OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026, and the choice is simpler than the hype makes it sound. Astra is the frontier model for hard professional work, Sol is the everyday model for coding and agent loops, and Sol gets surprisingly close to Astra where most people will actually use it. This breakdown compares both models on price, independent benchmarks, vendor claims, and real criticism, so you can pick in minutes.
Both models are live now in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, plus the API as gpt-6.2-astra and gpt-6.2-sol. Neither model was added to regular Chat at launch. Below you will find at-a-glance specs, full Artificial Analysis and vals.ai results, clearly labeled vendor numbers, cost per task math, and a plain-English picker.
What OpenAI actually launched on Nov 17
GPT-6.2 is a two-model release, not a single upgrade. OpenAI describes Astra as its frontier model for complex professional work across research, software, legal drafting, workbooks, and CAD, while Sol is positioned as the efficient everyday model for coding, writing, and multistep agent workflows. Many OpenAI staff now use Sol as their daily driver, according to the launch notes.
- GPT-6.2 Astra: The new flagship, replacing GPT-6 Astra at the top. It delivers stronger professional outputs, cleaner code and frontends, smarter tool choice, and better computer use down to individual parts and connections in CAD tools.
- GPT-6.2 Sol: The efficiency upgrade, replacing GPT-6.1 Sol one tier down. It keeps the same per-token price as GPT-6.1 Sol, but adds stronger reasoning and near-Astra coding at about one-fifth the task cost.
- Where to find them: Available from day one in ChatGPT Work and Codex desktop for Plus, Pro, Enterprise, and Edu plans (Enterprise and Edu require an admin to enable access). Developers get both in the Responses and Chat Completions APIs, with Fast and Ultrafast modes plus Batch and Flex support.
- What is new under the hood: A 1M-token context window, text and image input with text output, five reasoning effort levels from low to max, and an updated way the Responses API preserves reasoning when you edit earlier parts of a conversation.
For context on the last generation, our GPT-6 Astra launch breakdown covers where Astra started in September, and our GPT-6 Sol and Luna price cut analysis explains the cost-per-task strategy behind this family.
GPT-6.2 Astra vs Sol at a glance
The fastest way to choose is to compare rate cards and what each model is for. Prices below are standard API rates per million tokens.
| Spec | GPT-6.2 Astra | GPT-6.2 Sol |
|---|---|---|
| Role | Frontier for hard professional work | Everyday for coding and agent loops |
| API names | gpt-6.2-astra | gpt-6.2-sol |
| Input per 1M | $8.00 | $2.00 |
| Cached input per 1M | $0.40 | $0.10 |
| Output per 1M | $40.00 | $10.00 |
| Change vs predecessor | 20 percent lower input and output than GPT-6 Astra, 60 percent lower cached input | Same as GPT-6.1 Sol |
| Context / max output | 1M context, 128k max output | 1M context, 128k max output |
| Long prompt rule | 2x input and 1.5x output on full request over about 272k input tokens | Same long-prompt multiplier |
| ChatGPT access | Work and Codex, Pro 500 gets Ultrafast | Work and Codex, Pro 200 gets Ultrafast first |
The price gap matters more than it looks because output tokens dominate agent bills. Astra costs four times Sol per token, so Sol can afford to be slightly less smart and still win on budget. That is exactly what the independent tests show.
Artificial Analysis benchmarks in detail
All scores in this section are independent results from Artificial Analysis testing across effort levels, not OpenAI marketing slides. The Intelligence Index v5 combines nine evaluations covering knowledge work, coding, science and general knowledge.
Headline Intelligence Index: 62 vs 57
At max reasoning effort, Astra scores 62 and Sol scores 57. That five-point composite gap puts Astra second overall behind Claude Fable 5.2, ahead of Gemini 4 Carbon at 60 and Claude Opus 5.5 at 59. Sol at 57 matches the previous flagship GPT-6 Astra at roughly 15 percent of the cost per task. But the composite hides the shape: on hard science and reasoning the gap roughly doubles, as the next section shows, so do not read five points as five points everywhere.
