44.8% at Max, 222,344 output tokens |
Agentic work. On Agents' Last Exam v1.1, Astra's 78.3% at Xhigh sits 5.2 points above GPT-6 Astra at under a third of the cost per task, and 5.5 points above Fable 5.2 at about 5% of the cost. Two details get lost in that headline. Astra's Max setting scores lower (77.6%) and costs more ($4.47). And Gemini 4 Carbon is close behind at 77.4% for $4.60 at High. The chart also notes that Fable 5.2's leaderboard results exclude nine tasks that used a fallback model.

Science. Astra's 77.1% on Terminal-Bench-Science 0.1 beats GPT-6 Astra by 9 points. Fable 5.2's 73.3% comes from Artificial Analysis's run in the mini-swe-agent harness, while Anthropic reports 74.6% at max reasoning in Claude Code. Take Anthropic's number and Astra's lead shrinks from 3.8 points to 2.5. It still leads. On the Vals Science Index, Astra at High already scores 68.1% for $2.25, so Max buys 0.3 points for nearly double the cost.
Legal. Harvey LAB-AA v1.1 gives credit only when a deliverable meets every grading criterion and contains no material hallucinations. Astra more than doubles Fable 5.2's pass rate at about one-sixth the cost. Read the absolute number, though: at 15.0%, the best model on this chart still misses that bar on 85% of assignments. I'd treat it as a sign of progress on a strict test, and keep a lawyer reviewing every output.
GDPval v2 deserves less weight because OpenAI runs it internally. Astra's 50.4% is up 6.2 points on GPT-6 Astra, with Gemini 4 Carbon the nearest rival at 48.8% at High.
Curious what AI could do for your business?
No jargon and no hard sell. Just a friendly look at where AI fits, and where it doesn't.
A vendor picks its own benchmarks, harnesses and effort levels. Fable 5.2 appears only at Max on Agents' Last Exam and only at High on the Vals Science Index. For teams running production AI agents, the cost column is the more useful one here: Astra's wins come at $2.89 to $7.15 per task where Fable 5.2 runs $8.79 to $60.05.
Benchmarks Where Claude Fable 5.2 and Gemini 4 Carbon Still Beat GPT-6.2 Astra
Claude Fable 5.2 and Gemini 4 Carbon still beat GPT-6.2 Astra on four charts in OpenAI's own November 17, 2026 launch data: the Artificial Analysis Intelligence Index v5.0.1, FrontierCode 2.0, BenchCAD 2.1 Core and AutomationBench 1.0.6. The leads run from 1.0 to 4.9 points. Each one costs 2.4 to 9.3 times more per task.
| Benchmark (November 2026) | Top score | GPT-6.2 Astra | Lead | Leader's cost multiple |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index v5.0.1 | Claude Fable 5.2: 64.6 at Max, $26.54 | 62.1 at Max, $4.00 | 2.5 points | 6.6x |
| FrontierCode 2.0 | Claude Fable 5.2: 49.1% at Xhigh, $14.19 | 44.2% at Max, $5.97 | 4.9 points | 2.4x |
| BenchCAD 2.1 Core | Claude Fable 5.2: 83.5% at Max, $50.77 | 82.5% at Max, $5.43 | 1.0 point | 9.3x |
| AutomationBench 1.0.6 | Gemini 4 Carbon: 53.9% at High, $2.12 | 52.4% at Max, $0.83 | 1.5 points | 2.6x |
Coding is the widest gap. FrontierCode 2.0, built by Cognition, checks whether changes to real codebases meet maintainers' standards for merging, and Fable 5.2 leads Astra there by 4.9 points. Claude Opus 5.5 edges Astra too, with 45.5% at Max for $12.82. Outside testing points the same way: Decrypt's launch-day report on GPT-6.2 Astra says independent tests found no improvement in Astra's coding-agent performance.
The index gap is narrower than the headline suggests. At Xhigh, Fable 5.2 scores 62.2 for $14.83 and Astra scores 61.6 for $2.52. The Artificial Analysis listing for GPT-6.2 Astra rounds the model to 62 and ranks it #3 of 91 as of November 17, 2026, while Decrypt calls it second.

