Skip to content
, ,

GPT-6.2 Astra and Sol: Price, Benchmarks, ChatGPT Access and How to Use Them (2026)

11–17 minutes
Woman at an office desk with city views holding a phone showing the OpenAI logo, hook GPT-6.2 20% cheaper

OpenAI has updated both of its main GPT-6 models. On Tuesday 17 November 2026 it released GPT-6.2 Astra, its top model for hard professional work, and GPT-6.2 Sol, the faster and cheaper model most people will use every day.

The headline for developers is price: Astra is 20% cheaper than GPT-6 Astra, and Sol keeps GPT-6.1 Sol’s price. The headline for everyone else is mixed: the benchmark numbers went up, but several well-known power users called the launch a step back.

This guide puts everything in one place: what changed, full specs and prices, every benchmark we could verify (with who ran it), how to use both models in ChatGPT, Codex and the API, how they compare with Claude Fable 5.2 and Gemini 4 Carbon, and which one you should pick.

TL;DR

  • What: GPT-6.2 Astra (frontier model for hard projects) and GPT-6.2 Sol (efficient everyday model), released by OpenAI on 17 November 2026.
  • Who gets it: ChatGPT Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, plus the API. Not on the Free tier.
  • API price per 1M tokens: Astra $8 input / $40 output (was $10 / $50). Sol $2 / $10 (unchanged).
  • Specs: 1,050,000-token context, 128K max output, text and image input, text output.
  • Benchmarks: Astra is #2 on the Artificial Analysis Intelligence Index (62) behind Claude Fable 5.2 (65). Big gains in science and documents, flat on coding.
  • Catch: higher hallucination rates on hard questions and harsh early reviews (“essentially unusable”, said Matt Shumer).

What OpenAI Launched on 17 November

OpenAI describes the pair as models “built for professional work”. Astra is the frontier model “for your hardest projects”. Sol is the efficient model for everyday tasks, coding and agent workflows, and OpenAI says many of its own staff now use Sol as their default.

Some background helps with the confusing names. GPT-6 Astra launched in September. OpenAI planned a GPT-6.1 Astra, but OfficeChai reports it was cancelled in October after internal tests found an alignment regression. So GPT-6.2 Astra is the first real upgrade to Astra, while GPT-6.2 Sol replaces GPT-6.1 Sol.

What changed, in plain words

  • Better office work: Astra follows templates for slides, spreadsheets and documents more closely, including layout, colours and exact font sizes taken from a template or screenshot.
  • Professional conventions: OpenAI says Astra handles the habits of law, finance and accounting more reliably.
  • Fewer tokens: both models give answers with fewer output tokens. OpenAI says this lets you do more inside the same ChatGPT usage limits, and it lowers API bills.
  • Computer use: more precise control of CAD and simulation tools. OpenAI demos included a 3D car model, a Python physics solver animated in Blender and an animated RC car assembly guide.
  • Lower Astra price: 20% off GPT-6 Astra, with cached input down 60%.

Specs and API Prices

These figures come from OpenAI’s developer model pages for gpt-6.2-astra and gpt-6.2-sol, checked on 19 November 2026.

Spec GPT-6.2 Astra GPT-6.2 Sol
API model ID gpt-6.2-astra gpt-6.2-sol
Context window 1,050,000 tokens 1,050,000 tokens
Max input / max output 922,000 / 128,000 922,000 / 128,000
Knowledge cutoff 18 May 2026 1 June 2026
Input / output types Text + image in, text out Text + image in, text out
Reasoning Always on; none is rejected Always on; none is rejected
Fine-tuning No No
Tools (Responses API) Web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, tool search

Price per 1 million tokens

Item GPT-6.2 Astra GPT-6 Astra (old) GPT-6.2 Sol
Input $8.00 $10.00 $2.00
Cached input $0.40 $1.00 $0.10
Cache write (30 min) $10.00 – $2.50
Cache write (12 h, short context) $20.00 – $5.00
Output $40.00 $50.00 $10.00

Modifiers apply to both models:

  • Long context: above 272K input tokens, input and cache rates double and output is 1.5x, for the whole request. A long Astra call is $16 in / $60 out.
  • Batch: 50% of standard.
  • Fast mode: 2x standard.
  • Ultrafast: 3x standard until 28 February 2027, then 6x from 1 March 2027 (Pacific time).

Tip: keep prompts under 272K tokens where you can. Crossing that line doubles your input cost on every token of the request, not just the extra part. Our OpenAI API pricing guide has worked cost examples for the other GPT-6 models.

