Knowing models is knowing who builds them, what they optimize for, and where they live on the cost/quality curve.
OpenAI
- GPT-4o family (GPT-4o, GPT-4o-mini): general-purpose, strong tool use, native multimodal. The most copy-pasted "default".
- o-series (o1, o3, o3-mini): reasoning-tuned, slow, expensive, much better at math/code/planning.
- GPT-4.5 (Orion): conversational, less aggressive reasoning. Niche.
- Strengths: ecosystem (function calling, structured output, batch API, fine-tuning).
- Watch: aggressive deprecations, rate-limit politics, no on-prem.
Anthropic
- Claude 4.x family (Opus, Sonnet, Haiku): the standard for long-context (200K+) reasoning, strong coding/agent steerability.
- Strengths: coding (top of SWE-bench), tool use, faithful instruction following, safety alignment.
- Watch: smaller ecosystem; no on-prem.
- Gemini family (Gemini 3.x Pro, Flash, Ultra): native multimodal (image, video, audio), 1M+ context, integrated with Vertex.
- Strengths: massive context, multimodal, Google ecosystem (BigQuery, Vertex).
- Watch: tool-calling reliability historically behind OpenAI/Anthropic; gap closing fast.
Meta — Llama family
- Llama 4 (8B, 70B, 405B): open weights, commercial-permissive license. Strong instruction-following at 70B+, runs on-prem.
- Strengths: open license, customization, on-prem, no vendor lock.
- Watch: reasoning lags frontier closed models; ops cost is yours.
Mistral
- Mistral Large, Mistral Small, Mixtral 8x22B (MoE), Codestral.
- Strengths: efficient MoE, strong open / managed mix, EU data residency story.
- Watch: ecosystem smaller than OpenAI; quality varies sharply by tier.
Cohere
- Command R / R+: optimized for RAG out of the box (citations, structured outputs, multilingual).
- Embed v3, Rerank v3: best-in-class retrieval components.
- Strengths: enterprise, RAG-native, strong multilingual including Arabic.
- Watch: less general-purpose than GPT-4o for non-RAG tasks.
Qwen / DeepSeek
- Qwen 3 family (Alibaba): open weights, strong multilingual + tool use, top of many open benchmarks.
- DeepSeek-V3 / R1: open MoE, very strong on math/code at low cost.
- Strengths: cost, openness, increasingly competitive frontier.
- Watch: data residency / regulatory scrutiny in some regions.
IBM watsonx — Granite family
- Granite 3.x (8B, 13B, 20B): purpose-built for enterprise: code, RAG, function-calling, with full data lineage and IP indemnity.
- Strengths: regulated industry fit, IBM ecosystem (watsonx.data, watsonx.governance), audit trail, enterprise SLAs.
- When to pick: regulated finance, healthcare, government — anywhere indemnity and provenance matter more than absolute leaderboard rank.
The map you should keep in your head
- Frontier general: GPT-4o, Claude Opus 4.x, Gemini 3.x Pro.
- Frontier reasoning: o3, Claude 4.x with extended thinking.
- Cost-optimized: Claude Haiku, GPT-4o-mini, Gemini Flash, Llama 4 70B self-hosted.
- Agent / coding: Claude Sonnet 4.x.
- RAG-native: Cohere Command R+.
- Open / on-prem: Llama 4 70B, Qwen 3 72B, Mistral Large self-hosted.
- Enterprise / regulated: IBM Granite, Azure OpenAI with private endpoints.