In 2022, there were three AI language models worth knowing: GPT-3, a research model from Google called PaLM, and early LaMDA. In 2025, there are more than 50 significant models in active production use, spanning six major organisations, three continents, and every price point from free open weights to ₹7 per thousand words.
This is the reference you need: every model that matters, what it was built for, and how to choose.
How to Read This Article
Closed models are run by the company that built them — you access them through an API or product interface, and the company controls the weights, training data, and pricing. OpenAI, Anthropic, and Google run closed models.
Open-weight models are released with their weights publicly available — you can download, run, and modify them. Meta, Mistral, DeepSeek, and Alibaba have released open-weight models. This matters enormously for enterprises: you can run an open-weight model inside your own infrastructure, with no data leaving your servers.
Context window is how much text the model can "see" at once — its working memory. Larger is better for complex, multi-document tasks. The evolution from 4,000 tokens (GPT-3) to 10,000,000 tokens (Llama 4 Scout) happened in four years.
Benchmark caveat: Benchmarks measure specific capabilities under controlled conditions. A model that tops MATH-500 may still struggle with your specific task. Use benchmarks for rough ranking; test on your actual use case before committing.
OpenAI — The Company That Changed Everything
OpenAI launched ChatGPT in November 2022 and triggered the current AI race. Their model lineup spans general language, reasoning, image, video, and speech.
GPT-4o — The Everyday Workhorse
Released May 2024. The "o" stands for omni: text, image, and audio in a single model. GPT-4o can see an image you send, hear your voice in real time, and respond in kind — all within one unified model rather than stitching separate systems together. It is faster and significantly cheaper than GPT-4 Turbo.
Best for: Conversational AI products, content generation, coding assistance, multimodal tasks (describe this image, explain this chart). For most everyday tasks, GPT-4o is OpenAI's flagship recommendation.
Context window: 128,000 tokens (~300 pages of text)
o1 / o3 / o4 — The Reasoning Series
Released September 2024 (o1), December 2024 (o3). These models spend compute thinking before answering — they run an internal chain-of-thought that is hidden from the user, exploring multiple approaches before settling on a response.
The result is dramatically better performance on problems that require step-by-step reasoning: complex mathematics, competitive programming, graduate-level science. o3 scored 87.5% on the ARC-AGI benchmark (a test of general intelligence), compared to the human average of around 85%.
Best for: Hard math problems, competitive coding, scientific analysis, any task where a wrong answer is worse than a slow answer.
Important trade-off: o-series models are 5–10× slower and more expensive than GPT-4o. For a customer service chatbot or a content summary, they are overkill.
GPT-4.5 and GPT-5
GPT-4.5 (February 2025): More capable than GPT-4o on general knowledge, with improved emotional intelligence and conversational naturalness. Very expensive — intended for applications where quality justifies cost.
GPT-5 (2025): Unifies reasoning and general capabilities into a single model. Leads on creative writing benchmarks in early 2026 evaluations.
Specialised OpenAI Models
OpenAI builds more than language models. Three specialised models extend their platform into image, speech, and video — each now embedded in products used by hundreds of millions of people.
| Model | Type | What It Does |
|---|---|---|
| DALL-E 3 | Image generation | Text-to-image, integrated into ChatGPT and Copilot |
| Whisper | Speech recognition | Open-source, state-of-the-art transcription in 99 languages including Hindi, Tamil, Telugu |
| Sora | Video generation | Text-to-video; produces 60-second clips |
Anthropic Claude — Built With Safety at the Core
Anthropic was founded in 2021 by Dario Amodei, Daniela Amodei, and others who left OpenAI over disagreements about safety practices. Claude is trained using Constitutional AI — a set of principles the model uses to self-evaluate its own outputs, making it less likely to produce harmful content even under adversarial prompting.