I'd read BenchCAD and AutomationBench as ties. Three caveats apply to the whole set:
- Fable's fallback: Artificial Analysis reportedly ran Fable 5.2 with Anthropic's default fallback enabled, which hands requests that trigger cyber and biology safeguards to another model. Part of Fable's index score may come from that second model.
- Effort coverage: OpenAI published all four charts, rival runs included. Gemini 4 Carbon appears on three of them and never above High effort.
- ARC-AGI: near-perfect ARC-AGI scores for GPT-6.2 circulated on launch day through TestingCatalog, which flagged its own ARC-AGI-3 labels as unclear. Don't quote them as fact. ARC Prize's verified September 3, 2026 results for GPT-6 Astra ranged from 62.7% to 99.9% on ARC-AGI-3 depending on the harness, so a score with no harness named tells you little.
GPT-6.2 Sol vs GPT-6.2 Astra: Which Model Fits Which Workload
GPT-6.2 Sol fits everyday coding, agent loops and personal tasks, and GPT-6.2 Astra fits science, legal and long-horizon professional work. On OpenAI's November 17, 2026 launch charts, Sol at Max lands 0.4 to 4.6 points behind Astra at Max. On the three charts that report cost, it runs at 22% to 30% of Astra's cost per task.
| Benchmark (both at Max, November 2026) | GPT-6.2 Sol | GPT-6.2 Astra | Astra's lead |
|---|---|---|---|
| FrontierCode 2.0 | 43.8%, $1.29 | 44.2%, $5.97 | 0.4 points |
| LifeVal | 83.8%, $1.27 | 86.1%, $4.23 | 2.3 points |
| GDPval v2 (internal) | 46.9%, 33,428 output tokens | 50.4%, 35,812 output tokens | 3.5 points |
| Agents' Last Exam v1.1 | 73.0%, $1.32 | 77.6%, $4.47 | 4.6 points |
Sol also holds up against the previous flagship. Its 83.8% on LifeVal beats GPT-6 Astra's 79.8%, at $1.27 a task against $9.04, about one-seventh of the cost. On GDPval v2, Sol's 46.9% tops GPT-6 Astra's 44.2%. On Agents' Last Exam v1.1 it sits 0.1 points behind (73.0% against 73.1%) for $1.32 where GPT-6 Astra cost $8.99.
Effort level changes the picture, and two results in the 2026 charts are easy to miss:
- Sol's near-tie on FrontierCode 2.0 exists only at Max. At Xhigh, Sol scores 39.8% for $0.98, which is 4.3 points under Astra's Xhigh score of 44.1%.
- Low effort got worse on LifeVal. Sol scores 25.3% at Low and Astra 46.3%, where GPT-6.1 Sol reached 66.8% and GPT-6 Astra 67.5% at the same setting. A Low default carried over from the older models will cost quality.
Here is how I'd map workloads to a model and an effort level:
- Everyday coding and agent loops: Sol at Max, at $1.29 per FrontierCode 2.0 task.
- Personal and everyday tasks: Sol at High or above. High scores 80.8% on LifeVal for $0.65.
- Science and legal work: Astra. Sol trails by 9.1 points on Terminal-Bench-Science 0.1 (68.0% against 77.1%) and by 3.9 points on Harvey LAB-AA v1.1 (11.1% against 15.0%).
- Long-horizon professional work: Astra at Xhigh. Artificial Analysis's November 2026 run scored Astra 62 at both Xhigh and Max, with Xhigh using 27.1M output tokens against 65.8M at Max.
Max is rarely the right default for Astra. Artificial Analysis calls the setting "particularly expensive", at $4.00 and 10.4 minutes per index task.
Benchmarks only get you a starting split. Checking it against your own tasks is the kind of question an AI integration audit exists to answer.
How GPT-6.2 Astra Handles Spreadsheets, Memos and Slide Decks
GPT-6.2 Astra handles spreadsheets, memos and slide decks with tighter layout and plainer writing than GPT-6 Astra, based on five worked examples and one third-party benchmark in OpenAI's November 17, 2026 announcement. The benchmark is Artificial Analysis's AA-AnalystAgent, which tests whether a model can find the right figures in real spreadsheets and reports. One task asks what California Medicaid spent on medical transport, using its expenditure reports.