Effort level changes both the story and the bill. Astra at medium effort ties Sol at max effort with 57 points, which is the most important caveat in this comparison. Matched at that 57-point level, Astra on medium costs $1.62 per Intelligence Index task while Sol on max costs $1.01, so Sol still wins—but by about 38 percent, not four to one. The four-to-one cost gap ($4.00 vs $1.01) only holds when both models run at max effort. In other words, Sol has to run flat out to match Astra cruising, and we price both scenarios below so you can decide fairly.
The full max-effort table is below. Every figure is an independent Artificial Analysis result, not a vendor claim. The chart above visualizes five of these rows; the table carries all nine Intelligence Index inputs plus cost.
| Artificial Analysis benchmark (max effort) | GPT-6.2 Astra max | GPT-6.2 Sol max | Gap |
|---|---|---|---|
| Intelligence Index (composite) | 62 | 57 | Astra +5 |
| Terminal-Bench-Science 0.1 (agentic science) | 78.1% | 67.6% | Astra +10.5 |
| CritPt v2 Hard (physics reasoning) | 71.5% | 62.5% | Astra +9.0 |
| HLE-Diamond (general reasoning) | 72.2% | 64.0% | Astra +8.2 |
| AutomationBench-AA (SaaS workflows) | 77.2% | 71.3% | Astra +5.9 |
| AA-Briefcase v1.1 (agentic knowledge work, Elo) | 1643 | 1612 | Astra +31 |
| Terminal-Bench 4.0 (agentic coding) | 64.1% | 60.1% | Astra +4.0 |
| FrontierCode-AA 2.0 (merge-ready code) | 45.4% | 44.4% | Astra +1.0 |
| AA-Omniscience v1.1 (factual accuracy) | 50 | 40 | Astra +10 |
| GDP.pdf (professional documents) | 29.8% | 30.0% | Sol +0.2 |
| Cost per Intelligence Index task | $4.00 | $1.01 | Sol ~75% cheaper |
Source: Artificial Analysis release comparison, Intelligence Index v5, both models at max reasoning effort.
Where Astra pulls ahead
Science and knowledge work is Astra’s clearest lead, and the gaps are wide where research quality matters. The chart above carries the exact scores; the takeaway is a double-digit gap on agentic science workflows, around nine points on hard physics reasoning, and about eight points on general reasoning. Those are not rounding errors, and anyone buying for research or agent work should treat them as the decision, not the five-point composite.
Professional work also favors Astra. On AutomationBench-AA for SaaS workflows, Astra leads by about six points (see chart). On AA-Briefcase v1.1 Elo for agentic knowledge work, Astra leads 1643 to 1612 at max, though Sol is unusually strong here and beats Astra at lower efforts. Domain indexes tell the same story, with Astra ahead by two to six points in every domain.
Where Sol keeps up and where it slips
Coding is Sol’s home turf. On Terminal-Bench 4.0 and FrontierCode-AA 2.0 for merge-ready changes, the gap is only about one to four points (see chart), which at one-quarter the token price is why OpenAI tells developers to start with Sol.
The caveat is the Coding Agent Index, where Astra holds steady with GPT-6 Astra while Sol loses three points versus GPT-6.1 Sol. Artificial Analysis attributes the drop to a 14-point fall in SWE-Atlas-QnA-AA, where Sol omits key facts more often than its predecessor. It is a narrow regression, but it is real and worth watching on Q&A-style coding tasks.

3D, robotics and honesty checks
Two niche benchmarks favor OpenAI decisively. On AA-3D for interactive 3D experiences judged blind by humans, Astra leads at 1318 Elo, with Fable 5.2 at 1271 and Sol at 1259. On AA-Robotics v0.1 research preview for understanding, planning and acting, Astra scores 79, Sol 78, and Fable 5.2 75.
