Decision guide
Should you move from GPT-5.6 Sol to GPT-6 Sol?
GPT-6 Sol replaces GPT-5.6 Sol at half its $4/$20 promo price with the same 1.05M context — the question is whether your prompts behave the same, not whether it is cheaper.
Switch if
- You pay GPT-5.6 Sol rates today: GPT-6 Sol lists $2/$10 (cached $0.20) against $4/$20.
- OpenAI's claimed lower factual-error rate matters for your knowledge or research workload.
- You also route volume work to GPT-5.6 Luna — GPT-6 Luna ($0.10/$0.50) can move in the same change.
Stay if
- You have a validated GPT-5.6 Sol baseline and cannot re-run evaluations yet — GPT-5.6 Sol stays served, and its $4/$20 promo is guaranteed at least through 2026-11-21.
- Your hardest tasks need GPT-6 Astra-class quality anyway — test Astra instead of Sol.
- Structured-output or tool-call behaviour regresses on GPT-6 Sol in your pilot.
Check before production traffic moves
- 01Pin `gpt-6-sol` explicitly; the GPT-5.6 ids stay live and do not auto-upgrade.
- 02Check requests over 272K input tokens: they bill the whole request at $4/$15.
- 03Re-run refusal and structured-output tests at the reasoning effort you use in production.
This is a same-provider generation swap where the price change alone justifies a pilot. OpenAI reports GPT-6 Sol at 68.8% on DeepSWE 1.1 (max, vendor-run), and the independent AA Intelligence Index v4.3.2 scores it 47.5 against 47.0 for GPT-5.6 Sol at max effort — similar capability at half the cost.
Run a controlled trial first
A second option in a labelled pilot is cheaper than a production cutover. Score the trial, then migrate or roll back on evidence.
Pilot plan
- Replay a frozen prompt set on both ids with identical reasoning effort and validators.
- Record accepted-task rate, retries, latency, and output tokens per task.
- Blind-score a sample of long-form answers for factual errors, the change OpenAI highlights.
Score the trial
- Accepted-task rate
- Pass rate on the same validators and human sample.
- Cost per success
- Tokens and retries per accepted task at your effort level.
- Behaviour parity
- Structured-output validity and refusal rate versus the GPT-5.6 baseline.
Migration sequence
- Add `gpt-6-sol` behind the existing OpenAI adapter with GPT-5.6 Sol as the fallback.
- Shift traffic in stages and promote after two consecutive clean evaluation windows.
Rollback plan
- Keep the `gpt-5.6-sol` id configured for instant switch-back while it remains served.
- Replay any failed batch on the fallback id rather than mixing model outputs.
Hidden costs to price
- Long-context requests above 272K input double the input rate for the whole request.
- Prompt retuning and evaluation time are real even for a same-provider upgrade.
Evidence to check