I have been watching the frontier model race all year, and yesterday OpenAI made its loudest move yet. On 17 November 2026, the company launched GPT-6.2 Astra and GPT-6.2 Sol, a new flagship and a cheaper everyday model, and the pricing tells you exactly who OpenAI is aiming at: Anthropic, one day before its IPO, and every developer who has been quietly moving their agent workloads to the cheapest capable model they can find.
Both models are available now in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and through the API as gpt-6.2-astra and gpt-6.2-sol. Two limits worth knowing straight away: they are not in regular Chat, and free accounts do not get them.
What changed, in plain numbers
- GPT-6.2 Astra: $8 per million input tokens, $0.40 for cached input, $40 per million output tokens. That is 20% less than GPT-6 Astra on input and output, and 60% less on cached input.
- GPT-6.2 Sol: $2 input, $0.10 cached input, $10 output per million tokens, unchanged from GPT-6.1 Sol.
- Context: both models support around 1.05 million tokens of context, with up to 128,000 output tokens.
- Benchmarks: on Agents Last Exam, Astra scores 5.2 points higher than GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2 at just 5% of the cost. Sol comes within 0.5 points of Astra on Cognition FrontierCode 2.0 at about one-fifth the cost.
Read that last line again, because it is the story for most of us. Sol, the $2 model, is nearly matching the $8 flagship on real coding work. OpenAI says its own staff already use Sol as their default for technical and non-technical tasks, and after seeing these numbers I understand why.
The angle nobody is spelling out: Sol is the freelancer model
Here is how I would choose. If you are a solo developer, a small agency, or a creator running coding agents on client work, Sol is now the sensible default. You get frontier-adjacent coding at $2 per million input tokens, and the cached-input rate of $0.10 makes long agent sessions with repeated context genuinely cheap. Astra earns its price only when the work is genuinely hard: multi-file refactors, scientific analysis, legal or financial documents, or slides and spreadsheets that must match a template exactly. OpenAI says Astra is its best model yet at following slide templates down to the exact font size, and it set a new high on Terminal-Bench-Science for research work in a terminal.
There is also a pricing detail that changes the maths on long documents. The headline $8 and $40 rates apply to requests up to 272,000 input tokens. Go past that and Astra input and caching rates double, with output up 50%. If your workflow stuffs whole codebases or long reports into the context window, budget for that, or split the work so Sol handles the bulk and Astra only reviews the hard parts.
The safety footnote matters too
Astra is OpenAI first externally released model to reach the High threshold for AI self-improvement, which measures a model's ability to accelerate AI research itself. Both models also meet the Critical threshold for cybersecurity capability. OpenAI says it is shipping them with its strongest safeguards to date, including tighter restrictions on higher-risk cybersecurity assistance and a new Control Room system that can halt research workloads. That is honest framing, and it is also a reminder that these models are powerful enough that OpenAI felt it had to build new brakes before releasing them.
How this fits the price war
Put this launch next to what I covered yesterday and the pattern is hard to miss. Alibaba Qwen4-Max arrived with 2.4 trillion parameters and a coding score above Claude, and Tencent Hy4 Flash is selling tokens at $0.07 per million. OpenAI answer is not to be the cheapest, it is to make its flagship 20% cheaper and its everyday model nearly as good as the flagship. Astra now lists at exactly the same $8 and $40 as Anthropic Claude Fable 5.2, which turns the choice into a straight quality comparison instead of a price one. For small teams, that is good news: the frontier keeps getting cheaper, and the budget tier keeps getting smarter.
My take: if you pay for ChatGPT, open Work today and try Sol on your next real task before reaching for Astra. Most days, you will not miss the flagship, and your usage limits will stretch much further.
Sources: OpenAI, Introducing GPT-6.2 Astra and Sol, 17 November 2026, openai.com. Related reading on this blog: Alibaba Qwen4-Max Is Here: 2.4 Trillion Parameters, a 1 Million Token Context, and a Coding Score Above Claude and Tencent Hy4 Flash Costs $0.07 per Million Tokens: The AI Price War Just Hit Rock Bottom.
