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 3.6 Flash, Gemini Ultra | GPT-5.6 Sol, GPT-5.6 Luna | Claude Opus 5, Claude Sonnet 5 | Muse Glimmer (30B, MSL), Llama 4 | Phi-4, Azure OpenAI GPT-5.6 | |
| APIs / Platform | Gemini API, Google AI Studio, Vertex AI | OpenAI API (Responses API), Zero Data Retention tier | Anthropic API, Claude Enterprise (inference hooks, DLP) | Meta AI API, Llama API | Azure OpenAI Service, Azure AI Foundry | |
| Dev Tools | Gemini API Managed Agents (hooks + triggers) | Codex, GPT-5.6 builder guide, Responses API | Claude Code (auto mode, BYO-compute sessions) | Muse Glimmer via Ollama MLX, Meta Startup School | GitHub Copilot, Azure AI Studio | |
| Assistants / Apps | Gemini (consumer), Gemini for Workspace | ChatGPT (GPT-5.6 Luna free tier, ChatGPT Work) | Claude.ai (Sonnet 5 default) | Meta AI (WhatsApp, Instagram, Facebook) | Microsoft Copilot, Copilot+ PCs | |
| Research Tools | AMIE (clinical AI), NotebookLM | Deep Research (ChatGPT), Private Safety Processing preview | Claude Opus 5 (long-context research) | Meta FAIR research models | Microsoft Research AI models | |
| Creative / Media | Imagen 4, Veo 3 | DALL-E 4, Sora | Fable 5 (narrative/creative, redeployed) | Emu Video, Make-A-Video | Designer (DALL-E powered), Clipchamp AI | |
| Edge / On-device | Gemini Nano (Pixel on-device) | GPT-5.6 Luna (lightweight API tier) | Claude Haiku (smallest production tier) | Muse Glimmer via Ollama MLX (Apple Silicon), Llama 4 Scout | Phi-4 Mini (on-device, Copilot+ PC) | |
| Voice / Speech | Gemini Live, Google Search voice AI Mode | ChatGPT Voice (Advanced Voice Mode) | Claude voice (via third-party integrations) | Meta AI Voice (Ray-Ban glasses, AI Glasses for visually impaired) | Azure Speech, Copilot Voice | |
| Agent Frameworks | Gemini Managed Agents (3.6 Flash, hooks, triggers), Vertex AI Agent Builder | OpenAI Agents SDK, Codex CLI agents, Replit GPT-5.6 Luna agents | Claude Managed Agents (brain/hands decoupled), Claude Code agentic sessions | Muse Glimmer local coding agents (Ollama), LlamaIndex integrations | AutoGen, Semantic Kernel, Azure AI Agent Service |
| Vendor | Models | License | Practical Use |
|---|---|---|---|
| Meta | Muse Glimmer 30B (multimodal, MSL), Llama 4 Scout, Llama 4 Maverick | Apache 2.0 (Muse Glimmer, Llama 4) | Local coding agents with image input via Ollama MLX; on-device multimodal tasks; Apple Silicon accelerated inference with DFlash; general chat and reasoning |
| Mistral | Mistral Large 3, Shieldstral 3B (safety classifier) | Apache 2.0 (Shieldstral, select models); Mistral Research License (Large 3) | Shieldstral: drop-in multimodal safety classifier for moderation pipelines, outperforms 21B models at 3B; Mistral Large 3: general reasoning and multilingual tasks; Agentic Search for complex document retrieval beyond naive RAG |
| DeepSeek | DeepSeek V4 Pro 0813, DeepSeek-V4 Flash | MIT (DeepSeek V4 Flash); DeepSeek License (V4 Pro) | V4 Pro: best pass@4 on coding tasks (DeepSWE), beats GPT-5.6 Sol and Claude Fable 5 in coding cascades at lower cost; V4 Flash: 4.8x coding solves per dollar vs GPT-5.6 Luna — ideal for high-volume code generation cost routing |
| Alibaba/Qwen | Qwen3 72B, Qwen3 235B-A22B (MoE) | Apache 2.0 | Strong multilingual reasoning; MoE variant for cost-efficient large-scale inference; competitive on math and coding benchmarks; self-hostable on Together AI or local GPU clusters |
| Gemma 3 27B, Gemma 3 12B | Gemma Terms of Use (permissive commercial use allowed) | Lightweight on-device and edge inference; fine-tuning base for domain-specific tasks; strong instruction-following at small parameter counts; integrates with Keras and JAX pipelines | |
| Microsoft | Phi-4 14B, Phi-4 Mini 3.8B | MIT | Phi-4: strong reasoning and coding at small scale for on-device or CPU inference; Phi-4 Mini: Copilot+ PC on-device tasks, fine-tuning base for edge deployment; best-in-class reasoning per parameter for sub-14B models |
| NVIDIA | Nemotron 3.5 Lightning 30B (3B active MoE) | NVIDIA Open Model License | Long-running agentic tasks via Ollama; persistent tool-calling with low memory footprint thanks to MoE active-parameter efficiency; optimized for multi-turn agent sessions on local hardware |
| Moonshot AI (Kimi) | Kimi K3 (3T parameter class) | Open weights (research license) | First open 3-trillion-parameter-class model; accessible via Together AI API; frontier-scale capability for research teams; benchmark evaluation ongoing — not yet fully characterized for production use |
Includes models actively used or relevant as of Saturday, 22 August 2026. Experimental or pre-release models excluded.
