I have a rule for model launches: I don't switch my workflow on launch day. Too many benchmark charts have looked incredible on Tuesday and failed by Thursday. So when OpenAI dropped GPT-6.2 Astra and Sol on November 17, 2026, I read the announcement, then the reactions, then looked for the gap between the two.
That gap is real. On paper, it's one of OpenAI's strongest releases: a frontier model that got cheaper, a budget model nearly matching it on coding tests, and the first model OpenAI rates "High" for AI self-improvement. In practice, some power users who tested it before launch call parts of it "essentially unusable." Both can be true — and if you build WordPress sites, this launch matters to your bill and your workflow. Here's what I found, what I'd trust, and what I'd wait on.
What OpenAI Actually Shipped on November 17
The release is two models, and the split is the whole point. GPT-6.2 Astra is the frontier model, positioned for the hardest work: changing a complex codebase, building a working web application, or producing a document that must follow a company's templates exactly. It succeeds GPT-6 Astra from September 3.
GPT-6.2 Sol is the everyday model for routine coding, writing, and multi-step workflows, succeeding GPT-6.1 Sol from September 29.
Both are available now to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and to developers through the API as gpt-6.2-astra and gpt-6.2-sol. One catch: Enterprise and Edu workspaces need an admin to switch them on.
Both share the same limits: a 1.05 million-token context window and up to 128,000 output tokens. That context size matters most to me, because it lets you hand a model a whole plugin or theme instead of one file at a time. OpenAI also says Astra follows professional conventions more reliably — law, finance, accounting. The WordPress version is structured content: specs and accessibility statements, where wrong format equals wrong answer.
The Price Cut Is the Real Story
What caught my attention first wasn't a benchmark — it was the price. GPT-6.2 Astra costs $8 per million input tokens, $0.40 cached, and $40 output at standard API rates: 20% lower than GPT-6 Astra on input and output, with cached input down 60%. A frontier model getting cheaper is unusual.
Sol keeps GPT-6.1 Sol's prices exactly: $2 input, $0.10 cached, $10 output per million. So Sol costs a quarter of Astra — the most important number here for anyone paying API bills, because benchmarks say Sol is much closer to Astra in ability than in price.
One detail changes the math on long jobs: past 272,000 input tokens, OpenAI doubles input and cache rates and multiplies output by 1.5 for the whole request — Astra becomes $16 in and $60 out per million, Sol $4 and $15. The giant context window is real, but filling it isn't free. Batch runs at half standard price, Fast at twice, Ultrafast at three times until March 1, 2027, then six times; Ultrafast is live for Pro 500, coming to Pro 200 starting with Sol.
The comparison that frames the launch: Astra's list price now matches Anthropic's Claude Fable 5.2 exactly — $8 in, $40 out. OpenAI says the fight is now about tokens burned per job, not sticker price — harder to verify, worth skepticism. But if it's even mostly true, the effective gap is bigger than the price list suggests. For a WordPress agency: per-task cost just dropped switching from GPT-6 Astra to 6.2, and routine jobs get a better model at the same money.
The Benchmarks Look Great — Read the Fine Print
OpenAI's launch post has nine head-to-head charts. Astra beats GPT-6 Astra on all nine; Sol beats GPT-6.1 Sol on eight. The exception is Sol on the Vals Science Index, down slightly from 59.6% to 58.5% — a small regression worth noting, since vendors rarely publish their own.
The headline numbers, all OpenAI-reported: on Agents' Last Exam (multistep professional work), Astra scores 78.3% vs 73.1% at under a third of the cost, and leads Fable 5.2 by 5.5 points at 5% of the cost. On GDPval v2, Astra gains 6.2 points while using 22% fewer output tokens. On Harvey LAB-AA (legal work, counted only with every requirement met and no material hallucinations), Astra more than doubles Fable 5.2's pass rate at one-sixth of the cost. On FrontierCode 2.0 (whether code meets real maintainers' merge bar), Sol comes within 0.5 points of Astra at about one-fifth the cost. On AA-3D (human judges comparing blinded 3D experiences), Astra leads at 1,318 Elo, ahead of Fable 5.2 at 1,271 and Sol at 1,259. Terminal-Bench-Science: Astra +9 points at under a third of the cost; Vals Science: 68.4% vs 65.1%.