The Claude Family — Four Generations
| Model | Released | Key Capability | Context |
|---|---|---|---|
| Claude 3 Haiku | Mar 2024 | Fast and cheap — high-volume tasks | 200,000 |
| Claude 3 Sonnet | Mar 2024 | Balanced speed and quality | 200,000 |
| Claude 3 Opus | Mar 2024 | Most capable at launch — complex reasoning | 200,000 |
| Claude 3.5 Sonnet | Jun 2024 | Beat all GPT-4 class models on coding at launch | 200,000 |
| Claude 4 Sonnet | May 2025 | 72.7% on SWE-bench coding; hybrid reasoning mode | 200,000 |
| Claude 4 Opus | May 2025 | 72.5% SWE-bench; best for autonomous multi-hour tasks | 200,000 |
| Claude Sonnet 4.6 | Jul 2025 | Current production model; 98% of Opus quality at fraction of cost | 200,000 |
Claude 4 Opus can work autonomously on complex tasks for hours — running code, browsing files, running tests, iterating on its output — making it the current leader for agentic workflows where the AI operates more like a junior engineer than a chat partner.
Unique Claude advantage: 200,000-token context window across all models means Claude can read an entire large codebase, a full financial report, or a book-length document in one pass. For SAP consultants or finance teams working with long documents, this matters more than benchmark rankings.
Google Gemini — Native Multimodality and the Context Window Race
Google DeepMind merged their two AI research groups (Brain and DeepMind) in 2023 and released the Gemini series. Google's unique advantage: native multimodal reasoning (text, image, video, code, audio processed together from day one), and the deepest integration with Google Search, YouTube, and Workspace.
Gemini Model Lineup
| Model | Context | Pricing (per 1M tokens) | Standout |
|---|---|---|---|
| Gemini 1.5 Pro | 1,000,000 | $1.25 / $5.00 | First 1M-token window — entire codebases in one pass |
| Gemini 2.0 Flash | 1,000,000 | $0.10 / $0.40 | Ultra-cheap, fast, agentic tasks |
| Gemini 2.5 Flash | 1,000,000 | $0.30 / $2.50 | 84.0% MMMU, 98.1% MATH-500; thinking budget control |
| Gemini 2.5 Pro | 2,000,000 | $1.25 / $10.00 | Leads LiveCodeBench; 2M tokens |
Gemini 2.5 Pro has a 2 million token context window — the largest of any frontier model. That is roughly the equivalent of reading 50 full-length novels and then answering detailed questions about all of them.
Gemini Nano: A tiny version running directly on Android devices. Powers real-time transcription, smart reply, and spell-check on Pixel phones without sending data to Google's servers.
Meta Llama — Open Source as a Strategic Weapon
Meta's decision to release Llama model weights publicly is one of the most consequential moves in AI history. The reasoning is partly altruistic and partly strategic: if the world builds on Llama, Meta shapes the default infrastructure of AI. It is the same strategy Red Hat used with Linux.
Llama Family
| Model | Released | Key Facts |
|---|---|---|
| Llama 2 | Jul 2023 | Benchmark-competitive with Gemini 1.0 Pro. 13B, 70B variants. Free for commercial use. |
| Llama 3 | Apr 2024 | 8B and 70B. Significantly better than Llama 2. 128K context. |
| Llama 3.1 405B | Jul 2024 | 405 billion parameters. Competitive with GPT-4o. Truly open-weight. |
| Llama 3.3 70B | Dec 2024 | 70B model matching 405B quality through improved training. |
| Llama 4 Scout | Apr 2025 | Natively multimodal (text + image). 10 million token context. MoE architecture. |
| Llama 4 Maverick | Apr 2025 | Larger MoE model — outperforms GPT-4.5 on STEM benchmarks. |
Llama 4 Scout's 10 million token context window is extraordinary. At that context size, the model can read the entire source code of a large enterprise application — including all comments, history, and documentation — in a single inference call.
Why open weights matter for India: Indian enterprises and government bodies that cannot send sensitive data to foreign servers can run Llama 4 on their own infrastructure. The RBI, UIDAI, and defence ministries all have data residency requirements that make open-weight models not optional but necessary.
DeepSeek — The Sputnik Moment
On 20 January 2025, a Chinese startup called DeepSeek released DeepSeek R1 under the MIT Licence — unrestricted commercial use, free to download, free to modify.