Rate limits

Tier Requests per minute Tokens per minute
Build 5,000 1,000,000
Launch 10,000 4,000,000
Grow 15,000 40,000,000

The limits are the same for Astra and Sol, according to OpenAI’s model pages.

Benchmarks: Vendor vs Independent

There are two kinds of numbers here. OpenAI’s own charts (vendor-reported, OpenAI picked the tests and settings) and scores from independent groups such as Artificial Analysis, Vals AI and Epoch AI. We list them separately.

OpenAI-reported scores

OpenAI published these as charts. The figures below are approximate readings by OfficeChai, with cost per task in brackets.

Benchmark GPT-6.2 Astra GPT-6.2 Sol GPT-6 Astra Claude Fable 5.2 Gemini 4 Carbon
Agents’ Last Exam v1.1 ~78% (~$3) ~73% (~$1.3) ~73% (~$9) ~73% (~$60) ~77.5% (~$4.5)
Terminal-Bench-Science 0.1 ~77% (~$7) ~68% (~$2) ~68% (~$24) ~73%* not shown
Harvey LAB-AA v1.1 (legal) ~15% ~11% ~9% ~7% not shown
BenchCAD 2.1 Core ~83% ~73% ~63% ~84% not shown
AA-3D (blind human Elo) 1,318 1,259 – 1,271 –

*OpenAI ran Fable 5.2 on the mini-swe-agent harness and notes that Anthropic’s own Terminal-Bench-Science figure is higher. OpenAI also reports Astra gains 6.2 points on GDPval-v2 with 22% fewer output tokens, and that Sol is within 0.5 points of Astra on Cognition’s FrontierCode 2.0 at about one fifth of the cost.

Independent scores

Who tested Result
Artificial Analysis Intelligence Index Fable 5.2 65, Astra 62 (+5 vs GPT-6 Astra), Gemini 4 Carbon 60, Opus 5.5 59, Sol 57 (+2)
Artificial Analysis Coding Agent Index Astra flat vs GPT-6 Astra; Sol down 3 points vs GPT-6.1 Sol
AA-Omniscience hallucination rate (max effort) Astra 59% (GPT-6 Astra 50%), Sol 55%
Vals Index Astra 7th, Sol 9th (1.6 points apart); Fable 5.2 first
Vals cost per task Sol $3.18, Astra $10.68, GPT-6 Astra $21.28
Vals Science Index Astra first (also first on Terminal-Bench Science)
Vals Legal Research Astra 30th, Sol 52nd
Epoch Capabilities Index Astra 174, tied with Fable 5.2
FrontierMath Erdős (Epoch) Astra solved 7 of 68 open problems; previous record 3 of 68
ARC-AGI (ARC Prize, via TestingCatalog) ARC-AGI-2 95.8% at $0.50/task; ARC-AGI-1 98.5% at $0.17/task

Note: the AA-Omniscience questions are deliberately very hard, so Artificial Analysis says the 55-59% hallucination figures do not reflect everyday chat. They do show the models guess more often than GPT-6 Astra when they don’t know.

The pattern is clear. Astra is a big step up for science, documents and cost per task, roughly level for coding, and weaker than the leaders on legal research despite OpenAI’s legal chart.

How to Use GPT-6.2 in ChatGPT and Codex

Both models are rolling out to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. Free users don’t get them. On Enterprise and Edu, a workspace admin has to enable the models first.

  1. Update the app. Rollouts arrive in waves, so update the ChatGPT app or refresh chatgpt.com if you don’t see the new names yet.
  2. Open the model picker. Choose GPT-6.2 Sol for everyday chats and quick work, or GPT-6.2 Astra for long documents, slide decks, spreadsheets and research.
  3. Give Astra a template. For slides or documents, upload your template or a screenshot of it and ask it to match fonts, colours and layout. This is the area OpenAI improved most.
  4. In Codex, pick the model in the model menu, or in the Codex CLI use the /model command (or start with codex -m gpt-6.2-sol). Sol is the sensible default for routine edits.
  5. Pro subscribers: Pro 500 gets Ultrafast mode on both models. Pro 200 gets Ultrafast for the first time, starting with Sol.

Tip: because both models write shorter answers, the same usage cap now stretches further. Our ChatGPT and Codex usage limits guide explains how the caps work per plan, and ChatGPT Free vs Go vs Plus vs Pro helps if you’re deciding whether to upgrade.

How to Call GPT-6.2 from the API (Python)

Both models work with the Responses API and Chat Completions. If you already use GPT-6, switching is mostly a model-name change.

