AI FAQ
40 questions on reading the rankings, model choice, tools, cost and term misconceptions. Answers only cite data recorded on this site.
Reading the rankings
15Why can v5 and v4.3.2 scores not be compared?
The Artificial Analysis Intelligence Index moved to v5 on 2026-10-27 with harder tasks, so scores shifted down across the board. Column headers carry the version; cross-version comparisons are meaningless. Methodology & sources →
Does a dash mean zero?
No. A dash means the benchmark has not recorded that model, not that its capability is zero. Newer, open and minority-language models are covered less often, so sorting on the composite alone systematically underrates them. Methodology & sources →
How is the 3:1 blended price calculated?
It weights input 3 parts to output 1: (3 × input price + output price) / 4. It is closer to a real bill than input price alone, but output length varies by model, so budget on cost per task. Methodology & sources →
What is the Intelligence Index?
The main leaderboard uses the Artificial Analysis Intelligence Index, which combines several public benchmarks. Individual scores (GPQA, Terminal-Bench 4.0, LiveBench Design and others) are listed on each model page and should be read together. Methodology & sources →
What does GPQA Diamond measure?
Graduate-level science reasoning with relatively strong resistance to contamination, scored out of 100. It measures hard reasoning, not everyday writing or coding. Methodology & sources →
Why does Terminal-Bench 4.0 matter?
It runs software-engineering and terminal tasks in real containers and measures task completion, which is closer to how agents actually behave than chat-style Q&A. Below 20%, try it on a small scope first. Methodology & sources →
Does higher tok/s always mean faster?
tok/s is sustained output speed; time to first token (TTFT) decides how soon text appears. This site records both: TTFT matters more for live chat, tok/s for long generation. Methodology & sources →
Why does this site run no benchmarks itself?
In-house results cannot be independently checked. Only public, clearly defined, reproducible evaluations and official vendor pricing pages are used, with sources and version labels on every page. Methodology & sources →
What are the price units?
USD per 1M tokens, split into input and output. Cache-hit pricing is listed separately; subscription-only models without per-token pricing are excluded from price and value rankings. Methodology & sources →
How often is the leaderboard updated?
Roughly every two weeks; model and pricing data follow this site's data files, with the latest date in the footer. Corrections can be reported on the methodology page and are logged once verified. Methodology & sources →
Are open weights the same as open source?
No. Open weights mean you can obtain parameters and self-host, while the licence may restrict commercial use or redistribution. Open source also implies training code, data and full freedom. Model pages only record the licence name; check the official model card. Open-weight ranking →
Why does the same model behave differently across products?
System prompts, retrieval, tool chains and context management at the product layer all change results. This site records the model's public benchmark scores, not the experience of any particular chat product. Model library →
How do I read the capability radar on a model page?
Six dimensions are each normalised against the best value on this site. The chart shows relative position, not absolute score, and the area across dimensions is not comparable. Use the score table below for decisions. See a model page →
Why do some models have no composite index?
A composite index requires the model to be covered by that leaderboard. New, open or niche models often only have individual benchmark scores; missing values are shown as a dash and never estimated. Leaderboard →
What goes into the changelog?
Definition changes, model additions and removals, score and price corrections, and site features. Every change is recorded rather than made silently; the "recent updates" block on the home page is generated from this same log. Changelog →
Choosing a model
6Which models lead on composite intelligence?
The top three recorded here by points are: Claude Fable 5.2 (65), GPT-6.2 Astra (62), Gemini 4 Neon (61). These are single-dimension rankings; also weigh price, speed and context. Leaderboard →
Which model should I pick for research reasoning?
The top three recorded here by points are: Gemini 4 Argon (96.4), Claude Sonnet 5.5 (95.6), Gemini 3.8 Flash (95.4). These are single-dimension rankings; also weigh price, speed and context. Reasoning ranking →
Which model should I pick for terminal agents or coding?
The top three recorded here by points are: Kimi K3.1 Pro (67.2), Gemini 4 Carbon (65.2), GPT-6.2 Astra (64.1). These are single-dimension rankings; also weigh price, speed and context. Coding ranking →
Which model has the fastest output?
The top three recorded here by tok/s are: Claude Haiku 5.5 (255), DeepSeek V4.1 Flash (203), Gemini 4 Hydrogen (188). These are single-dimension rankings; also weigh price, speed and context. Speed ranking →
Which model gives the best value?
The top three recorded here by points are: Claude Haiku 5.5 (180), GPT-6 Luna (140), GLM-5.3 Flash (118.1). These are single-dimension rankings; also weigh price, speed and context. Value ranking →
Which model has the largest context window?