The shock: DeepSeek R1 matched OpenAI's o1 on mathematics and coding benchmarks. Specifically: 97.3% on MATH-500 (near perfect), comparable scores on AIME and Codeforces to o1. The cost to run it: approximately 95% less than equivalent OpenAI models.
NVIDIA's stock dropped 17% in a single day. Not because DeepSeek had done something illegal — but because it suggested frontier AI may not require the compute scale (and therefore the GPU hardware) that Wall Street had assumed.
How they did it: DeepSeek uses a Mixture of Experts (MoE) architecture: instead of activating all 671 billion parameters for every token, only 37 billion parameters are active at any given time. The model routes each token to the experts most relevant for that token. Efficient routing + careful training = frontier performance at a fraction of the compute cost.
DeepSeek Model Lineup
| Model | Speciality | Notable Fact |
|---|---|---|
| DeepSeek V2 | General | First model to shock western labs with cost-efficiency |
| DeepSeek V3 | General | 671B params, 37B active. Competitive with GPT-4o |
| DeepSeek R1 | Reasoning | MIT licence. Matches o1 at 5% cost. |
| DeepSeek V3-0324 | General | Updated V3 with improved instruction following |
Mistral — Europe's Answer
Paris-based Mistral AI, founded in 2023, has produced models that consistently outperform models twice their size — proving that architecture efficiency matters more than raw parameter count.
| Model | Released | Standout Feature |
|---|---|---|
| Mistral 7B | Sep 2023 | Outperformed Llama 2 13B at half the parameters using sliding window attention |
| Mixtral 8x7B | Dec 2023 | First major open-source MoE: 8 experts, 2 active = 13B active / 47B total. Near GPT-3.5 quality. |
| Mistral Large 2 | 2024 | Flagship. Competitive with GPT-4 class. |
| Codestral | 2024 | Fine-tuned for code generation and completion. Best in class for Python and JS. |
| Mistral Small 3.1 | 2025 | Compact but capable. Good for on-premise deployment. |
Why Mistral matters: European AI sovereignty. Models trained under GDPR constraints, hosted in EU data centres. For European and Indian enterprises with strict data residency rules, Mistral offers an alternative to US-controlled models.
xAI Grok — The Realtime Web Model
Elon Musk's xAI launched Grok in November 2023, with a key differentiator: realtime access to X (Twitter) data. Every other major LLM has a training cutoff; Grok has a live feed to current events.
| Model | Notes |
|---|---|
| Grok 1 | Open-weight, 314B MoE. Released under Apache 2.0. |
| Grok 2 | Frontier-class, closed API. Image understanding. |
| Grok 3 | 2025. "Thinks before answering." Competes with o1 on reasoning. |
Best for: Journalists, researchers, and social media analysts who need responses grounded in today's news rather than a training cutoff.
Alibaba Qwen — The Multilingual Champion
Alibaba's Qwen series is the strongest open-weight model for multilingual tasks — particularly important for Indian regional languages.
| Model | Context | Standout |
|---|---|---|
| Qwen 2.5 72B | 128K | Outperforms models of similar size on coding, math, multilingual |
| Qwen 3 235B-A22B | 128K | MoE. 88.4% on GPQA Diamond — exceeds many closed frontier models. Supports 100+ languages. |
India relevance: Qwen's multilingual training includes stronger coverage of Indic scripts than most western models. For Bengali, Gujarati, Tamil, Telugu processing tasks, Qwen 3 often outperforms Llama and Mistral.
Microsoft Phi — Small But Surprisingly Capable
Microsoft Research's Phi series challenges the assumption that bigger is always better.
| Model | Params | Standout |
|---|---|---|
| Phi-3 Mini | 3.8B | Fits on a smartphone. Matches GPT-3.5 on many benchmarks. |
| Phi-4 | 14B | Best-in-class small model. 84.8% on MATH-500. Runs locally. |
Use case: On-premise deployment in constrained environments — hospital workstations, factory floor terminals, government offices — where a full GPU cluster is not available and cloud API calls are not permitted.