| AI Work Type | Best Closed (Speed / Quality) | Best Open (Control / Cost) | When to Use |
|---|---|---|---|
| Chat | ChatGPT (GPT-5.6 Sol) for depth; Claude Sonnet 5 for nuanced reasoning and long context | Llama 4 Maverick (self-hosted) or Muse Glimmer 30B via Ollama for local privacy-first chat | Use GPT-5.6 Sol for complex multi-turn conversations requiring broad knowledge; Claude Sonnet 5 when tone, nuance, and instruction-following fidelity matter; open models when data must not leave your infrastructure |
| Coding | Claude Code (Sonnet 5, auto mode, BYO-compute) for agentic coding; GPT-5.6 Sol via Replit for interactive development | DeepSeek V4 Pro 0813 for pass@4 coding quality; DeepSeek-V4 Flash for cost-optimized high-volume code generation | Use Claude Code auto mode for long agentic sessions in regulated environments (BYO-compute removes data egress); DeepSeek V4 Pro when you need best-in-class coding quality at lower API cost; DeepSeek-V4 Flash for bulk code generation where cost-per-solve is primary constraint |
| RAG | Claude Sonnet 5 (long context, 200K tokens) with Anthropic inference hooks for DLP; Gemini 3.6 Flash Managed Agents with hooks and triggers | Mistral Agentic Search for complex document navigation; DeepSeek V4 Pro for retrieval-augmented coding tasks | Use Claude Sonnet 5 when documents are sensitive and inline DLP via inference hooks is required; Gemini Managed Agents for event-driven RAG pipelines already in Google Cloud; Mistral Agentic Search when you need structured multi-hop retrieval beyond naive vector search |
| Agents | Claude Code (auto mode, managed agents brain/hands architecture); Gemini Managed Agents (3.6 Flash, hooks + triggers) | Muse Glimmer 30B via Ollama for local coding agents; NVIDIA Nemotron 3.5 Lightning for persistent tool-calling agents on local hardware | Use Claude managed agents when you need decoupled orchestration with enterprise compliance (DLP hooks, BYO-compute); Gemini Managed Agents for Google Cloud-native pipelines with event-driven triggers; Muse Glimmer locally when agent tasks involve images and data cannot leave on-premise |
| Search | ChatGPT with Deep Research (GPT-5.6 Sol); Google AI Mode in Search (Gemini-powered) | Mistral Agentic Search for self-hosted enterprise search over internal documents | Use ChatGPT Deep Research for multi-source web synthesis; Google AI Mode for consumer search with Gemini reasoning; Mistral Agentic Search when building internal knowledge retrieval that must remain on-premise with accurate multi-hop navigation |
| Image | Google Imagen 4 for photorealism; OpenAI DALL-E 4 for creative and diverse styles | Muse Glimmer 30B (multimodal image input/understanding); Stable Diffusion 3.5 for local generation | Use Imagen 4 for product and marketing photography quality; DALL-E 4 for creative illustration and concept art; Muse Glimmer when you need multimodal understanding of images as agent input rather than generation; open diffusion models for unconstrained local generation |
| Video | Google Veo 3 for cinematic video generation; Sora (OpenAI) for creative video | No production-ready fully open video generation model at frontier quality as of today | Use Veo 3 for high-quality marketing and cinematic content with audio; Sora for creative and experimental video generation; expect open video models to mature in the next one to two quarters |
| Speech | OpenAI Advanced Voice Mode (GPT-5.6 powered); Google Gemini Live for real-time voice | Whisper large-v3 for transcription; benchmark-optimized ASR models tracked via Hugging Face leaderboard | Use OpenAI Advanced Voice Mode for low-latency conversational voice agents; Gemini Live for Google ecosystem integration; Whisper large-v3 when you need self-hosted, accurate transcription with no data sent to cloud |
| Edge | Gemini Nano (Pixel on-device); Microsoft Phi-4 Mini via Copilot+ PC | Muse Glimmer via Ollama MLX (Apple Silicon); Phi-4 Mini (MIT license); Gemma 3 12B | Use Gemini Nano for Android/Pixel on-device inference with Google ecosystem; Phi-4 Mini for Windows Copilot+ PC local inference; Muse Glimmer via Ollama MLX when you need multimodal capability locally on Apple Silicon with maximum license permissiveness |
| Fine-tuning | OpenAI fine-tuning API (GPT-5.6 Luna); Google Vertex AI supervised fine-tuning | Llama 4 Scout (Apache 2.0, efficient fine-tuning base); Phi-4 14B (MIT, strong reasoning base); Qwen3 72B (Apache 2.0, multilingual) | Use OpenAI fine-tuning when you want managed infrastructure with no GPU ops overhead; Llama 4 Scout for domain-specific tasks requiring full weight ownership under permissive license; Phi-4 for fine-tuning where inference must run on constrained hardware post-training |
| Multimodal | GPT-5.6 Sol (text, image, code); Claude Opus 5 (text, image, long-document); Gemini 3.6 Flash (text, image, video, audio) | Muse Glimmer 30B (text + image, Apache 2.0, Ollama-native); Kimi K3 (3T-class, text + multimodal, Together AI API) | Use GPT-5.6 Sol for broad multimodal tasks with strong reasoning; Claude Opus 5 for long-document analysis combining text and images; Gemini 3.6 Flash when video and audio modalities are needed in managed agent pipelines; Muse Glimmer locally when multimodal capability must run on-premise under a permissive license |
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.