The top three recorded here by tokens are: GPT-6.1 Terra (1,050,000), GPT-6 Luna (1,050,000), Gemini 3.8 Flash (1,048,576). These are single-dimension rankings; also weigh price, speed and context. Model library →
Tools and usage
8What tools are available for everyday Q&A, writing and information triage?
This site lists 9 tools under chat, including Claude, ChatGPT, Gemini. They fit everyday Q&A, writing and information triage, with pricing and caveats on each card. AI tools library →
What tools are available for coding, refactoring and repo-scale tasks?
This site lists 9 tools under code, including Claude Code, Cursor, Codex. They fit coding, refactoring and repo-scale tasks, with pricing and caveats on each card. AI tools library →
What tools are available for image generation, editing and asset batches?
This site lists 5 tools under image, including ChatGPT Images, Nano Banana, Midjourney. They fit image generation, editing and asset batches, with pricing and caveats on each card. AI tools library →
What tools are available for video generation, editing and post-production?
This site lists 5 tools under video, including Veo, Sora, Runway. They fit video generation, editing and post-production, with pricing and caveats on each card. AI tools library →
What tools are available for speech synthesis, transcription and music?
This site lists 3 tools under audio, including ElevenLabs, Suno, Whisper. They fit speech synthesis, transcription and music, with pricing and caveats on each card. AI tools library →
What tools are available for workflow orchestration and autonomous agents?
This site lists 4 tools under agent, including n8n, Dify, Manus. They fit workflow orchestration and autonomous agents, with pricing and caveats on each card. AI tools library →
What tools are available for literature search, reviews and source organisation?
This site lists 3 tools under research, including Perplexity, NotebookLM, Elicit. They fit literature search, reviews and source organisation, with pricing and caveats on each card. AI tools library →
What tools are available for model deployment, inference and API access?
This site lists 5 tools under infra, including Ollama, LM Studio, vLLM. They fit model deployment, inference and API access, with pricing and caveats on each card. AI tools library →
Cost and deployment
11Are open-weight models always cheaper?
Not necessarily. Open weights remove the API margin but you pay for GPUs, memory, inference serving and operations. The value ranking here uses API blended price and excludes self-hosting costs. AI glossary →
Is a bigger context window always better?
A long context holds more material, but "lost in the middle" means mid-document facts get overlooked. Put key conclusions near the start or end, or retrieve and compress first. AI glossary →
Why does cache-hit pricing matter?
Repeated prefixes such as system prompts and long document headers are billed at the cache rate, usually far below input price. Keeping stable prefixes first is the simplest cost lever. AI glossary →
Why is cost per task more important than unit price?
Two models at $2/$10 can differ 3× in output tokens for the same task. A stronger model often writes less and needs fewer retries, so a higher unit price can mean only slightly higher cost per task. AI glossary →
How do I tell if a model can be self-hosted?
Check the "weights" row on the model page: only open-weight entries with a licence name qualify. Closed API or subscription models cannot be self-hosted, and no weight URLs are recorded here. AI glossary →
What if a model is deprecated?
Vendors retire older models. Prefer options with same-price substitutes and decouple routing from prompts; the "same-price rivals" block on each model page is built for this. AI glossary →
What is the difference between the tools library and the AI directory?
The tools library is a curated set of 43 tools with commentary and pricing; the AI directory (nav.ppv9.com) is a full 320-site classified catalogue for exhaustive browsing. They share one source list. AI glossary →
Where can I look up unfamiliar terms?
The AI glossary on our blog covers 138 terms in 8 categories, each with a one-line definition, common misconceptions and further reading. "Related terms" on model and tool pages link straight there. AI glossary →
Can I use this data directly for procurement?
Use it for shortlisting, not as the sole basis. Scores come from public evaluations and prices from official pricing pages with sync lag; verify on the vendor site and run your own task set before buying. Methodology & sources →
How does the context window relate to tokens?
Context windows are measured in tokens and include both input and generated output. A Chinese character usually takes 1–2 tokens, so a 1M context does not hold one million Chinese characters, and room must be left for the answer. Context window term →
Is the "free" tag in the AI directory reliable?
It reflects public information at the time of collection and only means a free tier exists. Quotas, concurrency and commercial limits are not tracked; always confirm on the vendor site. AI tools library →
Data sources & methodology
Full methodology & sources →Every number on this page comes from published evaluations. We run no tests of our own and accept no vendor-submitted scores. Main sources:
Evaluation data in the wider industry also comes from public datasets, leaderboard platforms and their evaluation harnesses, private in-house sets, production logs and user feedback, and synthetic data. This site's current rankings use only the parts that are public, independently checkable and clearly defined.