  1. Install or update the SDK: pip install --upgrade openai
  2. Set your key: export OPENAI_API_KEY="sk-..." (on Windows PowerShell: $env:OPENAI_API_KEY="sk-...")
  3. Send a request with the code below.
from openai import OpenAI

client = OpenAI()

resp = client.responses.create(
    model="gpt-6.2-sol",          # or "gpt-6.2-astra" for hard work
    reasoning={"effort": "medium"},  # "none" is rejected on GPT-6.2
    input="Summarise the attached quarterly report in 5 bullet points."
)
print(resp.output_text)

Migration checklist if you are moving from GPT-6 Astra or GPT-6.1 Sol:

  • Reasoning effort: if your code sends none or minimal, change it to low. eesel’s testing notes OpenAI suggests medium as a starting point for new work.
  • Caching: keep the stable part of your prompt (system message, tool list, documents) at the start so cached input at $0.40 (Astra) or $0.10 (Sol) kicks in.
  • Re-run your evals: answers are shorter, so tests that check length or exact wording may need updating.
  • Budget for long context: anything above 272K input tokens is billed at the higher rate.

New to the API? Our OpenAI API Python tutorial covers streaming, structured outputs and tools step by step, and our reasoning effort guide shows how effort levels change your bill.

GPT-6.2 Astra vs Claude Fable 5.2 vs Gemini 4 Carbon

GPT-6.2 Astra GPT-6.2 Sol Claude Fable 5.2 Gemini 4 Carbon
API price (in / out per 1M) $8 / $40 $2 / $10 $8 / $40 $2 / $10 until 30 Nov, then $4 / $20
AA Intelligence Index 62 57 65 (first) 60
Vals Index rank 7th 9th 1st –
Epoch Capabilities Index 174 – 174 –
Strongest at Science, documents, slides, cost per task Cheap everyday and coding work Coding, writing, real-world projects Agents at low cost

Sources: OpenAI model pages, Artificial Analysis and Vals AI via The Decoder, Epoch AI, VentureBeat; Fable and Carbon prices from our Fable 5.2 vs GPT-6 Astra vs Gemini 4 Carbon comparison, Claude Fable 5.2 guide and Gemini 4 Carbon guide.

Which should you pick?

  • Best overall quality, especially coding and writing: Claude Fable 5.2 still leads most independent indexes.
  • Science, spreadsheets, slide decks and documents on a budget: GPT-6.2 Astra. Same list price as Fable, but Vals and OpenAI both show much lower cost per task because it uses fewer tokens.
  • High-volume everyday work, support bots and routine code: GPT-6.2 Sol. Vals calls it the cheapest model in its top 10.
  • Cheap agents right now: Gemini 4 Carbon, but its price doubles on 1 December.

Why Power Users Are Unhappy

The launch got an unusually rough reception from well-known testers, as reported by VentureBeat and Gizmodo:

  • Matt Shumer called the models “essentially unusable” at first and said Astra kept asking permission for trivial steps such as downloading a file.
  • Theo Browne called Astra “the worst model OpenAI has ever released” after a long Rust rewrite project produced a build that ran 56% slower.
  • Dan Shipper (Every) said Astra often stopped early on end-to-end builds and has a “noticeable house style”, and he preferred Fable 5.2.
  • Ethan Mollick praised its output on an ambitious 3D project but found it “very uneven” and harder to steer than rivals.

OpenAI’s Thibault Sottiaux said early reviewers had tested an older version and that many changes landed in the days before launch. Shipper disputed that he was told about any changes.

Known Problems and Fixes

Problem Why it happens What to do
API error when reasoning.effort is none GPT-6.2 models reject none Use low or higher
Model not in ChatGPT picker Staged rollout, Free plan, or admin hasn’t enabled it (Enterprise/Edu) Update the app, check your plan, ask your workspace admin
Astra keeps asking for permission Reported by Shumer and others; more cautious agent behaviour State up front which actions it may take without asking (downloads, file edits)
Agent stops early or doesn’t flag blockers OpenAI’s system card says both models more often failed to report obstacles like a missing file Ask it to list any missing files or tools before it finishes, and check logs
Confident wrong answers on niche facts Higher AA-Omniscience hallucination rate Turn on web search or file search and ask for sources
Bill higher than expected Requests above 272K input tokens are charged at long-context rates Trim context, use caching, use Batch for non-urgent jobs
Security research requests refused Stricter cyber safeguards Apply to OpenAI’s Daybreak programme for vetted security access

Safety Notes and What Is Still Unknown

Both models reach the “Critical” cybersecurity level and the “High” biological and chemical level in OpenAI’s Preparedness Framework. Astra is also the first model OpenAI has released publicly that reaches its “High” threshold for AI self-improvement.

