Frontier model comparison · Prices per million tokens · Oct 2026

METR = autonomous task horizon (50% success) · SWE-bench Verified · — = not publicly reported · Prices change frequently, verify at provider

Foundation Models	Frontier model reference Claude (Anthropic) — Opus, Sonnet, Haiku family; best for coding & reasoning GPT-4o / o3 (OpenAI) — strongest general-purpose closed model Gemini 2.5 Pro (Google) — best long context (1M tokens) Llama 3 (Meta) — best open-source, free to self-host Mistral Large 4 — open-weight Mixture-of-Experts (1.05T params, 49B active); fast, multilingual, vision-capable DeepSeek R1 — open-source MoE, matches GPT-4 at near-zero cost; major attention in 2026 OpenRouter — single API to access 200+ models for comparison
Reasoning Models	Test-time compute — models that think before they answer arXiv — Test-Time Scaling: Comprehensive Survey (ACL 2026) — sampling-based, feedback-based and search-based approaches to inference compute Meta — UniT: Unified Multimodal Chain-of-Thought Test-Time Scaling — reason, verify and refine across text and images in multiple rounds OpenAI o3 — flagship reasoning model; extended thinking before answering Claude Extended Thinking — Anthropic's reasoning mode; visible scratchpad before final response Gemini 2.5 Thinking (Google) — reasoning model with 1M token context
Multimodal AI	Models that understand and generate across text, image, audio and video GPT-4o (OpenAI) — text, image, audio and video understanding in one model Gemini 2.5 Pro (Google) — native multimodal; strongest for video and long documents arXiv — Test-Time Scaling in Multimodal Foundation Models (ACL 2026) — survey of generation and reasoning across modalities Hugging Face — Vision-Language Models — open-source multimodal model catalogue ElevenLabs — state-of-the-art audio/voice AI; text to speech and speech to speech
Learning Resources	Courses, books and hands-on guides Sebastian Raschka — LLM from Scratch Awesome LLM Apps GenAI / LLM case studies Andrej Karpathy — Neural Networks: Zero to Hero — most recommended free course for understanding LLMs from first principles fast.ai — Practical Deep Learning — free, hands-on, highly respected DeepLearning.AI Short Courses — bite-sized courses on RAG, agents, prompt engineering Hugging Face LLM Course — free, covers transformers, fine-tuning, deployment end to end TensorFlow — 10,000 Word Vector Visualisation
Prompt Engineering	Crafting instructions to get better model outputs DAIR-AI — Prompt Engineering Guide — guides, papers, notebooks, RAG, agents promptslab — Awesome Prompt Engineering — curated papers, tools, APIs, datasets Awesome ChatGPT Prompts System Prompts and Models of AI Tools Anthropic Prompt Library — official curated prompts prompts.chat — public community prompts library DSPy (Stanford) — auto-optimises prompts using examples instead of hand-tuning; major shift in how prompt engineering is done
Context Engineering	Constructing and managing what goes into the context window Anthropic — Effective Context Engineering for AI Agents — the definitive official guide; treats context as a finite, curated resource arXiv — A Survey of Context Engineering for LLMs — comprehensive academic survey mlnjsh — Complete Guide to Context Engineering — papers, blogs, videos, tools, roadmap OsherElhadad — Awesome Context Engineering — curated resources, papers, tools, best practices Zilliz — Top 10 Context Engineering Techniques for Production RAG LogRocket — The LLM Context Problem in 2026 — memory, relevance and scale strategies
Self-Evolving Ontology	Knowledge structures that organise and update themselves autonomously EvoOntology — Ontology Layer for LLM Data Agents via MCP — builder agent constructs ontology; refinement loop evolves it automatically (Sep 2026) OaK — Toward Effective and Reliable LLM Agents via Dynamic Ontology — ontology-as-a-kernel; makes domain concepts explicit and iteratively refined via judge feedback (Aug 2026) DEEP-JLU — Awesome Graph Engineering — survey covering ontology, graph, loop and context engineering together zjukg — KG-LLM Papers — paper list integrating knowledge graphs with LLMs OntoChat — Collaborative LLM-based Ontology Engineering — practical tool with peer-reviewed paper ontograph — LLM-powered Knowledge Graph Engine — ontological entity resolution and hybrid semantic-graph retrieval LLMs4KGOE 2026 Workshop — ESWC 2026 workshop on LLM-driven knowledge graph and ontology engineering