Specialist Models — Beyond Language
Reasoning
- OpenAI o3: Best reasoning model as of H1 2025 for complex math, science, and logic
- DeepSeek R1: Open-weight reasoning at 5% of o3 cost
- Gemini 2.5 Pro: Thinking mode with budget control — you set how long it "thinks"
Image Generation
| Model | Provider | Notes |
|---|---|---|
| DALL-E 3 | OpenAI | Integrated into ChatGPT. Good prompt adherence. |
| Midjourney v6 | Midjourney | Best aesthetic quality. Subscription only, no API. |
| Stable Diffusion 3.5 | Stability AI | Open weights. Run locally. Best for custom fine-tuning. |
| Flux 1.1 Pro | Black Forest Labs | Excellent text rendering in images — a historic weakness of image AI. |
| Adobe Firefly | Adobe | Commercially safe (trained on licensed data). Built into Photoshop. |
| Imagen 3 | Google's flagship image model. Part of Gemini ecosystem. |
Video Generation
| Model | Provider | Notes |
|---|---|---|
| Sora | OpenAI | Up to 60-second HD clips. |
| Veo 2 | Google DeepMind | Cinematic quality. Part of Workspace integration. |
| Kling | Kuaishou | Chinese competitor. Strong at realistic motion physics. |
Coding
| Model | Notes |
|---|---|
| GitHub Copilot (powered by GPT-4o) | Inline code completion inside VS Code, IntelliJ. 50,000+ enterprise customers. |
| Cursor (powered by Claude 4 Sonnet) | IDE built around AI — entire codebase context, multi-file edits. |
| Codestral (Mistral) | Open-weight code specialist — best for European on-premise deployments. |
Science
| Model | Notes |
|---|---|
| AlphaFold 3 (Google DeepMind) | Predicts 3D structure of proteins, DNA, RNA. Has accelerated drug discovery by years. |
| Med-PaLM 2 (Google) | Medical question answering at expert-physician level on USMLE. |
Indian AI — Built at Home
| Company | Model | Notes |
|---|---|---|
| Krutrim (Ola) | Krutrim LLM | India's first AI unicorn. 2 million users in 8 weeks. Indian language focus. |
| Sarvam AI | Sarvam-1, Saarika, Bulbul | Purpose-built for 22 Indian languages. Open-weight. Speech + text. Saarika: voice AI. Bulbul: text-to-speech. |
| CDAC / AIIMS | iOncology.ai | AI for cancer detection on radiological images. Deployed in government hospitals. |
| Qure.ai | qXR, qCT | AI radiology. Deployed across India and 50+ countries. 98% precision on TB screening. |
Sarvam AI is particularly significant: a startup built specifically to serve India's linguistic diversity, with models trained on Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, and Telugu alongside English — and released as open weights for Indian enterprises.
Context Window Evolution — The Working Memory Race
| Year | Model | Context Window | What Fits Inside |
|---|---|---|---|
| 2020 | GPT-3 | 4,000 tokens | A short article |
| 2023 | GPT-4 | 128,000 tokens | 300 pages of text |
| 2024 | Claude 3 | 200,000 tokens | A full novel |
| 2024 | Gemini 1.5 Pro | 1,000,000 tokens | 2,500 pages — entire codebase |
| 2025 | Gemini 2.5 Pro | 2,000,000 tokens | 5,000 pages — multiple large codebases |
| 2025 | Llama 4 Scout | 10,000,000 tokens | The entire Linux kernel source code |
How to Choose: The Right Model for Your Task
| Task | Best Choice | Why |
|---|---|---|
| Everyday writing and summarisation | GPT-4o or Claude Sonnet 4.6 | Best cost-performance balance |
| Complex coding, agentic tasks | Claude 4 Opus | Leads SWE-bench; works autonomously |
| Hard math and science | o3 or Gemini 2.5 Pro | Reasoning mode purpose-built for this |
| Long document analysis | Claude 4 or Gemini 2.5 Pro | 200K–2M token context |
| Cost-sensitive / high volume | DeepSeek V3, Gemini 2.0 Flash | Frontier quality at fractions of GPT-4o cost |
| On-premise (data stays in your server) | Llama 4, Mistral, DeepSeek R1 | Open weights — self-host |
| Indian languages | Sarvam AI, Qwen 3 | Specific Indic language training |
| Image generation | Midjourney v6 (art), DALL-E 3 (product) | Different aesthetic strengths |
| Video generation | Sora, Veo 2 | Similar quality; Sora has API |
| Realtime news / current events | Grok 3 | Live X data feed |
| Protein folding / drug discovery | AlphaFold 3 | Nothing else competes |
What It All Means
The AI model landscape of 2025–2026 is not one horse race — it is six parallel races. OpenAI leads on product experience and ecosystem. Anthropic leads on safety research and autonomous agent capability. Google leads on context window and multimodal depth. Meta has made open-source mainstream. DeepSeek proved that compute efficiency can neutralise a $100 billion advantage. And Mistral, Qwen, Sarvam, and Krutrim are demonstrating that frontier AI is not the exclusive property of Silicon Valley.
