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 4 Carbon (frontier, announced), Gemini 4 Argon (live) | GPT-6 Astra, GPT-6.1 Sol, Codex Ultrafast | Claude Opus 4, Claude Sonnet 4 | Meta Muse Spark Open (1M context), Llama 4 | Phi-4, Azure-hosted OpenAI GPT-6 series | |
| APIs / Platform | Gemini API, Google Cloud Vertex AI, Model Garden | OpenAI API, Decisions API, Codex API | Claude API (Google Cloud managed), Claude Platform | Meta AI API, Together AI (Muse Spark Open) | Azure OpenAI Service, Azure AI Foundry | |
| Dev Tools | EmbeddingGemma 2, Gemini Code Assist | Codex Ultrafast, OpenAI Daybreak, Decisions API | Claude Code Review (GA), Claude Code Mods, Claude Code Auto Mode | Thin Orchestrator pattern, Meta Andromeda ads SDK | GitHub Copilot (GPT-6 series), Azure AI Studio | |
| Assistants / Apps | Gemini App (AI Pro/Ultra), Gemini document/PDF/spreadsheet editing | ChatGPT, GPT-6 Astra voice/agent mode | Claude.ai, Claude Duet | Meta AI, Meta Muse (Singapore live), Meta Forum | Microsoft Copilot, Copilot 365 | |
| Research Tools | Genie 4 (world model, interactive simulation), SynthID standalone detector | OpenAI Daybreak (vulnerability/security research), misalignment detection methodology | Claude for Financial Advisors, Agentic AI Foundation skills | Meta university research partnerships (5 US universities), Meta AI medical/scientific | Azure Research AI, Turing models | |
| Creative / Media | Gemini image/presentation generation, SynthID watermarking, Immersive Explorations | DALL-E 3, Sora | Claude creative writing (Sonnet 4) | Meta Movie Gen, Meta AI image generation | Azure OpenAI DALL-E, Designer (Copilot) | |
| Edge / On-device | Gemma 3 (on-device), EmbeddingGemma 2 | GPT-6 Ultrafast (low-latency inference) | Claude Haiku 4 (lightweight tier) | Llama 4 Scout (edge-optimized), Meta device AI | Phi-4 (on-device, small models), Windows Copilot Runtime | |
| Voice / Speech | Gemini Live, Genie 4 native audio | GPT-6 Astra (voice, tops Retell benchmark), GPT-Live-1 | Claude voice (via API integrations) | Meta voice AI (Muse agent) | Azure Speech Services, Copilot voice mode | |
| Agent Frameworks | Gemini Agentic Coding (Argon/Carbon), Vertex AI Agent Builder | Codex agents, OpenAI Agents SDK, ServiceNow integration | Claude Code agents, Claude Duet, Cyber Verification Program | Thin Orchestrator pattern, Meta Muse personal agent | AutoGen, Semantic Kernel, Azure AI Agent Service |
| Vendor | Models | License | Practical Use |
|---|---|---|---|
| Meta | Llama 4 Maverick, Llama 4 Scout, Meta Muse Spark Open (1M context, on Together AI) | Llama 4 Community License (open weights, commercial with conditions) | General chat, agentic coding, long-context RAG, fine-tuning on custom data; Muse Spark Open strong for agentic pipelines |
| Mistral | Mistral Large 4 'Le Chonk' (1T params, multimodal, best open coding), Voxtral 3 (voice agent SDK, 50+ languages), Timestral (forecasting, 5M–209M params) | Apache 2.0 / Mistral Research License (model-dependent); EU-trained | Large 4: top open-weight coding and reasoning; Voxtral 3: production voice agents with tool use; Timestral: CPU-efficient time-series forecasting; all self-hostable, EU-compliant |
| Gemma 3 (on-device/self-host), EmbeddingGemma 2 (multimodal embeddings, open lightweight) | Gemma Terms of Use (open weights, commercial allowed) | Gemma 3: on-device inference, fine-tuning; EmbeddingGemma 2: self-hosted multimodal RAG embeddings without cloud API dependency | |
| DeepSeek | DeepSeek V3, DeepSeek R2 | MIT (open weights) | Strong reasoning and coding at low cost; self-hostable; competitive on math and science benchmarks; good for cost-sensitive production inference |
| Alibaba/Qwen | Qwen 3 (235B MoE flagship), Qwen 3 (dense variants 0.6B–32B) | Qwen License (open weights, commercial with conditions) | Multilingual (strong Chinese/English), coding, instruction following; MoE variant for high-throughput; small variants for edge deployment |
| Microsoft | Phi-4 (14B), Phi-4-mini (3.8B) | MIT | Small, efficient models optimised for reasoning on limited hardware; strong for on-device and edge scenarios; fine-tuning on domain-specific data |
| Reflection (Community) | Reflection Beam (501B open-weight, agentic coding and reasoning) | Open weights (Hugging Face); free on Ollama Cloud (limited window) | Large-scale agentic coding, long-running reasoning tasks; 1M-token context on Together AI serverless; self-hostable for research and enterprise evaluation |
Includes models actively used or relevant as of Saturday, 31 October 2026. Experimental or pre-release models excluded.
