AI Model Comparison — GPT, Claude, Gemini, Llama side by side

Picking an AI model is a multi-axis decision: input price, output price, context window, output cap, modalities, training cutoff, and whether the weights are downloadable. Most comparison pages on the web pick two of those axes and call it done. This page lays out all of them at once for the 18 models that almost every Sri Lankan developer, student, or startup ends up considering — drawn from 7 vendors.

1. The pricing formula

Every commercial LLM API charges separately for input and output tokens. The cost of one call is:

usd_per_call = (input_tokens ÷ 1,000,000) × input_$/M + (output_tokens ÷ 1,000,000) × output_$/M

Monthly cost is then per-call multiplied by the number of calls you make per month. The cost projection in the tool above uses this formula directly — no caching credits, no batch discounts, no enterprise rates. That is the published list price, which is what you actually pay on most plans.

2. Reasoning models bill hidden tokens

Models with extended thinking (OpenAI o1, o3-mini, Claude with extended thinking, DeepSeek R1) emit a chain-of-thought before the visible answer. The vendor bills all of those tokens as output. A 200-word visible reply can consume 2,000–5,000 output tokens on a hard problem. When you compare a reasoning model to a chat model on this page, multiply the reasoning model's output number by 3–10× before drawing conclusions.

3. Context window vs output cap

Context window is the size of the prompt the model can read. Output max is the size of the reply the model can write. These are different limits. Gemini 2.5 Pro reads 2M tokens but only emits ~65K. Llama 4 Maverick advertises 1M context with an 8K output cap. For long-form generation (article drafts, code refactors), the output cap is the one that bites first.

4. Capability flags

Vision, audio, function calling, reasoning, and open weights are independent dimensions. A model can support any combination. The comparison table marks each capability with a chip; struck-through chips mean the model lacks that capability. The vendor docs URL on each row is the authoritative source — capabilities sometimes ship in API tiers behind allowlists or paid plans.

5. Cross-check

The data module exports a deterministic verifyWorkedExamples() function that recomputes seven hand-derived test cases — including zero-input edges, the 1M-token boundary, and a 10⁹-token large input — to assert the cost math matches the file-header arithmetic to within a millionth of a cent. A second integrity check asserts unique ids, non-negative prices, positive context windows, and non-empty positioning notes. Both run at typecheck time. If a row drifts during a quarterly update, the build fails.

• OpenAI — API pricing
• OpenAI — Models documentation (context windows, capabilities)
• Anthropic — Claude API pricing
• Anthropic — Claude model cards
• Google — Gemini API pricing
• Google — Gemini model documentation
• Together AI — Llama 4 hosted pricing
• Meta — Llama model cards
• DeepSeek — API pricing
• xAI — Grok models and pricing
• Mistral AI — La Plateforme pricing

The dataset and its cited sources were last cross-checked on 2026-05-12. Pricing and capability flags are reviewed quarterly and whenever a vendor publishes a substantive pricing or model update.