Intent Engineering	Designing systems that correctly identify what the user actually wants Vincent Koc — 2026: The Year of Intent Engineering — from writing prompts to specifying goals, constraints and feedback loops GopenAI — Intent Routing for AI Agents — intent computed once, stored in agent state, not left to the LLM to decide; includes code bswen — AI Agent Routing: Intent Classification Guide — hands-on Python / LangGraph examples FutureAGI — Intent Classification LLM Pipeline 2026 LiteLLM — open-source router across 200+ LLMs; supports confidence-based intent routing
Decision Engineering	Governing what an agent decides to do — when to act, ask, or stop Anthropic — Building Effective Agents — when to use tools vs. ask vs. stop; definitive reference AI Builder Club — Loop Engineering Guide 2026 — designing the full agent control cycle cobusgreyling — loop-engineering — patterns, starters and CLI tools for agent loop design The Hitchhiker's Guide to Agentic AI — reason → act → observe → stop condition mental model FourWeekMBA — Agent Reliability Gap 2026 — why human-in-the-loop remains the dominant safe pattern
AI Safety & Interpretability	Understanding and aligning what happens inside the model Anthropic — Transformer Circuits — primary research publication for mechanistic interpretability; reverse-engineering how models work internally Anthropic — Core Views on AI Safety — constitutional AI, RLHF/RLAIF, scalable oversight OpenAI Safety — alignment research, red-teaming and policy work Alignment Forum — primary community for technical AI alignment research and discussion arXiv — Constitutional AI: Harmlessness from AI Feedback — Anthropic's method for training models to follow principles Simon Willison — The Lethal Trifecta for AI Agents — the dangerous combination of private data access, untrusted content exposure, and external communication that enables prompt injection attacks and data exfiltration
LLM Evaluation	Testing, benchmarking and measuring LLM output quality PromptFoo — open-source LLM testing and red-teaming; compare prompts, models and providers DeepEval — unit-testing framework for LLM outputs; metrics for RAG, hallucination, toxicity Braintrust — eval platform with dataset management, scoring and CI integration LangSmith — LangChain's eval and observability platform; trace, debug and test LLM chains OpenAI Evals — open-source framework for evaluating OpenAI models; large library of community evals Hugging Face Open LLM Leaderboard — live benchmark comparison of open-source models
Frameworks & Tools	Libraries for building LLM-powered applications and agents LangGraph — hottest in production; stateful, graph-based agent workflow orchestration LlamaIndex — go-to for RAG and document-heavy agents CrewAI — easiest way to build role-based multi-agent teams LangChain examples Pydantic AI — type-safe, structured LLM outputs in Python smolagents (Hugging Face) — minimal agent library, popular for simple agent tasks Microsoft Agent Framework — successor to AutoGen + Semantic Kernel merged; production-grade for Azure-centric teams (AutoGen now in maintenance) Hugging Face Transformers — standard library for running open-source models
Agents & Agentic Systems	Multi-agent frameworks, protocols and skill resources MCP — Model Context Protocol (Anthropic) — open standard for connecting LLMs to tools, data sources and APIs; rapidly becoming the universal standard OpenAI Agents SDK — lightweight framework for OpenAI-based agents (2026) Google ADK — Agent Development Kit — Google's agent framework, gaining momentum Anthropic Agent SDK — computer use as a first-class primitive; launched April 2026 Letta — agents with persistent, stateful memory across sessions Google Agent Skill GitHub repository AgentSkills website
Claude Code	Anthropic's agentic CLI and SDK for software engineering Mastering Claude Code in 30 minutes Every Claude Code Concept Explained Claude Code GitHub repository Claude Code Documentation
Visualisations & Playgrounds	Interactive demos, local runtimes and model explorers TensorFlow — 10,000 Word Vector Visualisation Hugging Face Spaces — thousands of live demos of open-source models LM Studio — run any open-source LLM locally on your Mac, completely offline Ollama — run Llama, Mistral, DeepSeek locally via terminal; extremely popular in 2026 OpenRouter — one API to access and compare 200+ models Anthropic Console Workbench — test and iterate prompts against Claude models