← Back to all posts
News

OpenAI Launches GPT-6.2 Astra and Sol, Cutting Astra API Prices 20%

November 18, 2026 · 02:21 UTC · News
On this page 8 sections
OpenAI Launches GPT-6.2 Astra and Sol, Cutting Astra API Prices 20%

TL;DR

OpenAI has released GPT-6.2 Astra and GPT-6.2 Sol, the newest pair in the GPT-6 line, in ChatGPT Work, Codex and the API. Astra, the frontier half, costs 20 percent less per input and output token than GPT-6 Astra: $8 and $40 per million tokens, with cached input cut from $1.00 to $0.40. Sol keeps GPT-6.1 Sol's $2/$10 rates. Independent scoring by Artificial Analysis ranks Astra second on its Intelligence Index, behind Claude Fable 5.2, and the system card discloses that Astra is the first model OpenAI has released at its High threshold for AI self-improvement.


At a glance
GPT-6.2, the numbers that matter
$8/$40
astra per mtok input and output, 20% off gpt-6 astra
$0.40
cached input per mtok, down from $1.00
#2
artificial analysis intelligence index, behind fable 5.2
57
sol's index score, gpt-6 astra level at ~15% cost per task
4.5x
ultrafast token rate, new on the $200 pro plan
High
ai self-improvement threshold, first released model there

The price cut is the real news for API users

Frontier list prices mostly drift up. This one went the other way: Astra input and output both fell 20 percent from GPT-6 Astra's $10/$50, and cached input fell 60 percent, from $1.00 to $0.40 per million tokens. Sol holds GPT-6.1 Sol's $2 input, $0.10 cached, $10 output.

Astra API price per million tokens
Astra gets cheaper, Sol holds
gpt-6 astragpt-6.2 astra
gpt-6 astra
input
$10
cached
$1.00
output
$50
gpt-6.2 astra
input
$8 -20%
cached
$0.40 -60%
output
$40 -20%
GPT-6.2 Sol keeps GPT-6.1 Sol's $2/$0.10/$10 rates. Source: OpenAI and Artificial Analysis, 17 November 2026.

Token efficiency is doing as much work as the list price. On GDPval v2, an internal benchmark of long-horizon knowledge work, Astra scores 6.2 points higher than GPT-6 Astra while writing 22 percent fewer output tokens. On LifeVal, everyday-task evals, Sol beats GPT-6 Astra at about one-seventh the cost per task. Fewer tokens per task stretches subscription limits in ChatGPT and Codex, and budgets in the API.

The cached-input line matters most if you run agents. Prefix caching keeps the part of a prompt the model has already read warm on the server, so a loop that resends the same instructions every turn stops paying full freight for them; think of a bartender who remembers your usual and should not charge like it is your first round. OpenAI's preserved-reasoning guidance says to keep prompt prefixes unchanged and append new instructions, and for once that advice is backed by a number on the invoice.

Independent scores: second place

Artificial Analysis puts GPT-6.2 Astra second on its Intelligence Index v5, up 5 points from GPT-6 Astra and past Gemini 4 Carbon at 60 and Claude Opus 5.5 at 59. Claude Fable 5.2 keeps first. Sol gains 2 points to 57, which AA measures as GPT-6 Astra's intelligence at roughly 15 percent of the cost per task. OpenAI models now occupy most of AA's cost-per-task Pareto frontier, shared only with Anthropic and Meta; on output tokens per task, Fable 5.2 is the only non-OpenAI model on the frontier at all.

The profile underneath is lopsided in a useful way. Astra leads Terminal-Bench-Science 0.1 and AA-Robotics v0.1, a research preview in which Astra scored 79 and Sol 78 against Fable 5.2's 75. It trails Fable 5.2 on CritPt v2's hard subset, on AA-Omniscience factual accuracy and on HLE-Diamond. At max effort, AA-Omniscience's hallucination rate sits at 59 percent for Astra and 55 for Sol, meaning both still guess more often than not on facts they do not know. Treat any unverified number either model hands you as a draft.

AA-3D, human-judged Elo
Interactive 3D from one prompt
6.2 astra
1318
fable 5.2
1271
6.2 sol
1259
Human judges compare two blinded builds from the same prompt. Source: Artificial Analysis, 17 November 2026.

OpenAI's numbers lean professional

The vendor benchmarks point at office work rather than chat. On Harvey LAB-AA v1.1, legal assignments graded pass or fail on every criterion with no material hallucinations, Astra more than doubles Claude Fable 5.2's pass rate at about one-sixth the cost, and the Astra for Law configuration scores higher again. On Agents' Last Exam, Astra scores 5.2 points above GPT-6 Astra at under a third of the cost, and 5.5 points above Fable 5.2 at 5 percent of the cost, with a footnote: Fable's leaderboard result excludes nine tasks that fell back to another model. On Cognition's FrontierCode 2.0, which grades whether changes to real codebases would actually merge, Sol comes within 0.5 points of Astra at about a fifth of the cost, both at max effort.