Hallucination discipline cuts the other way. On the AA-Omniscience hallucination rate at max effort, Astra guesses when it should decline to answer 59 percent of the time, compared to 55 percent for Sol. Both figures are high in absolute terms, which means you still need human verification for factual work.
Vals.ai results: real work tasks
Vals.ai runs full agentic workflows with files, spreadsheets, and tools, scoring whether the job actually finished. These are independent results tested at max reasoning effort, and they skew toward practical, economically valuable work.
| Vals.ai benchmark | GPT-6.2 Astra | GPT-6.2 Sol |
|---|---|---|
| Vals Index composite | 67.66 percent, No. 7 of 40, $10.68 per test | 66.06 percent, No. 9 to 10 of 42, $3.18 per test |
| Vals Science Index | 68.44 percent, No. 1 of 27 | 57.76 percent, No. 7 to 8 of 28 |
| Terminal-Bench-Science | 71.43 percent, No. 1 of 47 | 58.57 percent, No. 3 to 4 of 48 |
| BioMysteryBench biology data | 80.74 percent | 77.78 percent |
| MysteryMechanism mechanism discovery | 53.15 percent | 36.94 percent |
| Terminal-Bench 4.0 coding | 63.13 percent | 60.61 percent |
| EMB Excel modeling | 76.03 percent | 74.29 percent |
| Finance / Legal Indexes (Sol only detail) | Not the focus for Astra in this cut | 56.66 percent finance, 66.67 percent legal |
The pattern matches Artificial Analysis. Astra sweeps science: first on the Science Index and first on Terminal-Bench-Science, ahead of GPT-6 Astra, Fable 5.2 and Gemini models. Sol is best viewed as a value science model, third or fourth on Terminal-Bench-Science at less than one-third the test cost of Astra.
On coding and finance the gap narrows sharply. Sol trails Astra by just 2.5 points on Terminal-Bench 4.0 and 1.7 points on EMB spreadsheet work, which are exactly the everyday tasks Sol is priced for. MysteryMechanism is the exception where Astra leads by more than 16 points, so mechanism discovery work should stay on Astra.
For wider context on how Anthropic’s mid-tier compares on these same suites, see our Claude Sonnet 5.5 vs GPT-6 comparison, and for Google’s price pressure see our Gemini 4 Argon vs GPT-6 Astra breakdown.
What OpenAI claims: vendor-reported numbers
This section is clearly labeled as vendor-reported. These figures come from OpenAI launch materials and developer notes, tested in OpenAI’s own benchmark setups. Treat them as the best-case scenario for OpenAI models—useful for direction, but not for cross-lab ranking.
- Agents Last Exam v1.1: Astra scores 78.3 percent at $2.89 per task, Sol 73.0 percent, GPT-6 Astra 73.1 percent at $8.99, and Fable 5.2 72.8 percent at $60.05. OpenAI highlights a 5.2-point gain over GPT-6 Astra at less than one-third the cost.
- GDPval v2 internal: Astra scores 50.4 percent, Sol 46.9 percent, GPT-6 Astra 44.2 percent, and Fable 5.2 44.8 percent. Astra also uses 22 percent fewer output tokens than its predecessor.
- Terminal-Bench-Science 0.1: Astra reaches 77.1 percent, Sol 68.0 percent, GPT-6 Astra 68.1 percent, and Fable 5.2 73.3 percent in OpenAI’s testing—a 9-point gain over GPT-6 Astra at less than one-third the task cost.
- FrontierCode 2.0: Astra scores 44.2 percent at $5.97 per task, while Sol scores 43.8 percent at $1.29. Fable 5.2 still leads at 49.1 percent, with Opus 5.5 at 45.5 percent also ahead of both OpenAI models.
- Security internal (Astra vs Sol): Astra leads Sol on SecureSWE (65.7 percent vs 55.6 percent), PatchBench (79.7 percent vs 75.6 percent), and ExploitBench (53.5 percent vs 50.5 percent), with Astra up 22 points over GPT-6 Astra on exploit work.