| AI Work Type | Best Closed (Speed / Quality) | Best Open (Control / Cost) | When to Use |
|---|---|---|---|
| Chat | ChatGPT (GPT-6 Astra) — best general-purpose; Claude Sonnet 4 — nuanced long-form and collaborative tasks via Claude Duet | Mistral Large 4 (self-hosted); Meta Muse Spark Open (1M context, Together AI) | ChatGPT for broad consumer and enterprise chat; Claude Sonnet 4 when collaborative, document-heavy, or safety-sensitive; open models when data privacy or cost at scale is the constraint |
| Coding | Codex Ultrafast / GPT-6.1 Sol (OpenAI) — top agentic coding and incident response; Claude Code with Code Review GA (Anthropic) — strong PR review and autonomous sessions | Mistral Large 4 'Le Chonk' — best open-weight coding model; Reflection Beam 501B — agentic long-running tasks | Codex/GPT-6.1 Sol for speed-critical CI and security agents; Claude Code for opinionated review pipelines; Mistral Large 4 when you need open-weight coding parity with EU data residency |
| RAG | Gemini 4 Argon via Gemini API (long context, native multimodal retrieval); Claude Sonnet 4 (200K context) | EmbeddingGemma 2 (self-hosted multimodal embeddings); Meta Muse Spark Open (1M-token context) | Gemini 4 Argon for large multimodal corpora on GCP; Claude for long legal/financial docs; EmbeddingGemma 2 when you need self-hosted embeddings to eliminate cloud API costs |
| Agents | GPT-6 Astra + Codex (OpenAI Agents SDK); Claude Code Auto Mode + Duet (Anthropic); Gemini 4 Argon agentic coding | Meta Thin Orchestrator pattern + Llama 4 / Muse Spark Open; Mistral Large 4 with Voxtral 3 for voice agents | GPT-6 Astra for voice + tool-use agents at scale (Retell validated at 100M calls/month); Claude Code for code-focused autonomous agents; open stack when sovereign deployment or cost control is required |
| Search | ChatGPT with web search (GPT-6 Astra); Gemini App with Google Search grounding | Perplexity (open API), self-hosted RAG with EmbeddingGemma 2 + open LLM | ChatGPT/Gemini for real-time web-grounded answers; self-hosted RAG for private knowledge bases where data cannot leave your environment |
| Image | DALL-E 3 (OpenAI/Azure); Gemini image generation (native in Gemini App) | Stable Diffusion 3.5 (self-hosted); FLUX.1 variants (Hugging Face) | DALL-E 3 for polished consumer/enterprise creative; Gemini for integrated document+image workflows; open models for high-volume generation or fine-tuned style control |
| Video | Sora (OpenAI); Veo 2 (Google DeepMind) | Wan 2.1 (open-weight video generation, Hugging Face) | Sora for high-quality short-form video generation; Veo 2 for cinematic and longer outputs; open models for local experimentation or pipelines requiring full model control |
| Speech | GPT-6 Astra / GPT-Live-1 (OpenAI) — tops Retell benchmark; Voxtral 3 via Mistral hosted API — 50+ languages, emotional TTS | Phonon via Ollama (local TTS, 5 languages, 100x real-time on Apple Silicon); Voxtral 3 (self-hosted, 50+ languages) | GPT-6 Astra for production voice agents at scale; Voxtral 3 for multilingual or EU-hosted voice; Phonon/Ollama for private local TTS on Apple Silicon without cloud dependency |
| Edge | Claude Haiku 4 (low-latency API tier); GPT-6 Ultrafast (OpenAI) | Phi-4-mini (3.8B, MIT, on-device reasoning); Gemma 3 (2B/4B, on-device); Qwen 3 dense (0.6B–7B) | Phi-4-mini or Gemma 3 for on-device inference on constrained hardware; Qwen 3 small variants for multilingual edge use; closed APIs for edge-adjacent low-latency when local deployment isn't feasible |
| Fine-tuning | OpenAI fine-tuning API (GPT-4o-mini base); Google Vertex AI fine-tuning (Gemma 3 / Gemini) | Mistral Large 4 (full weights, EU-trainable); Llama 4 Scout (efficient fine-tuning); Phi-4 (MIT, strong base for domain adaptation) | OpenAI/Vertex for managed fine-tuning without infrastructure overhead; open weights when you need full control, data sovereignty, or iterative experimentation on custom hardware |
| Multimodal | Gemini 4 Argon/Carbon (text, image, audio, video, code); GPT-6 Astra (text, image, voice, tool use) | Mistral Large 4 (multimodal, 1T params, open weights); EmbeddingGemma 2 (multimodal embeddings, self-hosted) | Gemini 4 series for deep Google ecosystem integration and document+vision pipelines; GPT-6 Astra for voice+vision agent workflows; Mistral Large 4 when open-weight multimodal is required for compliance or self-hosting |
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.