Anthropic — Claude
Opus (deepest reasoning), Sonnet (balanced workhorse), Haiku (fast/cheap), Fable (creative). Strong at careful, long-form, tool-using work.
claude-opus · sonnet · haiku · fable
Reference hub
All that is AI, in one place.
Frontier models, the agents that run them, the tools that research with them, the concepts that explain them, and the protocols that connect them. A living reference — curated, opinionated, current.
Frontier models
The labs and their current families. Names move fast — treat this as a map, not a spec sheet.
OpenAI — GPT
GPT-5 family incl. Pro/reasoning tiers. Pro/reasoning models are served by the Responses API, not chat-completions.
gpt-5 · gpt-5-pro · o-seriesGoogle — Gemini
Gemini 2.5 Pro / Flash and successors. Huge context, multimodal, deep Google integration.
gemini-2.5-pro · flashxAI — Grok
Real-time data via X; strong on current events / live context.
grokMeta — Llama
Leading open-weight family; self-host or fine-tune.
open weightsNVIDIA — Nemotron
Nemotron 3 (incl. Super, ~120B): open, agentic-reasoning; available via Perplexity Agent API, OpenRouter, build.nvidia.com.
open · agenticOthers worth tracking
Mistral, DeepSeek, Qwen (Alibaba), Cohere — strong open / specialist models.
open · specialistCoding agents
Terminal / IDE agents that read a task and ship code, headless or attended.
Claude Code
Anthropic's CLI agent — deep multi-file work, repo-aware, careful. Native MCP.
CLI · MCPOpenAI Codex
OpenAI's CLI coding agent — fast, cheap contained slices and quick scaffolding.
CLICursor CLI
Headless cursor-agent -p for CI/scripts; the IDE agent in your terminal. JSON output, MCP-aware.
Google Antigravity
Go-based agy CLI + desktop app; command mode agy -p. Official replacement for the Gemini CLI.
Aider
Open-source pair-programmer; model-agnostic, git-native.
open · CLIResearch & tools
Where AI answers, searches, designs, and builds beyond raw chat.
Perplexity
Live web research with citations; Sonar models + an Agent API (hosts open models incl. Nemotron).
researchStitch (Google)
UI generation / design-to-code — fast visual mocks.
designManus
Agentic multi-step deliverables; branding / research.
agenticAI gateways
Portkey, Cloudflare AI Gateway, Kong, LiteLLM — routing, observability, guardrails over model calls.
infraMemory & connectors
Mem0, Zep, Letta (memory); Composio, Zapier MCP, Smithery, Pipedream (tool/MCP aggregation).
infraCore concepts
The vocabulary, plainly.
Tokens
The chunks models read/write in. Pricing and limits are counted in tokens, not words.
Context window
How much a model can "see" at once. Bigger = more input, but cost + latency scale with it.
Reasoning models
Models that "think" before answering (hidden reasoning tokens). Better at hard problems, slower/pricier.
RAG
Retrieval-Augmented Generation — fetch relevant documents, feed them in, so answers are grounded in your data.
Embeddings
Text → vectors, so you can search by meaning. The backbone of RAG and semantic memory.
Tool use / function calling
The model calls functions/APIs you define, instead of just talking. The basis of agents.
Agents
A model in a loop: plan → call tools → observe → repeat, until a goal is met.
Fine-tuning
Further-training a base model on your data to specialize it. Powerful, but prompt + RAG often get you there cheaper.
Confluence (multi-LLM)
Fan one prompt across several models, then synthesize — diversity catches what one model misses.
Prompt engineering
Shaping inputs (context, examples, constraints, format) to steer outputs reliably.
Prompt caching
Reuse a fixed prefix (system prompt, docs, canon) across calls so the provider skips re-processing it — large cost and latency cuts when the same context repeats.
Structured output
Constrain a model to valid JSON or a schema (or a tool call) so its output is machine-parseable, not free text. The reliable bridge from model to code.
Multimodal
Models that take or emit more than text — images, audio, video, PDFs — inside the same context.
Evals
Systematic tests (graded datasets, LLM-as-judge) that measure whether a prompt, model, or agent actually works — before and after a change.
Mixture of Experts (MoE)
A model split into many "expert" sub-networks; each token routes to only a few, so a very large model runs at a fraction of the compute.
Quantization & distillation
Two ways to shrink a model for cheaper, faster inference — store weights at lower precision (quantization), or train a small model to mimic a big one (distillation).
Hallucination & grounding
Models can state false things confidently. Grounding — RAG, citations, tool calls — ties answers to real sources to reduce it.
Reasoning effort
A dial on reasoning models: spend more hidden "thinking" tokens (slower, pricier) for better answers on hard problems, fewer for quick ones.
Protocols & interop
The plumbing that lets models, tools, and agents talk — and increasingly, talk to each other.
MCP — Model Context Protocol
An open standard (from Anthropic) for connecting AI assistants to external tools and data. An MCP server exposes tools/resources; an MCP client (Claude Desktop/Cowork, Cursor, Antigravity, …) connects and uses them. Increasingly the universal connector across the ecosystem — one integration, many clients.
open standardA2A — Agent2Agent
An open protocol for agents to discover and coordinate each other across vendors and platforms. Started by Google, donated to the Linux Foundation (2025); v1.0 landed early 2026 with signed Agent Cards for verifiable identity. The complement to MCP: MCP connects an agent to its tools; A2A connects agents to one another.
open standard · 150+ orgsOpenAI-compatible API
The /v1/chat/completions shape became the de-facto wire format. xAI, many gateways, and local servers (Ollama, vLLM) all speak it — so one client can target many backends.