Now the fine print: several are OpenAI's own benchmarks, run by OpenAI on OpenAI's terms — GDPval is internal, Agents' Last Exam is OpenAI's test. When a vendor claims it beats the rival at 5% of the cost, the right response is "show me the independent test." OpenAI's footnotes admit Fable's leaderboard results leave out nine tasks that used a fallback model.
Independent scores are more modest. On the Artificial Analysis Intelligence Index (nine evaluations across knowledge work, coding, science, and general knowledge), Astra scores 62 — second behind Fable 5.2 at 65. Epoch AI scored it 174 on its Capabilities Index, tying Fable 5.2. Honest shape: near the top, wins some tests outright, doesn't sweep the board. A good launch, not a coronation.
The Reception Has Been... Complicated
VentureBeat's coverage leads with the opposite of the press release: several power users who tested the models before launch are sharply critical — the headline phrase is "essentially unusable." The complaints are concrete: missed instructions, compressed explanations, agents that stop before finishing.
Bloomberg adds a wrinkle: its headline notes OpenAI unveiled the new Astra after scrapping a prior release. I don't know what was scrapped or why, but a flagship launch following a discarded build says something about this cycle's rush — three flagship-adjacent releases in under three months means real progress and real rough edges.
To be fair, early-tester complaints aren't unusual and aren't a verdict — power users push models to the edge of multi-hour agentic work, where every model stumbles. But the pattern matters: "stops before finishing" and "misses instructions" are the two failure modes that cost a WordPress developer most, turning ten minutes of automation into twenty of debugging. My read: benchmarks probably real, price cut definitely real, agentic reliability genuinely unproven. Test on your own work — small jobs first — before moving anything a client pays for onto the new models.
What This Means for WordPress and Elementor Work
The most direct impact is through Codex, where both models are live now — scaffolding a custom plugin, refactoring a theme's functions file, writing a migration script, debugging an Elementor widget that breaks on the front end. Astra is for the hard jobs: a complex refactor across a whole plugin, a WooCommerce customization touching checkout, a performance problem where the fix must be right first time. Sol is for everything else.
Here's the split I'm using. Sol at $2 in and $10 out is cheap enough to be generous with: first plugin drafts, boilerplate, client docs, routine jobs. Astra at $8 and $40 is the escalation — when Sol's answer isn't good enough, or a wrong answer costs more than the tokens saved. Cheap model first, frontier model as escalation: that's how you capture this launch's value. Run everything on Astra and you pay 4x for jobs Sol would have done fine; run everything on Sol and you'll hit its ceiling.
There's a less obvious angle. Elementor just shipped Agent Ready features in version 4.4, including llms.txt support and tools letting AI agents read and work with your site. WordPress is moving toward agents that don't just help build pages — they read your site, understand it, and act on it. A cheaper, stronger coding model arriving the same month is the other half of that shift.
Sol Is the Model I'll Actually Use Most
Astra is the headline, but Sol changes my week. The FrontierCode 2.0 result says it best: Sol within half a point of Astra on whether code changes would actually be merged by real maintainers — at about one-fifth of the cost. On LifeVal, OpenAI's everyday-tasks test, Sol beat the older GPT-6 Astra at about a seventh of the cost. Those are OpenAI's numbers — but even if the real gap is twice as wide, Sol is still the value pick for routine work.
For a WordPress developer, "routine work" is the bulk of the job: a custom post type, a shortcode, a filter changing WooCommerce checkout text, a functions.php grown organically for three years — jobs where the difference between a frontier model and a good everyday model is invisible in the output and very visible in the invoice. My plan: Sol as my Codex default for day-to-day WordPress work; Astra as the escalation when I find myself re-prompting three times. I'll know within a week whether that split holds — from my own projects, not a benchmark chart.
The "High Self-Improvement" Rating Deserves a Calm Read
One line will be quoted out of context for the next month: GPT-6.2 Astra is the first model OpenAI has rated "High" for AI self-improvement on its own internal scale — dramatic-sounding, and the kind of phrase that generates alarming headlines.
Here's what it means, as far as public reporting goes: OpenAI's internal evaluation places Astra at the "High" threshold on a scale measuring capability to contribute to AI research and development. It's a capability rating, not a report that the model is doing anything on its own — a meaningful milestone.
