AI Models Compared

An AI model is a software system trained on large amounts of data so it can generate predictions, text, images, audio, or video. In product language, “AI model” usually means a foundation model — a large neural network trained once at enormous cost, then adapted to many tasks through prompting, fine-tuning, or tool use. Large language models (LLMs) such as GPT-5.5, Claude 4, and Gemini 2.5 take text in and return text out. Multimodal models also accept or produce images, audio, or video. Open-weights models such as Llama and DeepSeek can be downloaded and run on your own hardware; proprietary models are typically reached through a vendor API.

GPT-5.5 currently leads our intelligence rankings, while DeepSeek V3 offers one of the lowest API costs among strong open-weights options. Claude 4 Opus remains a top choice for writing and long-context code, and Gemini 2.5 Pro excels with long-context workflows thanks to its 2M-token window.

When people ask which model is “best,” they are usually asking several questions at once: which reasons most accurately, which answers fastest, which costs the least at production volume, which can read the longest documents, and which handles code, writing, or vision most reliably. No single score answers all of those. That is why this page compares models across a small set of transparent axes instead of crowning one universal winner — or open a GPT-5.5 vs Claude 4 Opus matchup for a head-to-head view.

We compare models on public, reproducible metrics whenever possible. For language models the core columns are intelligence (a composite of academic benchmarks), generation speed in tokens per second, time-to-first-token latency, context window size, input and output API prices, and a blended price that reflects a typical traffic mix. Image, video, and speech models use modality-appropriate pricing — per image, per second of video, or per minute of audio — and appear on their image, video, and speech category pages with columns that match how those products are sold.

Side-by-side matchups let you pit any two models against each other on the same axes, with notes on strengths, weaknesses, and who each option is best for. Use the rankings below to scan the field, then open individual model pages or the comparator when you are ready to decide.

The Intelligence Index is our 0–100 summary of how strong a language model is on widely cited public benchmarks, including MMLU, MMLU-Pro, GPQA, MATH, and HumanEval. Higher is better — today GPT-5.5 leads, with Claude 4 Opus and Gemini 2.5 Pro close behind. The index is a rough ordering for scanning the catalog — not a substitute for reading the underlying scores. A model can lead on coding while trailing on knowledge-heavy exams and still land in the middle of the pack.

Reasoning models that spend more compute at inference time — such as OpenAI o1 and DeepSeek R1 — often score higher on hard math and planning tasks; chat-oriented models may look closer on everyday writing. Always treat the index as a starting point, then check task-specific benchmarks and, when stakes are high, run your own evaluation set.

Speed on this site means median output tokens per second during steady-state generation across major API providers — not the time until the first token appears. Latency (time to first token, or TTFT) is listed separately because interactive apps feel sluggish when the first token is slow even if later tokens stream quickly.

Reasoning models that pause to “think” before answering will look slower on tokens-per-second; that is expected and does not mean the API is broken. Fast tiers like GPT-4o mini, Claude 3.5 Haiku, and Gemini 2.0 Flash usually top the speed board. Hosted open-source endpoints can also vary by provider, so treat speed as a directional signal rather than a guaranteed SLA.

Pricing for token-based models is shown as dollars per million input tokens and dollars per million output tokens, using list API rates from each vendor. Blended price mixes those rates at a 3:1 input-to-output ratio, which approximates many production workloads where prompts and retrieved context outweigh generated replies. That single number makes it easier to compare models whose input and output prices diverge sharply — for example Llama 3.1 8B at the budget end versus frontier rates on GPT-5.5.

Image, video, and audio models are priced in their native units. Promotional credits, enterprise discounts, and batch or cached-input rates can change effective cost — our tables reflect public list pricing so comparisons stay apples-to-apples.

Benchmark scores differ across leaderboards for several good reasons. Contaminated or overlapping evaluation sets reward models that memorized similar questions. Prompt formatting, temperature, tool use, and whether chain-of-thought is allowed can swing results by large margins. Some vendors report self-reported scores; others are third-party replications. Multimodal and agentic tasks are still inconsistently measured.

When two sources disagree, prefer transparent methodology, look at the full score suite rather than a single headline number, and validate on your own prompts before locking in a production choice. See how we rank for the full methodology behind this catalog.

Rankings and prices on What’s The Big Data are refreshed monthly. Each model entry carries a last-updated stamp; we re-pull API prices and public benchmark figures in the first week of each month. New model releases are added as soon as enough reliable data exists to place them fairly. If you spot a stale price or missing release, tell us — keeping the catalog current is the point of a living comparison page.