Skip to content
← All posts
Release report·4 min read

GPT-6.2 Astra and Sol: what OpenAI’s new tiers mean for testing

OpenAI shipped GPT-6.2 in two tiers — Astra for hard work, Sol for everyday. The reported numbers, the pricing, and where they fit once we test them.


OpenAI shipped GPT-6.2 on November 17, 2026 in two tiers: Astra, aimed at complex professional and research work, and Sol, the everyday model. It lands about a week after Anthropic’s Claude 5.2 and Fable 5.2 — close enough that the two launches will be read against each other, which is exactly the comparison this site exists to settle on real tasks rather than press slides.

Both tiers are available now 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. One caveat up front: every figure below is OpenAI’s own or drawn from launch-day coverage. The first outputs are already running in our arena, but we have not scored GPT-6.2 against the field yet — so treat the numbers here as reported, not measured.

Two tiers, two jobs

The split is the structural story. Instead of one model with a size dial, OpenAI is shipping Astra for the hard end — multi-step professional and research work — and Sol for the everyday end, where volume and latency matter more than peak depth. If that framing sounds familiar, it is the tiering logic the whole field has converged on; what differs is the price attached to each rung.

TierReported price per 1M tokens (in / cached / out)Positioned forvs previous
Astra$8 / $0.40 / $40Complex professional and research work~20% below GPT-6 Astra
Sol$2 / $0.10 / $10Everyday, high-volume tasksUnchanged from GPT-6.1 Sol

OpenAI puts Astra at $8 per million input tokens, $0.40 cached and $40 output — roughly 20% under GPT-6 Astra — while Sol holds exactly where GPT-6.1 Sol was, at $2 / $0.10 / $10. The number that matters is not the sticker but the ratio OpenAI claims alongside it: that Astra does the same class of professional and research work at under a third of GPT-6 Astra’s cost per task. Price per token and cost per task are different things — a cheaper token that burns more of them on a job is not cheaper — which is precisely the gap a same-task test closes.

What OpenAI reports

On benchmarks, OpenAI reports Astra scoring 5.2 points higher than GPT-6 Astra on Agents’ Last Exam and 22 points higher on ExploitBench, its security-exploitation evaluation. Those are the two figures OpenAI chose to lead with; we are quoting them, not confirming them, and we are setting aside the other numbers circulating in launch coverage until there is something to check them against.

“Regaining the lead” is positioning, not a result

OpenAI frames GPT-6.2 as retaking the frontier. Independent reporting is more measured: 9to5Google describes the results as broadly comparable to Anthropic’s latest at a much lower cost — comparable, not clearly ahead. A lower price at similar quality is a real competitive claim on its own; the lead claim is OpenAI’s, and that only settles on identical tasks.

Safety

OpenAI describes Astra as its most capable model yet for cyber tasks and says it ships with tighter safeguards for higher-risk use. That is consistent with the reported ExploitBench jump: a model that is better at finding and writing exploits is also one that needs more guardrails around who can point it at what. How those safeguards behave in practice is its own test, separate from the capability score.

What’s live, and what isn’t yet

The first published batch contains ten first-attempt artifacts: both tiers ran the five most-voted coding briefs (landing page, dashboard, palette tool, arcade game and news front page) at requested ultra effort with client delegation disabled. Current CLI source/catalog imply xhigh inference effort for that selection; the per-call wire setting was not captured. The results update tracks current coverage across the effort ladder. We recorded CLI wall time and token usage, with $0 in incremental subscription charges. Timing affected by controller sleep is withheld from public generation-time values; the original observations remain in the run evidence. These observations do not establish a general speed ranking, a scored verdict or a measured API cost-per-task comparison against other models.

GPT-6.2 Astra · requested ultra, single turn
GPT-6.2 Sol · requested ultra, single turn
Requested-ultra outputs with delegation disabled. Current source/catalog imply xhigh inference effort; per-call wire resolution was not captured. Open “Landing” in the arena to run both full-screen, go blind, and vote.

The coding arena runs every model on the same briefs, live and blind-judged, and the leaderboard has the standings. The GPT-6.2 ultra pilots are in now; the scored verdict follows as the votes land.

Sources

Frequently asked questions

What is the difference between GPT-6.2 Astra and Sol?

They are the two tiers of OpenAI’s GPT-6.2 generation, launched together on November 17, 2026. Astra is aimed at complex professional and research work; Sol is the everyday, higher-volume tier. Both are available 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.

How much do GPT-6.2 Astra and Sol cost?

Per OpenAI’s reported pricing, Astra is $8 per million input tokens, $0.40 cached and $40 output — about 20% below GPT-6 Astra. Sol is $2 input, $0.10 cached and $10 output, unchanged from GPT-6.1 Sol. OpenAI also claims Astra does comparable professional and research work at under a third of GPT-6 Astra’s cost per task; we have not verified that.

Is GPT-6.2 better than Claude 5.2?

OpenAI positions GPT-6.2 as retaking the frontier, but independent coverage frames the results as broadly comparable to Anthropic’s latest at a lower cost, without confirming a clear lead. We have first-attempt outputs live but no scored comparison yet, so we cannot say. The comparison only settles on identical tasks, which is what our arena is for.

Has testingmodels.com tested GPT-6.2 yet?

Yes. The first published batch has ten first-attempt artifacts: Astra and Sol at requested ultra effort with delegation disabled on five coding tasks. Current CLI source/catalog imply xhigh inference effort; per-call wire resolution was not captured. Current effort coverage is listed in our coding-results update. CLI wall time, token usage and $0 incremental subscription charges are recorded. Public timing is withheld for runs affected by controller sleep. We have not established a scored comparison or measured API cost ranking.

Don’t take the post’s word for it

The arena runs every model’s real output live. Pick a challenge, go blind, and cast a vote that counts in the public tally.

Open the arena