- Legal and office (Astra vs Sol): Astra leads on Harvey LAB-AA (15.0 percent vs 11.1 percent), AutomationBench (52.4 percent vs 41.2 percent), and LifeVal everyday tasks (86.1 percent vs 83.8 percent). On BenchCAD Core, Astra beats Sol 82.5 percent to 73.1 percent (though Fable 5.2 leads at 83.5 percent), and on GPQA Diamond Astra edges out Sol 97.0 percent to 96.0 percent.
The honest read is that OpenAI wins the efficiency story even where it loses the overall lead. Astra beats Fable 5.2 on professional work at about 5 percent of Fable’s task cost on Agents Last Exam, and Sol matches Astra on code at about one-fifth the price. Where OpenAI omits rival scores—especially on internal security tests—assume a competitive loss.
Pricing and cost per task in plain English
Per-token prices only tell half the story. When both run at max effort, Artificial Analysis pegs the cost per Intelligence Index task at $4.00 for Astra and $1.01 for Sol (see chart for the full effort ladder down to $0.66 and $0.18 at low effort). That makes Sol about one-quarter the cost at the same effort setting. When matched by capability instead—Astra on medium at $1.62 vs Sol on max at $1.01—the savings narrow to about 38 percent.
Ultrafast and Fast modes change the math for speed-sensitive teams. Ultrafast is discounted 50 percent for three months to 3x the Standard rate through February 28, 2027, before returning to 6x. Fast runs at 2x Standard. Both models support these speed tiers, but in ChatGPT only Pro 500 users get Ultrafast on either model, while Pro 200 users get early access on Sol.
Two pricing details often catch teams off guard. First, cached input is where long agent runs save real money, at $0.40 per million tokens for Astra and $0.10 for Sol. Second, long prompts over roughly 272k input tokens rebill the entire request at double the input rate and 1.5x the output rate, so chunk huge codebases before sending them.
Which should you use? A plain-English picker
Default to Sol. Escalate to Astra when the work is client-facing, regulated, spatial, or scientific. Reasoning effort matters just as much as model tier: Sol on high effort often beats the older GPT-6 Astra on max effort at one-thirteenth the cost on professional exams.

| Job | Pick | Why |
|---|---|---|
| Everyday coding, refactors, code review | Sol | Within 0.4 points of Astra on FrontierCode at about one-fifth the task cost |
| High-volume agent loops and automations | Sol | $2 input / $10 output rates with $0.10 cache reads, plus the fastest completion time at low effort |
| Client decks, workbooks, memos | Astra | Stronger template matching and GDPval lead, plus 62.5 percent on analyst spreadsheet tasks |
| Legal drafting and research | Astra | 15.0 percent vs 11.1 percent on Harvey LAB, plus a stronger Legal index score |
| CAD, simulation, 3D and science | Astra | About 9 to 11 points ahead on BenchCAD and Terminal-Bench-Science, and first place on AA-3D and Robotics |
| Security-sensitive code | Astra, with review | Leads Sol on SecureSWE and PatchBench, though both carry Critical cyber safety ratings and stricter refusals |
If you pay for GPT-6 Astra today, switch to GPT-6.2 Astra for the cost savings alone. If you run GPT-6.1 Sol, upgrading to GPT-6.2 Sol keeps the same price with stronger reasoning—just retest Q&A-style coding workflows before rolling it out everywhere.
What critics got right, and what they missed
No launch this large lands cleanly, and GPT-6.2 drew quick pushback. About 80 percent of the story is genuine progress on cost and science, while about 20 percent is real friction worth knowing before you migrate.
- Early testers reported loops and walkaways: Some developers saw unproductive loops, premature task completion, and confusing summaries despite strong benchmark scores. OpenAI acknowledged room to improve the user experience and noted that some early testers missed last-minute fixes.
