How We Rank AI Models
Published 23 September 2026 by Jake Hissitt, Founder of Stob.AI.
Reviewed 6 November 2026 to clarify weighted factors, retired entries and the historical September example.
Our AI model comparison tool ranks more than 230 models. People ask how the order is decided, so here it is in full.
The six scores
Every model carries six scores from 1 to 10:
- Reasoning — multi-step problem solving, analysis and judgement under ambiguity.
- Coding — code generation, debugging and agentic software work.
- Speed — practical latency and throughput for interactive use.
- Cost efficiency — capability delivered per dollar at published rates.
- Context length — usable input window and maximum output.
- Multimodal — image, audio and video handling.
How the rank is calculated
Rank is capability-first. Reasoning and coding carry triple weight, context length and multimodal single weight, and cost efficiency and speed contribute smaller weighted amounts to the total:
Score = (reasoning x 3) + (coding x 3) + context + multimodal + (cost efficiency x 0.4) + (speed x 0.3) − availability penalty
The availability penalty is what stops the list filling up with models you cannot responsibly build on:
| Status | Penalty |
|---|---|
| Superseded by a newer model | −6 |
| Active previous generation | −3 |
| Limited, invitation-only or trusted access | −2 |
| Preview | −1.5 |
Retired models are kept at the bottom of the comparison as historical references.
A superseded model with perfect capability scores therefore sits well below a current model with slightly lower ones. That is deliberate. A ranking that recommends a model its own provider has replaced is not useful.
Why cost has a smaller weight
Weight cost too heavily and the cheapest model in the world tops a list titled "best". Weight it at zero and the list becomes a procurement hazard. Its smaller weight can change the order of models with similar capability scores, alongside speed and availability.
In our September 2026 recalibration, GPT-6 Sol at $2 / $10 sat above Astra at $10 / $50. That is a historical example: GPT-6 Sol has since been superseded by GPT-6.1 Sol, and the current capability scores and order differ.
Recalibration
Scores are relative to the current frontier, not fixed for life. When a new generation lands, the previous one moves down even though nothing about it got worse. In the September 2026 pass we reduced reasoning and coding scores for GPT-5.6 Sol, GPT-5.5, GPT-5.4, GPT-5 Pro, o3-pro, Claude Opus 5, Fable 5, Mythos 5, Gemini 2.5 Pro, Qwen3.8 Max and DeepSeek V4 Pro, because a perfect 10 has to mean "at the current frontier".
Pricing, context windows and availability are re-verified alongside every launch, and each model carries its own verification date.
What the ranking is not
- Not a benchmark leaderboard. Published benchmarks inform the scores; they are not copied into them.
- Not a recommendation for your workload. Rank 1 is a reasonable default, not an answer to a specific problem. Filter by the use case that matters to you.
- Not a measure of total system cost. The model is one line in the bill. Architecture usually costs more than model choice.
The measure that decides real deployments is cost per accepted outcome: total workflow cost divided by outcomes that meet your standard.
Frequently asked questions
How often is the comparison tool updated?
Within days of a major launch, and pricing is re-verified at the same time. Each model shows its own verification date.
Why is a cheaper model ranked above an expensive one?
A cheaper model can rank higher when its capability, cost efficiency, speed and availability produce a higher weighted total.
Why did a model drop without changing?
Scores are relative to the current frontier. A new generation pushes the previous one down, and a superseded status carries a ranking penalty.
Do you take money from model providers?
No. The rankings are not sponsored and no provider pays for placement.
What to do next
Use the rank as a starting shortlist, then filter by the capability your workload depends on and test the top two against your own tasks. Open the AI model comparison tool, or start with the pillar guide, Best AI Model 2026.