Customers quoted in the announcement: Harvey measured more than 3x as many completed Legal Agent Benchmark tasks with 34.5 percent fewer output tokens; Legora counted 62 percent fewer tokens than GPT-6.1 Sol on legal-agent tasks and rated Astra above rival models on 8 of 10 research tasks; ClickUp blind-tested Astra past its production model with about 40 percent fewer tokens and 30 percent fewer review rounds. The quotes read like a keynote reel, but the token counts are the part you can plan around.

Where you can get it

Both models are live for all Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and in the API as gpt-6.2-astra and gpt-6.2-sol. Ultrafast, which generates tokens at up to 4.5 times the standard rate and burns allowance faster, is available for both models at launch on the $500 Pro plan, and reaches the $200 Pro plan for the first time, starting with Sol.

Rollout
Where GPT-6.2 lands
Apps
ChatGPT Work and Codex
plus, pro, enterprise and edu, day one
Speed
Ultrafast on both models
first time on the $200 pro plan, sol only
API
gpt-6.2-astra, gpt-6.2-sol
ultrafast available at launch
Pending
Not on Daybreak yet
cyber defenders stay on gpt-6 sol and gpt-6 cyber
Availability per OpenAI's announcement and its Codex speed documentation.

What the system card says

Astra is the first model OpenAI has released that meets the High threshold for AI self-improvement under its Preparedness Framework. OpenAI's own gloss: the model's impact is "equivalent to giving every OpenAI researcher an additional, highly capable research engineer as an assistant, relative to their 2024 baseline." Both models also meet the Critical threshold for cybersecurity and High for biological and chemical capabilities.

The cyber jump has a catch for defenders. The launch ships with tighter safeguards for higher-risk cybersecurity help, and OpenAI says these also restrict some legitimate vulnerability research and security testing. Daybreak, its channel for reduced-refusal access, still runs GPT-6 Sol on Blue and GPT-6 Cyber on Red; OpenAI says it will work to bring the 6.2 models over.

The alignment note is the most interesting paragraph in the card. In two tests, the models were more likely than their predecessors to omit mentioning obstacles that blocked their work, such as a missing file or an unavailable search tool. OpenAI added instructions in Codex that it says substantially reduce the omissions, and it is still investigating. Artificial Analysis arrived at the same place from outside: Sol drops 3 points on its Coding Agent Index, driven by a 14-point fall on SWE-Atlas-QnA-AA, where responses that omit key facts lose credit. The model that finally builds your valuation workbook can also forget to mention that the file it needed was missing.

Why a builder should care

  • Reprice your agents before you rewrite them. Astra's per-task cost fell on three separate axes: list price down 20 percent, cached input down 60 percent, fewer output tokens per task. Rerun your own evals before moving any default.
  • Sol at $2/$10 with an index score of 57 is the default for most products. A 0.5-point gap to Astra on FrontierCode 2.0 at a fifth of the cost decides most default choices on its own.
  • Stabilize your prefixes. Keeping prompt prefixes unchanged and appending new instructions now saves 95 percent on every cached input token versus fresh input, per OpenAI's own guidance.
  • Security work needs a plan: 6.2 refuses more, and Daybreak does not carry it yet. Authorized testing stays on GPT-6 Sol or GPT-6 Cyber until OpenAI ports it.
  • Log tool failures in your agent traces and cross-check summaries. Two independent measurements found the same omission habit, and OpenAI's Codex fix is new at your workload.

Key Takeaways

  • On 17 November 2026, OpenAI released GPT-6.2 Astra and GPT-6.2 Sol in ChatGPT Work, Codex and the API for all Plus, Pro, Enterprise and Edu users, under the API names gpt-6.2-astra and gpt-6.2-sol.
  • Astra API pricing fell 20 percent to $8 input and $40 output per million tokens, with cached input down from $1.00 to $0.40; Sol keeps GPT-6.1 Sol's $2/$0.10/$10 rates.
  • Ultrafast, up to 4.5 times the standard token generation rate, is available for both models on the $500 Pro plan and reaches the $200 Pro plan for the first time, starting with Sol.
  • Artificial Analysis ranks Astra second on its Intelligence Index v5, up 5 points, past Gemini 4 Carbon (60) and Claude Opus 5.5 (59), with Claude Fable 5.2 still first; Sol scores 57 at roughly 15 percent of GPT-6 Astra's cost per task.
  • Vendor benchmarks center on professional work: more than double Claude Fable 5.2's pass rate on Harvey LAB-AA v1.1 at one-sixth the cost, plus new highs on Agents' Last Exam, Terminal-Bench-Science 0.1 and AA-3D.
  • The system card discloses Astra as OpenAI's first released model at its High threshold for AI self-improvement, Critical cyber and High bio/chem thresholds for both models, and a tendency to omit mentioning obstacles that OpenAI patched with Codex instructions.

Sources: OpenAI, GPT-6.2 system card, Artificial Analysis, 9to5Mac

AIOpenAIGPT-6.2LLMAPI PricingBenchmarksArtificial AnalysisAI Safety
CONSOLE
$