A structured reference map of the AI vendor landscape, open model ecosystem, and best tools by work type. Published every Saturday
| Capability Layer | OpenAI | Anthropic | Meta | Microsoft | Popularity | |
|---|---|---|---|---|---|---|
| Core Models | Gemini 2.5 Pro, Gemini 2.5 Flash | GPT-5, GPT-5.6 Sol, GPT-5.5-Cyber | Claude Opus 4, Claude Sonnet 4 | Llama 4 Scout, Llama 4 Maverick | Phi-4, Phi-4-mini | |
| APIs / Platform | Vertex AI, Google AI Studio | OpenAI API, Responses API, Realtime API | Anthropic API, Claude.ai Teams | Meta AI API, Llama API | Azure OpenAI Service, Azure AI Foundry | |
| Dev Tools | Gemini Code Assist, Firebase Genkit | Codex, Codex Security, Codex CLI | Claude Code, MCP SDK | Meta Llama Stack, Code Llama | GitHub Copilot, Azure AI SDK | |
| Assistants / Apps | Gemini App, Google One AI Premium, NotebookLM | ChatGPT (GPT-5, GPT-5.5 Instant), ChatGPT Enterprise | Claude.ai, Claude for Enterprise | Meta AI (WhatsApp, Instagram, Facebook) | Microsoft 365 Copilot, Copilot+ PC | |
| Research Tools | NotebookLM Plus, AMIE (medical AI), Google Scholar AI | Deep Research, GeneBench-Pro, LifeSciBench | Claude Research Mode, Projects | Llama 4 research preview, Meta FAIR releases | Microsoft Research AI, Azure AI Search | |
| Creative / Media | Imagen 4, Veo 3, MusicFX | DALL-E 4, Sora, Voice Mode | Claude Artifacts, Claude creative writing | Meta Movie Gen, Emu Video | Microsoft Designer, Copilot Image Creator | |
| Edge / On-device | Gemini Nano (Android, Chrome), Google AI Edge SDK | GPT-4o mini (edge-optimized), OpenAI on-device via HP Frontier | Claude Haiku (lightweight API tier) | Llama 4 Scout (on-device), Llama.cpp ecosystem | Phi-4-mini (on-device), Windows Copilot Runtime | |
| Voice / Speech | Gemini Live, Google Cloud Speech-to-Text v2, NotebookLM audio | Advanced Voice Mode (GPT-5), Realtime API, GPT-5.5 Instant health voice | Claude voice via third-party integrations | Meta AI voice (Ray-Ban glasses), SeamlessM4T | Azure AI Speech, Copilot voice mode | |
| Agent Frameworks | Vertex AI Agent Builder, Agent Garden, Agentspace | Assistants API, Swarm, Operator, Responses API with tools | Claude Agents, Model Context Protocol (MCP) | Llama Agentic System, Meta Llama Stack agents | AutoGen, Semantic Kernel, Copilot Studio |
| Vendor | Models | License | Practical Use |
|---|---|---|---|
| Meta | Llama 4 Scout (17Bx16E MoE), Llama 4 Maverick (17Bx128E MoE), Llama 4 Behemoth (preview) | Llama 4 Community License (open weights, commercial use with conditions) | Best-in-class open weights for chat, coding, and multimodal tasks; Scout fits consumer GPUs; Maverick targets server-scale inference; strong choice for self-hosted enterprise deployments |
| Gemma 3 (1B, 4B, 12B, 27B), Gemma 3n (mobile-optimized) | Gemma Terms of Use (permissive commercial use) | Excellent for on-device and edge inference; Gemma 3 27B competes with much larger models; Gemma 3n optimized for phones and embedded; strong multilingual support | |
| Mistral | Mistral Large 2, Mistral Small 3.1, Codestral 2, Devstral | Apache 2.0 (Small/Codestral variants); Mistral Research License (Large) | Codestral and Devstral are top choices for code generation and agentic coding; Mistral Small 3.1 is highly efficient for RAG and summarization pipelines; strong European compliance posture |
| DeepSeek | DeepSeek-V3, DeepSeek-R1, DeepSeek-Coder-V2 | MIT (DeepSeek-R1, DeepSeek-V3 weights) | DeepSeek-R1 is a leading open reasoning model rivaling GPT-5 on benchmarks at fraction of cost; V3 excels at general tasks; Coder-V2 strong for code; best value-per-parameter in open ecosystem |
| Alibaba/Qwen | Qwen3 (0.6B to 235B MoE), Qwen3-Coder, QwQ-32B | Apache 2.0 (most variants) | Qwen3 235B-A22B MoE matches frontier models at low activation cost; QwQ-32B is a strong open reasoning model; excellent multilingual especially Chinese; broad size range suits edge to server |
| Microsoft | Phi-4 (14B), Phi-4-mini (3.8B), Phi-4-multimodal | MIT | Phi-4 punches far above its weight for reasoning and coding given its size; Phi-4-mini ideal for on-device and low-latency inference; multimodal variant handles vision tasks on constrained hardware |
| Other (Community) | Falcon 3 (TII), Command R+ (Cohere), OLMo 2 (AI2) | Apache 2.0 / various permissive | Falcon 3 strong for multilingual and sovereign AI use cases; Command R+ purpose-built for RAG and tool use with grounding; OLMo 2 fully open training stack including data, best for research reproducibility |
Includes models actively used or relevant as of Saturday, 4 July 2026. Experimental or pre-release models excluded.
