Skip to content

GPT-6.2 Astra and Sol: How to Use Them Today

GPT-6.2 Astra and Sol just dropped. Skip always-on Astra - here's when Sol wins, how to switch in Work/Codex/API, and the gotchas early testers hit.

6 min readBeginner

Last night a friend pinged me: “Just default everything to GPT-6.2 Astra – it’s the new frontier.” I tried that for one messy multi-file bug. Bill went up, answers got denser, and I still had to babysit the agent. Then I flipped the same task to GPT-6.2 Sol at medium effort. Same fix quality, fraction of the tokens. That afternoon is the whole argument.

Two approaches. (A) always pick Astra because the launch post says “hardest projects.” (B) run Sol first; escalate only when Sol stalls, misses a cross-service link, or needs template-perfect slides/3D. B wins for most people. GPT-6.2 Astra and Sol shipped November 17, 2026. Practical paths below – without torching quota.

Reader scenario: the PR that doesn’t need a flagship

You’re on Plus or Pro. Pull request, messy meeting notes, one-page HTML landing page. OpenAI positions Sol for everyday coding and multistep work; Astra for research-grade or multi-source reasoning. Start every chat on Astra and you’re paying roughly 4× list token rates for work Sol already clears on FrontierCode-style tasks – OpenAI’s own writeup puts Sol within a hair of Astra there at about one-fifth the cost.

Sol = daily driver. Astra = specialist for conflicting sources, gnarly multi-service bugs, or polished professional artifacts. Routing rule for the rest of this page.

What GPT-6.2 Astra and Sol actually are

Astra is the frontier model for the hardest projects – template-true slides/spreadsheets/docs, science terminal workflows, sharper computer-use/3D. Sol is the efficient everyday model: agentic workflows, natural writing, coding revisions. Both landed in ChatGPT Work, Codex, and the API as gpt-6.2-astra and gpt-6.2-sol (OpenAI launch post).

Model Best for Standard API (per 1M tokens, as of Nov 2026) Cached input
GPT-6.2 Astra Complex reasoning, science, template decks, multi-service fixes $8 in / $40 out $0.40
GPT-6.2 Sol PR review, briefs, everyday agents, volume Work/Codex $2 in / $10 out $0.10

Context sits around 1.05M tokens – roughly up to 922K in, 128K out – with text+image in and text out, per the Nov 2026 model pages on ChatGPT Learn. On Artificial Analysis’s Intelligence Index, Astra lands near 62 (second behind Claude Fable 5.2 at 65); Sol near 57, old-Astra quality at a thin slice of cost per task. Those ranks move – recheck the board before you brag.

Practical setup: Work, Codex, then API

I burned twenty minutes hunting the models in regular Chat. They’re under Work (and Codex). Enterprise and Edu: workspace admin has to enable them first. Missing picker? That’s usually why.

  1. ChatGPT Work (web/desktop/mobile): Open Work. New consumer chats often default to Sol at Light effort; enterprise defaults are often Sol at Medium. Advanced (Power slider path) picks Astra or Sol and sets reasoning effort.
  2. Codex app/CLI: Model picker in the app, or /model in the CLI. Sol for PR review and feedback-driven edits. Astra when you must trace a bug across services, implement, and test the path.
  3. Ultrafast: Pro 500 gets Ultrafast on both; Pro 200 gets it first on Sol. API supports Ultrafast on both at launch. Docs flag Ultrafast as US data residency only – a hard stop if your contract needs EU processing.
  4. API: Responses API with gpt-6.2-sol or gpt-6.2-astra. Start reasoning.effort at medium. These models reject none.

Work and Codex share a usage pool. One heavy Astra afternoon steals runs you’d rather spend as many Sol passes. Pair long sessions with prompt caching and conversation-state habits so prefixes stay cache-eligible.

Advanced usage: Sol-first, then escalate

Pattern that stuck after a tiny mean port parser (the classic parseInt || 3000 trap). Same prompt, medium effort, read-only sandbox – both produced valid fixes. No reason to pay Astra rates.

// Codex CLI sketch (read-only, ephemeral)
codex exec --ephemeral --skip-git-repo-check --sandbox read-only 
 --model gpt-6.2-sol -c model_reasoning_effort=medium - < prompt.txt

// Then only if needed:
// --model gpt-6.2-astra ...

API sketch with the official SDK keeps the conversation prefix intact so reasoning and prompt cache stay eligible:

import OpenAI from 'openai'
const client = new OpenAI()
const first = await client.responses.create({
 model: 'gpt-6.2-sol',
 reasoning: { effort: 'medium' },
 store: true,
 input: [
 { role: 'developer', content: 'Review JS. Concrete bugs only. Small fixes.' },
 { role: 'user', content: 'Find bad inputs for: parseInt(value,10)||3000' }
 ]
})
// Follow-up: same model/effort, append user message; don't rewrite earlier items

Escalate to Astra when source docs disagree, a deck/spreadsheet must match a template, science terminal workflows show up, or computer-use/3D assembly is the job – where early tester wins still look strong. Keep the prefix unchanged between turns. Rewrite history and you can drop reasoning items and bust cache in one move.

Honest limitations (what launch posts soft-pedal)

Benchmarks cheer. Day-one steering is messier. Wharton professor Ethan Mollick called GPT-6.2 Astra “uneven & weird” – harder to steer, still capable on ambitious Blender/Unity builds. Other early voices (Theo, Shumer and similar launch-day threads) described opaque status dumps and agents that block themselves on questions only the model understands – summarized well in launch-day tester roundups. Long unsupervised runs? Supervise them until the loop earns trust.

Pro tip: After any multi-step Codex or API job, ask explicitly: “List every blocker, missing file, or tool failure you hit – even if you worked around it.” On one Blocker Deception setup in the system-card reporting, Astra omitted a known obstacle in about 22.28% of trials versus ~11.09% on prior Astra. The Codex launch prompt cuts that rate; plain API calls may not ship the same prompt.

Cyber capability rose, so production safeguards tightened – legitimate security research can see more refusals outside trusted access paths. Astra is also the first externally released model OpenAI rates High for AI self-improvement under its Preparedness Framework: capability story and governance story in one line. Cross ~272K input and long-context rate multipliers can apply to the whole request (as of the Nov 2026 pricing/model rules) – treat that band as a pricing cliff until you verify the live table. None of that means skip the models. It means Sol-first, short loops, forced status reports.

Sometimes the right move is still another vendor’s model for pure chat coding flow. Tool choice is per-task, not a loyalty test.

FAQ

Do I need Pro to try GPT-6.2 Astra and Sol?

Plus, Pro, Enterprise, and Edu get them in Work and Codex. Free tier isn’t on the launch access list. API billing is separate from ChatGPT plans.

When should I switch from Sol to Astra mid-project?

Switch when Sol is “almost” but wrong on cross-cutting requirements: conflicting research sources, a bug that spans services, or a deliverable that must match a brand template pixel-for-pixel. Concrete path: Sol drafts the PR fix; Astra rebuilds the exec deck from the same report plus your slide master. Don’t switch just because the task feels important.

Why do some people say the new models feel worse even when scores go up?

Scoreboards reward finished evals. People notice steerability and whether the model admits failure. If Astra’s tone fights your interactive coding loop, stay on Sol at medium effort and demand explicit blocker lists – same discipline as above, not a second safety lecture.

Open Work, set Sol + medium, run one real task you already know the answer to. Bring Astra in on the same prompt only if Sol misses. That single A/B beats another hour of benchmark Twitter.