GPT-6
text model · Announced
OpenAIFrontier reasoning with an effort dial, from instant to deliberate.
Announced model. All pricing, benchmarks and evals are estimates, replace before this page is indexed.
Fit and price
Recruitment fit
93
Per hour
$0.698
Per minute
$0.012
30-min interview
$0.349
About
GPT-6 is the current frontier of the GPT line. The headline change over GPT-5 is not raw benchmark score but controllability: a single effort parameter moves the same model from sub-second replies to multi-minute deliberate reasoning, so one integration covers both a live chat widget and an overnight batch job. For hiring teams that matters because screening and structured interview scoring have opposite latency budgets, and you no longer need two models to serve them.
- Reasoning effort control
- Tool use
- Structured outputs
- Vision
- Streaming
- Prompt caching
- 90+ languages
Strengths and watch-outs
Strengths
- Full transcript plus job spec in context
- One integration, effort dial
- ATS-ready structured outputs
Watch out for
- 24× the cost of Mini at max
- Still needs audited rubrics
- Unpublished pricing
Best for Rubric-based interview scoring · Shortlist vs job spec · Draft candidate feedback
5 variants
Max effortMax
Longest deliberation. Reserve it for work where a wrong answer is expensive, final scoring, offer-risk analysis, contract review.
42.0s
latency
24×
rel. cost
98
quality
HighHigh
The default for anything a human will act on. Strong multi-step reasoning without the max-effort latency tax.
9.0s
latency
8×
rel. cost
94
quality
MediumBalanced
The workhorse. Good judgement at a price that survives volume, this is the rung most production traffic should sit on.
3.2s
latency
3×
rel. cost
88
quality
LowFast
Near-instant. Skips deliberation entirely, right for classification, routing and anything a user is waiting on.
900ms
latency
1.4×
rel. cost
79
quality
MiniMini
Smaller sibling. The cheapest way to run GPT-6-shaped prompts when the task is mechanical rather than judgemental.
600ms
latency
1×
rel. cost
71
quality
The numbers
HumanLike evals
- CV field extraction (F1)
- 0.94Estimate
- Time to first token (medium effort)
- 640 msEstimate
Projection. Not yet run against GPT-6.