What it does not mean is that a model somewhere is rewriting itself in the background. Capability ratings are about what a model could do with the right setup and access, and OpenAI publishes them so safety work keeps pace with capability work. If you build WordPress sites, the practical consequence is close to zero today — it matters for the long-term AI safety conversation, not for how you use ChatGPT or Codex tomorrow.
How to Try GPT-6.2 Today
Plus, Pro, Enterprise, and Edu users should find both models in ChatGPT Work and Codex now. On Enterprise or Edu, ask your workspace admin if they're missing. For the API, the IDs are gpt-6.2-astra and gpt-6.2-sol, with the 1.05M-token context and up to 128K output tokens.
Remember the long-context tier: past 272,000 input tokens, input and cache rates double and output rises 1.5x for the whole request — run the cost math before sending a large codebase, or use batch processing at half price.
How I test every new model: never on a client project — my own site or a staging copy. Give it a job you've already done with the old model, for a baseline. Test the failure modes, not just the successes: a genuinely multi-step job — does it finish? A detailed instruction — does it follow all of it? The early complaints are about exactly those two things. And watch token usage for the first week: a cheaper model that burns twice as many tokens is not a cheaper model.
Conclusion
GPT-6.2 is a genuinely good launch with an uncertain edge. The price cut is real and helps everyone who pays for AI-assisted development. Independent scores put Astra near the top — second on the Artificial Analysis index, tied on Epoch's, first on Vals Science — without sweeping the board. Sol looks like the value pick of the year for routine coding: a quarter of Astra's price, within half a point on the coding test that matters most. And it lands the same month WordPress and Elementor opened their doors to AI agents.
But the early reception is a warning: missed instructions, agents that stop before finishing, a launch following a scrapped prior release — that deserves a test drive, not a blind migration. My approach is the boring one: Sol as your default for everyday WordPress work, Astra as the escalation for hard jobs, both tested on your own projects before they touch a client's site. Let your own results, not a benchmark chart or a complaint thread, decide where they fit.
Frequently Asked Questions
When did OpenAI release GPT-6.2 Astra and Sol?
Tuesday, November 17, 2026. Both are available now to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and to developers through the API as gpt-6.2-astra and gpt-6.2-sol.
How much do GPT-6.2 Astra and Sol cost?
At standard API rates, Astra costs $8 per million input tokens, $0.40 cached, and $40 output — 20% cheaper than GPT-6 Astra. Sol keeps GPT-6.1 Sol's prices: $2 input, $0.10 cached, $10 output per million. Past 272,000 input tokens, input and cache rates double and output rises 1.5x for the whole request.
Which is better for WordPress development, Astra or Sol?
For most day-to-day WordPress work — plugin scaffolding, theme edits, shortcodes, WooCommerce tweaks — Sol is the better value, within 0.5 points of Astra on the FrontierCode 2.0 coding benchmark at about one-fifth of the cost. Use Astra for the hard jobs: complex refactors, checkout customizations, work where a wrong answer costs more than the tokens saved.
Is GPT-6.2 Astra better than Claude Fable 5.2?
It depends on the test. On OpenAI's own benchmarks Astra leads Fable 5.2 on several, including Agents' Last Exam and Harvey LAB-AA. On the independent Artificial Analysis Intelligence Index, Fable 5.2 leads 65 to 62. Astra leads Vals Science, and the two tie on Epoch AI's Capabilities Index at 174. List prices now match exactly, so the real comparison is tokens used per task.
What does OpenAI's "High" self-improvement rating for Astra mean?
It's an internal capability rating: Astra is the first model OpenAI has released that reaches the "High" threshold for AI self-improvement — capability to contribute to AI research and development. It is not a report that the model is modifying itself. It matters for the long-term AI safety conversation, not for how you use ChatGPT or Codex day to day.
Are there known problems with GPT-6.2?
Some early power users who tested the models before launch reported missed instructions, compressed explanations, and agents stopping before finishing — VentureBeat's coverage led with "essentially unusable." OpenAI's own data also shows one small regression: Sol slipped on the Vals Science Index, from 59.6% to 58.5%. Test on your own staging sites before using them on client work.
Sources: OpenAI launch announcement (via The Decoder, The Parameter, VentureBeat, Bloomberg, The Verge, MadRobot, Kingy.ai), Artificial Analysis Intelligence Index, Epoch AI Capabilities Index. Benchmark figures are OpenAI-reported unless noted as independent.