  • OpenAI added a monitoring system called Control Room that can halt research tasks, plus wider monitoring of reasoning and actions.
  • Cyber scores such as ExploitBench (Astra about 54%) were measured with safeguards switched off, so they are not what normal users see.
  • Epoch AI found Astra does more reasoning steps inside a single token than earlier models (12.3 vs 10.5 for GPT-6 Astra), which makes its thinking harder to monitor.

Warning: treat Astra and Sol like any assistant. Check facts in legal, financial and medical work, because independent tests show they guess more often than GPT-6 Astra on hard questions.

Still unknown

  • Exact ChatGPT message limits per plan for each model.
  • When, or if, GPT-6 Astra and GPT-6.1 Sol leave ChatGPT and the API.
  • Whether OpenAI will ship a fix for the early-stopping and permission complaints.
  • Whether Free and Go users will get Sol later.

My Take: Who Should Switch

On paper this is a strong release. Astra costs 20% less, scores 5 points higher on the main independent index, and does each task for roughly half the money of GPT-6 Astra on Vals. For API users already on GPT-6 Astra, switching is close to a free upgrade.

The reviews are the warning sign. When testers who use these models all day say an update feels worse, it usually points to behaviour that benchmarks miss: stopping early, asking too often, a fixed writing style. Those things matter more in long agent jobs than in one-shot questions.

My suggestion: move routine API traffic to Sol now, since the price is the same and tokens are fewer. Test Astra on your own documents and science-style tasks before moving anything important. For coding and writing, Fable 5.2 still has the edge by most independent measures.

FAQ

When was GPT-6.2 released?

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Tuesday 17 November 2026, in ChatGPT, Codex and the API.

Is GPT-6.2 free in ChatGPT?

No. It is available on Plus, Pro, Enterprise and Edu plans. Free users don’t get it at launch.

How much does GPT-6.2 Astra cost in the API?

$8 per million input tokens, $0.40 cached and $40 per million output tokens. Above 272K input tokens the rates rise to $16 input and $60 output.

What is the difference between Astra and Sol?

Astra is the most capable model for hard professional and scientific work. Sol is faster and four times cheaper per token, and OpenAI recommends it for everyday tasks, coding and agents.

Is GPT-6.2 Astra better than Claude Fable 5.2?

Not overall. Fable 5.2 leads the Artificial Analysis and Vals indexes and is preferred by most reviewers for coding and writing. Astra wins on several science benchmarks and on cost per task.

What is the GPT-6.2 context window?

Both models have a 1,050,000-token context window, with up to 922,000 input tokens and 128,000 output tokens.

Why does my API call fail with reasoning effort “none”?

GPT-6.2 models always reason, so OpenAI rejects none. Set effort to low, medium or higher.

What happened to GPT-6.1 Astra?

According to OfficeChai, OpenAI cancelled GPT-6.1 Astra in October after internal tests showed an alignment regression, so Astra went straight from 6 to 6.2.

Sources

  • OpenAI developer docs, model pages for gpt-6.2-astra and gpt-6.2-sol (specs, prices, rate limits), checked 19 Nov 2026
  • OpenAI, “Introducing GPT-6.2 Astra and Sol” (17 Nov 2026)
  • The Decoder, “OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains” (17 Nov 2026), with Artificial Analysis, Vals AI and Epoch AI results
  • OfficeChai, “OpenAI Releases GPT 6.2 Astra & GPT 6.2 Sol” (benchmark chart readings)
  • TestingCatalog, “OpenAI launches GPT-6.2 Astra and Sol on ChatGPT and APIs”
  • VentureBeat, “‘Essentially unusable’: OpenAI launches GPT-6.2 Astra and Sol to negative reception from AI power users”
  • Gizmodo, “Anthropic Could Be the Real Winner in OpenAI’s Latest Model Release” (18 Nov 2026)
  • eesel AI, “GPT-6.2 Astra: what’s new, pricing, benchmarks, and who should switch”
  • No Code MBA AI release tracker, November 2026
Free weekly email

Get new AI guides every week

One short weekly email with our newest local AI, ComfyUI and AI tool guides. No spam, no pop-ups, just the good stuff.

Free. Unsubscribe anytime in one click. We only use your email to send this newsletter. See our privacy policy.

Woman at an office desk with city views holding a phone showing the OpenAI logo, hook GPT-6.2 20% cheaper