The practical implication: for most tasks, you now have options across every price point, every data residency requirement, and every language. The question is no longer "can AI do this?" but "which AI, deployed how, at what cost?"
The next article in this series covers how to evaluate and test any model for your specific use case — before committing to a production deployment.
Quick Reference — All Major Models
| Model | Organisation | Type | Context | Best For |
|---|---|---|---|---|
| GPT-4o | OpenAI | Closed | 128K | General, multimodal |
| o1 / o3 | OpenAI | Closed | 128K | Reasoning, math |
| GPT-5 | OpenAI | Closed | 128K | General flagship |
| Claude 4 Sonnet | Anthropic | Closed | 200K | Coding, balanced |
| Claude 4 Opus | Anthropic | Closed | 200K | Autonomous agents |
| Claude Sonnet 4.6 | Anthropic | Closed | 200K | Cost-effective daily use |
| Gemini 2.5 Flash | Closed | 1M | Fast, affordable | |
| Gemini 2.5 Pro | Closed | 2M | Long context, multimodal | |
| Llama 4 Scout | Meta | Open | 10M | Open, massive context |
| Llama 4 Maverick | Meta | Open | 1M | Open, STEM |
| DeepSeek R1 | DeepSeek | Open | 64K | Reasoning, free |
| DeepSeek V3 | DeepSeek | Open | 128K | General, low cost |
| Mistral Large 2 | Mistral | Closed | 128K | European compliance |
| Mixtral 8x7B | Mistral | Open | 32K | On-premise, MoE |
| Codestral | Mistral | Open | 32K | Code specialist |
| Grok 3 | xAI | Closed | 131K | Realtime news |
| Qwen 3 235B | Alibaba | Open | 128K | Multilingual, Indian scripts |
| Phi-4 | Microsoft | Open | 16K | On-device, constrained |
| DALL-E 3 | OpenAI | Closed | — | Image generation |
| Midjourney v6 | Midjourney | Closed | — | Art-quality images |
| Stable Diffusion 3.5 | Stability AI | Open | — | Custom image fine-tuning |
| Sora | OpenAI | Closed | — | Video generation |
| Veo 2 | Closed | — | Video generation | |
| Whisper | OpenAI | Open | — | Speech transcription, 99 languages |
| AlphaFold 3 | Google DeepMind | Closed | — | Protein structure prediction |
| Krutrim | Krutrim/Ola | Closed | — | Indian languages, India-only |
| Sarvam-1 | Sarvam AI | Open | — | 22 Indian languages |
Sources: Anthropic public blog (Claude 4 announcement, May 2025); OpenAI system cards (GPT-4o, o3); Google DeepMind (Gemini 2.5 technical report); Meta AI (Llama 4 release notes); DeepSeek technical report (January 2025); Mistral AI documentation; Vellum LLM Leaderboard (2026 edition); Artificial Analysis model comparison (June 2026); Sarvam AI press release (2025); Krutrim launch announcement (2024). Benchmarks: MATH-500, MMMU, SWE-bench, GPQA Diamond, LiveCodeBench sourced from respective technical reports.