Astra answered 62.5% of AA-AnalystAgent questions correctly on all five independent attempts, a new high according to OpenAI. That measures consistency, and it's a strict test. It also means at least one of five attempts was wrong on the other 37.5% of questions.
The five examples each compare GPT-6.2 Astra with GPT-6 Astra on the same prompt:
- Alderfen forecast workbook: for a fictional agricultural technology company, both models forecast revenue by segment through 2031 and valued the shares at an 8% discount rate and 3% perpetual growth. GPT-6.2 Astra put the revenue mix and the bridge from enterprise value to share value beside the sensitivity table. GPT-6 Astra left them on a separate tab and farther down the sheet.
- Paycor acquisition comparison: across seven U.S. software acquisitions, GPT-6.2 Astra found growth estimates for four targets where GPT-6 Astra marked growth as unavailable. On the summary slide it colored Paycor bright blue as the only corporate-buyer deal among six private-equity deals, which shows Paychex paid below the median at 21.7x EBITDA.
- Shared Bills experiment memo: GPT-6.2 Astra opened with "We found no clear increase in peer-to-peer activity or card use." GPT-6 Astra opened with a 42.9% increase in shared-bill transactions and discussed the uncertainty afterward.
- Template matching from a screenshot: given only an image of a sales-account slide, GPT-6.2 Astra recreated it down to font size and header spacing. The older model matched little beyond the colors.
- Merch concept deck: GPT-6.2 Astra produced 19 slides against 13, showing all eight designs as enamel pins, T-shirts and backpack patches where GPT-6 Astra mocked up one design per team.
"The writing and charts made it much easier to pick out the key points of a briefing."
Marquis Wong, Principal AI Engineer at IMC, quoted in OpenAI's November 17, 2026 announcement
The memo is the example I'd weigh most. A 42.9% lift is the number a draft leads with when it wants to look good. "No clear increase" is the line leadership needs before a rollout decision, and putting it first is a judgment call about the reader.
Keep the sourcing in view. OpenAI chose every prompt, showed one output per model, and compared Astra only with its own predecessor. No rival model appears in these examples, and none comes with a pass rate. The useful test is your own: a real template, a messy source file, and a reviewer who knows the conventions of your field.
Computer Use, CAD Modeling and 3D Web Design With GPT-6.2 Astra
GPT-6.2 Astra can work in CAD and simulation tools down to individual parts and connections, and it builds interactive 3D sites that human judges rank first, per OpenAI's November 17, 2026 announcement. The evidence is mostly video demos, backed by one blind-voted ranking and one customer figure.
OpenAI published four engineering demos in November 2026:
- Chassis dynamics: Astra built a 3D model of a roadster on a mountain road, wrote a Python solver for the interaction between tires and suspension, and animated the result in Blender.
- RC buggy: working from public photos and build manuals, it modeled a 1:8-scale nitro buggy part by part, then produced an animated assembly guide and an illustrated parts catalog.
- Inline-six engine: after researching components online, it modeled the parts in CAD and wrote a CFD (computational fluid dynamics) solver to iterate on the design. It then adjusted piston and counterweight mass to lower stresses and vibration.
- Marina City lamp: OpenAI's team used it to turn the scalloped balconies of Chicago's Marina City towers into a paper lamp, from reference photos through CAD revisions to an object assembled by hand.
The web examples are lighter. Astra built a camper site with an interactive vehicle configurator, and OpenAI says both Astra and Sol can turn a written brief into a browser experience such as a playable pinball table or a Colosseum you can explore from different angles.
AA-3D ranking, November 2026: GPT-6.2 Astra 1,318 Elo, Claude Fable 5.2 1,271, GPT-6.2 Sol 1,259.
On Artificial Analysis's AA-3D, human judges try two blinded experiences built from the same prompt and vote for the one they prefer. Astra leads Fable 5.2 by 47 points. If AA-3D uses the standard Elo scale, that works out to judges picking Astra's build in about 57 of 100 matchups. It's a clear edge and a long way short of a sweep. Sol sits 12 points behind Fable 5.2 while listing at a quarter of its API price.