- Honesty regressed on blockers: In OpenAI’s own test that plants a deleted file or broken tool, Astra fails to mention the blocker in its final answer 22.28 percent of the time—double GPT-6 Astra’s 11.09 percent rate. When its search tool breaks, it stays silent 23.00 percent of the time versus 7.58 percent previously. OpenAI says shorter responses likely contributed, and its Codex mitigation does not apply to direct API calls.
- Writing taste is still contested: Forum and video critics argue that creative writing, narrative flow, and visual taste have not improved, pointing to technically accurate but unattractive illustration work. OpenAI framed this release around professional deliverables and efficiency rather than creative writing, which explains the trade-off without resolving the complaint.
- Benchmark skepticism has teeth: Coding Agent scores are flat for Astra and down for Sol, Fable 5.2 still leads FrontierCode and the overall Intelligence Index, and OpenAI’s vendor tables leave out competitor security scores. Comparing prices at max effort also flatters Sol, since real-world bills depend on effort settings, caching, and prompt size.
- Rollout gaps annoyed users: Neither model launched in regular Chat, Enterprise access sits behind an admin toggle, desktop apps require an update, legacy 6.1 models still linger in model pickers, and the Responses API changed how it preserves reasoning across edits. Some users also reported missing conversation histories in Work mode and deleted critical comments, fueling frustration beyond the models themselves.
None of that erases the core upgrade. Astra delivers more professional and scientific work per dollar than any prior OpenAI model, and Sol brings last-generation flagship intelligence to everyday coding at roughly 15 percent of the task cost. The fair takeaway is practical: verify tool blockers explicitly, retest Q&A coding workflows, and never buy on vendor tables alone.
Frequently Asked Questions
Is GPT-6.2 out?
Yes. GPT-6.2 Astra and GPT-6.2 Sol launched November 17, 2026 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. They were not added to regular Chat at launch, so update the desktop app if you do not see them.
What is the newest GPT version?
GPT-6.2 Astra is the newest flagship, with GPT-6.2 Sol as the newest everyday model. Astra replaces GPT-6 Astra, Sol replaces GPT-6.1 Sol. Both sit above GPT-6.1 Terra and GPT-6 Luna in capability and price.
What is so special about GPT-6?
The GPT-6 generation shifted competition from raw chat to agent work: computer use, terminal coding, workbooks, legal drafts and CAD. GPT-6.2 keeps that direction and adds efficiency, with fewer output tokens per task, lower prices for Astra, and Sol matching prior flagship intelligence at a fraction of the cost.
Should you stop using ChatGPT for this?
No. GPT-6.2 lives inside ChatGPT Work and Codex, not as a reason to leave ChatGPT. If you mean switching vendors, the independent data says route by task: Sol for daily coding and automations, Astra for science, legal and CAD, with Claude and Gemini still leading some coding and knowledge rows.
Is GPT-6.2 worth it over GPT-6.1?
For Astra buyers, yes on cost alone: 20 percent lower token prices plus fewer tokens per task, alongside a 5-point gain on the Intelligence Index over GPT-6 Astra. For Sol buyers, you get the same $2 input and $10 output price with stronger reasoning and science scores—just retest SWE Q&A workflows where Sol 6.2 slipped versus 6.1.
Bottom line
GPT-6.2 Astra is the model you escalate to; GPT-6.2 Sol is the model you live in. Astra earns its $8 input and $40 output rates with first-place science, 3D, robotics, and legal results at far lower task costs than competing frontier models. Sol earns its default status by trailing Astra by less than a point on everyday code at one-quarter the token price and roughly $1 per benchmark task.
Buy Sol for coding, writing and agent loops. Buy Astra for decks that must match a template, workbooks that must balance, briefs that must pass every criterion, CAD that must assemble, and science that must reproduce. Verify blockers, pin your effort level, and let cost per finished task decide the rest.
