OpenAI released GPT-6.2 on November 17, and the interesting decision was not the model. It was shipping two of them, at prices five times apart, and putting both into the developer surfaces first.

Two products, one release

GPT-6.2 Astra is positioned for hard reasoning and high-difficulty professional work: debugging across services, comparing conclusions drawn from multiple sources, tasks where a wrong answer is expensive. GPT-6.2 Sol is positioned for everyday coding, writing and multi-step workflows.

Both launched into ChatGPT Work, Codex and the Responses API rather than the standard chat interface. That ordering is the tell. OpenAI is selling this to people whose usage is metered, which means the pitch has to survive a cost comparison.

The price sheet

Astra: $8 per million input tokens, $0.4 on a cache hit, $40 per million output. Against GPT-6 Astra that is roughly 20% lower on input and output, and 60% lower on cached input.

Sol: $2 input, $10 output, about one fifth of Astra, unchanged from GPT-6.1 Sol. Cache pricing for Sol was not broken out in the same detail.

The gap between the two is the whole strategy. At five times the price, Astra only makes sense if it saves more than four times its cost in retries, review time or wrong decisions.

What Astra is actually better at

OpenAI's own number: on Terminal-Bench-Science, Astra beats GPT-6 Astra by nine percentage points at less than a third of the cost. That combination — better and cheaper than the previous flagship — is what usually triggers migration, and it is also the kind of claim that needs a third party.

Third-party reports suggest Sol comes close to Astra on some coding tests. If that holds broadly, the honest reading is that a lot of "hard" work is not actually hard, and was only being sent to the expensive model because there was no cheaper option with the same interface.

The Ultrafast mode

A separate lever: an Ultrafast mode that generates tokens up to 4.5 times faster than standard, at the cost of burning quota faster. It is tied to subscription tier — the $200 Pro tier gets it on Sol first, the $500 tier on both.

Faster generation is not free speed. It usually means smaller batches and more expensive serving, which is why it is gated by subscription rather than priced per token. For interactive coding, latency is worth paying for; for batch jobs it is not.

The real change is routing

The release makes the routing decision explicit rather than hiding it behind one model. Enterprises will end up with a policy: classify the task by difficulty, risk, latency budget and cost ceiling, then send it to Astra or Sol. High-value work escalates; the bulk runs cheap.

That is a better architecture than a single model, and a harder one to operate. A router that misclassifies sends a hard task to Sol, gets a plausible wrong answer, and nobody notices until it is in production. The fix is not a better classifier but a verification step: cheap model first, escalate on a failed check.

What is missing

No published context window comparison, no detail on tool-calling differences between the two, and no independent cost-per-completed-task study — which is the only number that matters for the routing decision.

How to decide

Take one workflow you currently run on a frontier model. Run it on Sol, and only escalate to Astra when a deterministic check fails. Log both the escalation rate and the cost per finished task. If the escalation rate is under 20%, you have just cut your bill by more than half. If it is over 50%, the task was never a Sol task and the cheaper tier was a distraction.