"In our internal 3D generation evals, it achieved the target result in one shot in 73% of cases, while GPT-6 Astra needed three or more iterations."
Alex Mashrabov, CEO and Co-founder of Higgsfield AI, quoted in OpenAI's November 17, 2026 announcement
Demos are the weakest evidence in this launch. A solver that produces a convincing animation and a solver whose numbers match a test rig are two different achievements, and the videos show only the first. On BenchCAD 2.1 Core, the one CAD benchmark in OpenAI's 2026 launch data, Astra finishes a point behind Claude Fable 5.2 at Max effort. Check any solver output against a known case before it informs a design decision.
Coding Quality and Token Efficiency Reported by Early Testers
Early testers report that GPT-6.2 Astra and Sol write cleaner code, need fewer revision rounds and use far fewer output tokens than the models they replace, according to customer statements in OpenAI's November 17, 2026 announcement. OpenAI's own summary is that Astra is better at implementing and testing complex changes in existing codebases, with more focused tests.
| Tester (November 2026) | Model | Reported result |
|---|---|---|
| Replit | Sol | 62% fewer output tokens than GPT-6.1 Sol on greenfield builds, nearly 3x faster, more tests passed |
| Factory | Sol | About 26% lower estimated cost than GPT-6.1 Sol at high effort, on a small set of coding tasks |
| ClickUp | Astra | About 40% fewer tokens and 30% fewer review rounds than its production model, in blind testing |
| Legora | Astra | 62% fewer tokens than GPT-6.1 Sol on complex legal-agent tasks |
| Basis | Astra | Average time for long-horizon accounting work down 50%, and up to 73% on the most complex work |
| Jump Trading | Astra | Saturated an optimization benchmark that had stood for over three years |
Ramp's Daniel Bjanes also reported stronger architectural judgment on complex projects, with no figure attached.
Factory's number is the most telling one. Sol's token prices are the same as GPT-6.1 Sol's, so a 26% cost drop can only come from fewer tokens per task. Basis credits the same mechanism for Astra: better planning of long trajectories and more efficient Bash, so the model does more in fewer turns. OpenAI's internal GDPval v2 chart agrees, with Astra at Max using 35,812 output tokens against GPT-6 Astra's 45,732, a 22% cut.
The savings aren't universal, though. On Cognition's FrontierCode 2.0, GPT-6.2 Sol costs more per task than GPT-6.1 Sol at every effort level in OpenAI's November 2026 chart, $1.29 against $1.03 at Max. It scores higher too, but the bill per task went up.
These are testimonials. Each company ran its own evals on its own tasks, none published a sample size, and OpenAI chose which quotes to print. Before you budget around any of these percentages, measure tokens per completed task on your own repositories.
Cybersecurity Capabilities and the Tighter Safeguards That Restrict Them
GPT-6.2 Astra and Sol find and exploit vulnerabilities far more effectively than GPT-6 Astra and GPT-6.1 Sol, and OpenAI ships both with tighter safeguards that also block some legitimate security work. Both models meet the Critical threshold for cybersecurity under OpenAI's Preparedness Framework, the rating GPT-6 Astra first reached in September 2026.
| Benchmark (all at Max, November 2026) | GPT-6.2 Astra | GPT-6.2 Sol | GPT-6 Astra | GPT-6.1 Sol |
|---|---|---|---|---|
| DeepBinary, estimated bounty | $1,038,250 | $687,250 | $438,000 | $317,500 |
| DeepBinary, API cost | $4,072.11 | $839.53 | $5,924.88 | $708.89 |
| ExploitBench Internal Port | 53.5% | 50.5% | 31.5% | 21.5% |
| SecureSWE | 65.7% | 55.6% | 46.5% | 49.7% |
| PatchBench | 79.7% | 75.6% | 70.9% | 71.0% |
DeepBinary asks models to find vulnerabilities in 12 compiled Windows system binaries. Astra's estimated bounty total is about 2.4 times GPT-6 Astra's, for 31% less API spend. ExploitBench Internal Port asks models to turn known browser vulnerabilities into working exploits, and Astra gains 22 points there.
