Independent · Source-cited · No sponsored rankingsSources linked in every article
AIToolRanked
AI News · 5 min read

GPT-6.2 Astra and Sol: Price, Benchmarks and Who Gets Them

OpenAI launched GPT-6.2 Astra and Sol on November 17, 2026. Astra lists at $8 and $40 per million tokens. Sol stays at $2 and $10, in Work and Codex.

RA
· Founder, AIToolRanked
A large faceted crystal and a smaller warm sun disc, each above a dashboard panel, with descending bars and a shrinking stack of coins on a deep navy field
On this page

What OpenAI shipped on November 17

On November 17, 2026, OpenAI published Introducing GPT-6.2 Astra and Sol ↗. The Learn page ↗ has the same date, and the changelog ↗ has two November 17 entries. OpenAI ↗ and OpenAIDevs ↗ posted it. The Verge ↗ (Emma Roth, 5:45 PM UTC) reported it. The picture above is an original graphic created by aitoolranked. We have not used either model.

OpenAI calls Astra the frontier model for the hardest projects and Sol the efficient everyday model, both returning better results with fewer output tokens. Roth writes that OpenAI says Astra is better at slides, spreadsheets, and documents that "read less like AI," follows templates more closely, and codes better for websites, gaming, and 3D. Roth reports OpenAI calling Sol "smarter and more efficient" than GPT-6.1 Sol.

Where they run, and the token price

From November 17, Plus, Pro, Enterprise, and Edu get both in Work and Codex. The Learn page also lists Business. Enterprise and Edu admins turn them on. Its Chat line is: "To use GPT-6.2 Astra or Sol in ChatGPT, open Work." Consumer Work chats default to Sol at Light. Enterprise workspaces with access default to Sol at Medium. Codex and Work share usage. Choose the model in the Codex picker or with /model. Pro 500 gets Ultrafast on either model. Pro 200 gets it for the first time, on Sol. Enterprise and Edu get it after an owner enables it.

gpt-6.2-astra and gpt-6.2-sol are on Responses and Chat Completions. Tool calling requires Responses. Ultrafast is on for both. Editing a prefix, or the top-level instructions field, changes how Responses keeps reasoning. OpenAI says to leave prefixes unchanged and append. Data residency is US and EU for Standard, Fast, and Batch, and US only for Ultrafast.

ModelInputCached inputOutput
GPT-6.2 Astra$8$0.40$40
GPT-6.2 Sol$2$0.10$10

Per million tokens, Astra's input and output are 20% below GPT-6 Astra, and cached input is 60% below. Sol keeps the GPT-6.1 Sol rates. A second November 17 changelog item, in public beta, prices a two-hour background window at 30% below Standard and an eight-hour window at 40% below, for gpt-6.2-astra, gpt-6.2-sol, gpt-6.1-terra, and gpt-6-luna.

The scores, with the lab named

OpenAI's post: on Agents' Last Exam, Astra is 5.2 points above GPT-6 Astra at less than a third of the cost, and 5.5 ahead of Claude Fable 5.2 at 5% of the cost. Fable 5.2's leaderboard results exclude nine tasks that used a fallback model. Terminal-Bench-Science 0.1 is 9 percentage points above GPT-6 Astra, at less than one-third the cost. On Cognition's FrontierCode 2.0, Sol is within 0.5 points of Astra at about one-fifth the cost. Internal GDPval v2: Astra up 6.2 percentage points versus GPT-6 Astra, with 22% fewer output tokens, and Sol up 5.3 points versus GPT-6.1 Sol. On LifeVal, Sol outscores GPT-6 Astra at about one-seventh the cost.

Artificial Analysis ↗ wrote that the release "delivers the #2 model in the Artificial Analysis Intelligence Index and most of the Cost per Task Pareto frontier." Next in that thread: "GPT-6.2 Astra is flat with GPT-6 Astra, and GPT-6.2 Sol loses 3 points relative to GPT-6.1 Sol in the Artificial Analysis Coding Agent Index (Codex harness at max effort)." Then: "GPT-6.2 Astra and GPT-6.2 Sol hallucinate at a rate of 59% and 55% at max effort respectively, as measured by AA-Omniscience Hallucination Rate." And: "GPT-6.2 Astra leads AA-3D at 1318, with Claude Fable 5.2 at 1271 and GPT-6.2 Sol at 1259."

Perplexity ↗ says both models are in Search and Computer. On Perplexity's own WANDR benchmark: "On our WANDR benchmark, GPT-6.2 Astra scored 0.783 at about $12.54 per task, up 0.095 from GPT-6 Astra."

Under the Preparedness Framework, OpenAI says both meet Critical for cybersecurity and High for biological and chemical work. The tighter cyber safeguards also restrict some legitimate vulnerability research. Astra is the first external model at High for AI self-improvement. In two alignment tests both were more likely to omit an obstacle, such as a missing file. Codex instructions substantially reduce that, OpenAI says, and the lab is still investigating. Daybreak in the post is GPT-6 Sol on Blue, a GPT-6 Astra alpha, and GPT-6 Cyber on Red. OpenAI says it will work to bring 6.2 there.

What people are saying

  • "Over 21x cheaper and almost 4x faster than Fable 5.2." — Tony ↗ (@EnvolDev), his own single-prompt comparison.

  • "GPT-6.2-Astra: 77/100 (tied for best)" — Nick ↗ (@nickbaumann_), on his own GeoGuessr test of 100 locations.

  • "Coding remains flat for GPT-6.2 Astra vs. GPT-6 Astra, while GPT-6.2 Sol drops 3 points vs. GPT-6.1 Sol." — Vishal Singh ↗ (@vishalsingh2972)

  • "GPT-6.2 Astra: $11.49 · 38m 23s" and "GPT-6.2 Sol: $2.65 · 17m 12s" — soloDavid ↗ (@itsolodavid), one task.

  • "GPT-6.2 Astra costs like Claude Fable 5.2. GPT-6.2 Sol does not." — Parik ↗ (@pparik__).

Reddit was not reachable for this story.

Our take

Open Sol. Consumer Work defaults to it, Plus can use it, and $2 / $10 is unchanged. Astra is now cheaper per token than GPT-6 Astra (table above). The November 16 note recorded the old $10, $1, and $50 rates.

Leave the coding harness you already run. The Coding Agent Index is flat for Astra and lower for Sol. Fable 5.2 at $8 and $40 in our November 10 note matches Astra's input and output list rates, which is separate from Parik's one-task bill. See Codex vs Claude Code and Pro Codex limits. Work and Codex share a pool.

Use Astra on documents, decks, 3D, and science. Deepnote tested the previous pair. Gemini 4 Hydrogen is the other everyday model this month. Plans: ChatGPT vs Claude vs Gemini. Security work meets the tighter refusals, and AA's hallucination rates are a reason to check every figure. Watch Daybreak, regular Chat, and a later coding eval.

Verdict: Sol for daily Work and Codex. Astra for a document, a deck, or 3D or science work when the cache will hit.

More from the news desk · About Rai Ansar