| AI Work Type | Best Closed (Speed / Quality) | Best Open (Control / Cost) | When to Use |
|---|---|---|---|
| Chat | ChatGPT (GPT-5) — deepest general reasoning, best tool integration; Claude Opus 4 — superior for nuanced long-form conversation and instruction following | Llama 4 Maverick — best open chat model for self-hosted; Qwen3 235B-A22B MoE — strong multilingual chat at scale | GPT-5 for power users needing research, tools, and breadth; Claude Opus 4 for writing-heavy or sensitive professional contexts; open models when data privacy or cost at scale is the constraint |
| Coding | Claude Sonnet 4 via Claude Code — leading agentic coding benchmark performance; Codex (OpenAI) — strong for repo-level autonomous tasks | Devstral (Mistral) — top open-source agentic coding model; DeepSeek-Coder-V2 — strong code completion and generation | Claude Code / Sonnet 4 for complex multi-file agentic tasks; Codex for CI pipeline integration; Devstral for self-hosted coding assistants; DeepSeek-Coder-V2 when MIT license and cost matter |
| RAG | Gemini 2.5 Pro (2M token context, native retrieval); Claude Sonnet 4 (strong instruction following for cited responses) | Command R+ (Cohere, purpose-built for RAG with grounding citations); Mistral Small 3.1 (efficient embedding + generation pipeline) | Gemini 2.5 Pro when corpus fits in context and latency allows; Claude for compliance-sensitive RAG needing precise citation; Command R+ for production RAG with grounding out-of-box; Mistral for cost-efficient high-volume RAG |
| Agents | GPT-5 via OpenAI Responses API + Operator — best tool-calling reliability; Claude Opus 4 via MCP — strong for multi-step reasoning agents | Llama 4 Maverick via Llama Stack — best open agentic backbone; Qwen3-235B — strong tool-use and function calling | OpenAI stack for production agents needing broad tool ecosystem; Claude+MCP for structured agentic workflows with external integrations; open models for sovereign or air-gapped agent deployments |
| Search | Perplexity Pro (real-time web, citations); ChatGPT Search (GPT-5 with web browsing); Gemini with Google Search grounding | OLMo 2 + retrieval pipeline; Mistral + SearXNG for self-hosted search augmentation | Perplexity for research-grade real-time search with citations; ChatGPT Search for integrated conversational search; Gemini for Google-ecosystem grounded answers; open stack when search data must stay on-premises |
| Image | Imagen 4 (Google) — photorealism and prompt fidelity; DALL-E 4 (OpenAI) — creative generation with ChatGPT integration | FLUX.1 (Black Forest Labs) — top open image generation; Stable Diffusion 3.5 Large — flexible fine-tuning | Imagen 4 for commercial photorealistic output; DALL-E 4 for rapid ideation inside ChatGPT; FLUX.1 for self-hosted production pipelines; SD 3.5 when fine-tuning on custom brand style is required |
| Video | Veo 3 (Google) — state-of-the-art video generation with audio; Sora (OpenAI) — cinematic long-form video generation | Wan 2.1 (Alibaba) — best open video generation model; CogVideoX (THUDM) — open, commercially usable | Veo 3 for highest-quality video with synchronized audio; Sora for creative long-form storytelling; Wan 2.1 for self-hosted video workflows; CogVideoX for open research and prototyping |
| Speech | OpenAI Realtime API (GPT-5 voice, lowest latency); Google Cloud Speech-to-Text v2 (best transcription accuracy); ElevenLabs (best TTS expressiveness) | Whisper large-v3 (OpenAI, open ASR); Kokoro TTS (open, lightweight); Coqui XTTS-v2 (voice cloning) | OpenAI Realtime API for live conversational voice agents; Google STT v2 for high-volume transcription pipelines; ElevenLabs for premium TTS narration; Whisper for offline or privacy-sensitive transcription |
| Edge | Gemini Nano (on-device Android/Chrome); GPT-4o mini via HP Frontier edge deployments | Phi-4-mini (Microsoft, MIT, 3.8B — best reasoning per parameter at edge); Gemma 3n (Google, mobile-optimized); Llama 4 Scout (fits consumer GPU) | Gemini Nano for Android-native on-device AI; Phi-4-mini for edge servers and mobile apps needing strong reasoning; Gemma 3n for battery-constrained mobile inference; Llama 4 Scout for local workstation deployments |
| Fine-tuning | OpenAI fine-tuning API (GPT-4o mini, GPT-4o); Google Vertex AI supervised tuning (Gemini 2.0 Flash) | Llama 4 Scout + TRL/Unsloth — most popular open fine-tuning target; Qwen3-7B — efficient LoRA fine-tuning; Mistral Small 3.1 — strong base for domain adaptation | OpenAI fine-tuning for production with minimal MLOps overhead; Vertex for GCP-native teams; Llama 4 Scout for fully custom models on owned hardware; Qwen3 when multilingual domain adaptation is needed |
| Multimodal | Gemini 2.5 Pro (text, image, video, audio, code natively); GPT-5 (text, image, audio, video understanding); Claude Opus 4 (text and vision) | Llama 4 Scout/Maverick (text + image natively); Qwen3-VL (strong vision-language); Phi-4-multimodal (vision + speech on edge) | Gemini 2.5 Pro for broadest native modality support in one model; GPT-5 for multimodal reasoning with tool use; Llama 4 for self-hosted image+text pipelines; Phi-4-multimodal for constrained-device vision tasks |
Closed = prioritises performance and ease of use. Open = prioritises control and cost. Updated weekly.
The AI tool landscape shifts weekly — vendor moves, model releases, capability changes. Get the daily brief that tracks what's actually shipping. Five issues a week, in your inbox.
No spam. Unsubscribe in one click. Powered by Resend.