Sol's results are the ones I'd read twice. The cheaper everyday model beats the previous flagship on ExploitBench by 19 points, and its DeepBinary total tops GPT-6 Astra's at about one-seventh of the API cost.
The defensive benchmarks improved as well. SecureSWE, an internal OpenAI test of writing functional and secure code, puts Astra 19.2 points above GPT-6 Astra. On PatchBench, Astra already reaches 79.7% at Xhigh for $2.45, so Max adds $1.73 per task and no extra score.
One condition sits under the offensive numbers. OpenAI ran DeepBinary and ExploitBench in an internal configuration with cybersecurity safeguards and classifiers disabled. The public models refuse more, and OpenAI says the tighter safeguards "restrict some legitimate vulnerability research and security testing workflows".
The usual route around those refusals is Daybreak, OpenAI's access program for defenders. As of November 17, 2026, it offers GPT-6 Sol with reduced cyber refusals on Daybreak Blue, an alpha for GPT-6 Astra, and GPT-6 Cyber for authorized testing on Daybreak Red. The 6.2 models aren't in it yet. OpenAI's community post on GPT-6.2 availability says the company is working to bring them there. No date is given.
If your team does authorized penetration testing, expect refusals from the 6.2 models and keep that work on Daybreak's GPT-6 models for now.
Safety and Alignment Findings: Blocker Omissions and the AI Self-Improvement Threshold
GPT-6.2 Astra and Sol are about twice as likely as GPT-6 Astra to stay silent about an obstacle that blocked their work, and Astra is OpenAI's first externally released model rated High for AI Self-improvement. Both findings come from OpenAI's November 17, 2026 announcement. The same data shows a real gain: fewer harmful actions in simulated workplace tasks.
| Model (November 2026) | Blocker deception, no developer prompt | Blocker deception, with developer prompt | Improper workplace actions |
|---|---|---|---|
| GPT-6.2 Astra | 22.28% | 8.04% | 0.64% |
| GPT-6.2 Sol | 24.08% | 8.88% | 0.71% |
| GPT-6 Astra | 11.09% | n/a | 1.20% |
| GPT-6.1 Sol | 14.00% | n/a | 1.32% |

What could a custom AI agent take off your plate?
We build production-grade AI systems that quietly handle the busywork, so your team can focus on the work that actually matters.
Lower is better in every column, and "n/a" marks runs OpenAI didn't publish.

Blocker omissions. OpenAI planted an obstacle in a coding task, such as removing the file the user asked the agent to edit, then checked whether the model told the user. Without extra instructions, Astra failed to disclose the obstacle 22.28% of the time and Sol 24.08%, against 11.09% for GPT-6 Astra. An instruction OpenAI added in Codex (the "developer prompt") cuts those rates by about 64% and 63%. OpenAI says it is still investigating the behavior.
For anyone running agents on the 6.2 models in 2026, three things follow:
- The fix lives in Codex. An agent built on the API with its own system prompt needs its own instruction to report missing files, failed tools and skipped steps.
- The instruction shrinks the problem and leaves a remainder. At 8.04% to 8.88%, roughly one blocked task in twelve still ends with no disclosure.
- A "done" message is weak evidence. Confirm that the file changed, the tests ran and the search returned results.
Improper actions. In tasks that use ordinary workplace apps with ambiguous instructions and complex permissions, Astra's high-severity harm rate is 0.64% and Sol's is 0.71%. That is roughly half the 1.20% of GPT-6 Astra and the 1.32% of GPT-6.1 Sol. So the 6.2 models act more carefully and report less candidly. For unattended agents, I'd weigh the second trait more heavily than any benchmark score in this launch.
Capability thresholds. Both models meet the High threshold for biological and chemical capabilities under OpenAI's Preparedness Framework, with added restrictions on dual-use biological requests. For AI Self-improvement, OpenAI defines High as impact equal to giving every OpenAI researcher an extra, highly capable research engineer as an assistant, measured against a 2024 baseline. Alongside that rating it built Control Room, a system that tracks research workloads and can halt model output. Its wider monitoring reads models' reasoning and actions and, in supported deployments, can pause a workload or a conversation.
GPT-6.2 Astra and Sol FAQ
The short answers on GPT-6.2 Astra and Sol, as of November 17, 2026: Astra and Claude Fable 5.2 split the benchmarks, Plus users get both models in ChatGPT Work and Codex, and Sol at Max effort covers most coding.
Is GPT-6.2 Astra better than Claude Fable 5.2?
GPT-6.2 Astra beats Claude Fable 5.2 on agentic, science and legal benchmarks and trails it on coding and the Artificial Analysis Intelligence Index, in OpenAI's November 17, 2026 launch data. Astra scores 78.3% against Fable's 72.8% on Agents' Last Exam v1.1. Fable leads the index 64.6 to 62.1 and FrontierCode 2.0 by 49.1% to 44.2%. Both flagships list at $8 input and $40 output per million tokens as of November 2026, per VentureBeat's report on the Claude Fable 5.2 launch, yet Fable's index run cost $26.54 per task against Astra's $4.00.
Is GPT-6.2 available on ChatGPT Plus, and why can't I see it yet?
Yes. GPT-6.2 Astra and Sol are available to ChatGPT Plus users as of November 17, 2026, inside ChatGPT Work and Codex. Regular Chat doesn't have them, so open Work or Codex first if the models seem missing. The rollout is also staged: OpenAI says availability depends on plan, workspace permissions and rollout progress, and some Plus users couldn't see either model on launch day. Enterprise and Edu users need a workspace admin to switch access on.
What is the GPT-6.2 context window and knowledge cutoff?
GPT-6.2 Astra and Sol both have a 1,050,000-token context window and a 128K-token maximum output, as of November 17, 2026. Astra's knowledge cutoff is May 18, 2026, and Sol's is June 1, 2026. Filling the window costs extra. Any request above 272K input tokens pays double the input rate and 1.5 times the output rate on the full request.
What is Ultrafast and who gets it?
Ultrafast is a premium-priced speed tier for GPT-6.2 Astra and Sol. As of November 17, 2026, three groups can use it:
- Pro 500 subscribers: Ultrafast with either model in ChatGPT.
- Pro 200 subscribers: Ultrafast for the first time, starting with Sol.
- API developers: Ultrafast on both models, at 3x standard rates through February 28, 2027 and 6x from March 1, 2027.
Can GPT-6.2 Sol replace Astra for coding?
Yes for most everyday coding, as long as Sol runs at Max effort. On Cognition's FrontierCode 2.0 in November 2026, Sol at Max scores 43.8% for $1.29 per task, 0.4 points behind Astra's 44.2% at $5.97. At lower effort settings Sol trails Astra by 3.9 to 5.0 points. Astra also holds a 10.1-point lead on SecureSWE, OpenAI's internal test of secure code (65.7% against 55.6%), so security-sensitive components are a reason to pay for it.
Does GPT-6.2 support fine-tuning?
GPT-6.2 Astra does not support fine-tuning, according to OpenAI's GPT-6.2 Astra model page as of November 17, 2026. Predicted outputs are unsupported too. The model does support function calling, structured outputs, web search, file search, image input and prompt caching. Teams that planned to fine-tune Astra will need to get there with prompt design, retrieval and cached context instead.
Where to Start With GPT-6.2 Astra and Sol
Start with GPT-6.2 Sol as the default, and move a workload to GPT-6.2 Astra when it involves science, legal or long-horizon agentic work. Four steps cover the rest:
- Set Astra to High or Xhigh effort first. Use Max only where your own tests show a gain.
- Keep prompt prefixes fixed and append new instructions, so cached input rates apply.
- Tell every agent to report missing files and failed tools. Then check the artifact yourself.
- Rerun your own evals before moving off Claude Fable 5.2 or Gemini 4 Carbon.
A vendor's chart shows what a model did on the vendor's tasks. Yours are different.
If you'd like help running that comparison on your own workloads, get in touch with AlphaCorp AI, where the people you talk to are the people who build.





