Reference notes.

Info

Last reviewed: November 2026. Model versions, context windows, and pricing change rapidly — check provider docs for current specs.

Foundation models are large AI models trained on broad data that can be adapted to many downstream tasks. The term encompasses LLMs, vision models, and multimodal systems.

Commercial Models

Anthropic Claude

ModelContextStrengths
Claude Fable 5.21MFlagship for open-ended technical reasoning, coding, and scientific research (Nov 2026)
Claude Mythos 5.21MSame intelligence as Fable 5.2 with dual-use research access; gated behind verification programmes (Nov 2026)
Claude Opus 5.51MDaily-driver flagship for heavy coding and long-running agent tasks
Claude Sonnet 5.51MBest balance of speed, judgement, and capability
Claude Haiku 5.51MFastest, cost-effective tier
  • Adaptive and extended thinking for complex reasoning
  • Up to 128K max output tokens
  • Strong instruction following and tool use
  • Constitutional AI training
  • API Documentation

OpenAI GPT

ModelContextStrengths
GPT-6.2 Astra~1.05MFlagship for complex reasoning, agentic workflows, scientific research, and coding (Nov 2026)
GPT-6.2 Sol~1.05MHigh-volume, cost-efficient model for everyday coding, reviews, and professional workflows (Nov 2026)
  • Unified architecture with adjustable reasoning effort, superseding standalone o3 and GPT-5.x models
  • Up to 128K max output tokens
  • Strong function calling, coding, and tool use
  • API Documentation

Google Gemini

ModelContextStrengths
Gemini 4 Carbon1M+Deepest reasoning tier for long-horizon planning, complex software engineering, and science
Gemini 4 Neon1M+Mid-tier balance of deep reasoning and token efficiency (Nov 2026)
Gemini 4 Argon1M+General-purpose all-rounder for research and complex knowledge work
Gemini 4 Hydrogen1M+High-speed, cost-efficient model for everyday execution and high-volume agents
  • Native multimodal (text, images, audio, video)
  • Configurable deep reasoning across tiers
  • Strong agentic, coding, and scientific performance
  • API Documentation

Others

  • xAI Grok 4.7 — 500K context, adjustable reasoning effort, strong coding and long-horizon agentic execution
  • Meta Muse — Meta’s proprietary frontier family (Wonder, Spark), succeeding the open Llama line for frontier tasks (accompanied by the open-weight 30B Glimmer variant)
  • Qwen4-Max (Alibaba, Nov 2026) — Proprietary API-only flagship (2.4T total / 71B active MoE, 1M context, native multimodal); see Qwen 3.8 below for open-weight releases
  • Cohere Command A — 111B parameters, 256K context, enterprise focus and RAG-optimised
  • Amazon Nova 2 — AWS Bedrock native multimodal family (Lite, Pro, Omni, Sonic), up to 1M context

Open-Source Models

Meta Llama

ModelParametersContextNotes
Llama 4 Scout109B total / 17B active (16 routed + 1 shared)10MLargest context window available, iRoPE (April 2025)
Llama 4 Maverick400B total / 17B active (128 routed + 1 shared)1MLarger expert pool (April 2025)
Llama 3.3 70B70B dense128KText-only instruct
  • Mixture-of-experts architecture (Llama 4); planned Llama 4 Behemoth was shelved as Meta shifted frontier development to its proprietary Muse family (Wonder, Spark, with open-weight 30B Glimmer)
  • Native multimodal (text + images)
  • Llama 4 Scout uses Interleaved RoPE (iRoPE) — trained at 256K, extrapolates to 10M
  • Gated open licence (Llama Community Licence — not OSI open-source)
  • Llama Downloads

Mistral

ModelParametersContextNotes
Mistral Large 4~1.05T total / ~52B active (MoE)1MFlagship open-weight sparse MoE, native vision, adjustable reasoning
Mistral Small 4119B total / 6.5B active (MoE)256KCompact hybrid reasoning and instruction MoE
Ministral 3 (3B/8B/14B)Dense128KSmall models with vision baked in (base, instruct, reasoning)
Mistral OCR 4.1——Document digitisation with layout structure and confidence scores
  • European AI company
  • Strong efficiency/performance ratio
  • Multilingual with native vision support
  • Le Chat consumer product
  • Mistral AI

Qwen (Alibaba)

ModelParametersContextNotes
Qwen 3.8 (27B / Omni-Flash)27B dense & multimodal Flash256K+Latest open-weight Qwen generation
Qwen3-235B-A22B235B total / 22B active (MoE)128KWidely deployed open MoE
Qwen3-VLVarious128K+Vision-language family
QwQ-32B32B128KEarlier reasoning baseline
  • Flagship Qwen4-Max (Nov 2026; 2.4T total / 71B active MoE, 1M context, native multimodal) is API-only; Qwen 3.8 and earlier generations remain open-weight under Apache 2.0
  • Strong multilingual (100+ languages)
  • Extensive model family (code, maths, embedding, reranking variants)
  • Qwen3-Embedding series leads open-weight multilingual retrieval on MTEB
  • Qwen

DeepSeek

ModelParametersNotes
DeepSeek-V4-Pro~1.6T MoE (49B active)Native 1M context, integrated reasoning
DeepSeek-V4-Flash / V4.1-Flash284B (13B active) / 552B multimodal MoEFast, cost-efficient variants with native 1M context
DeepSeek-V3.2 / R1671B MoE (37B active)Earlier open-weight baselines with distilled variants
  • Competitive with frontier models at a fraction of training and inference cost
  • Cost-efficient MoE architecture with multi-head latent attention and integrated reasoning
  • Open weights under MIT licence — free for commercial use
  • DeepSeek

Others

  • Gemma 4 (Google, April 2026) — Open weights under Apache 2.0; spans E2B, E4B, 12B (multimodal including audio), 26B MoE, and 31B dense
  • OLMo 3 / 3.1 (AI2) — Fully open (weights, training code, checkpoints, and Dolma 3 data) in 7B and 32B (Base, Instruct, Think) under Apache 2.0
  • Phi-4 (Microsoft) — Small dense model with strong reasoning
  • GPT-OSS 20B / 120B (OpenAI, Aug 2025) — OpenAI’s open-weight reasoning release under Apache 2.0; gpt-oss-120b targets a single 80 GB data-centre GPU (e.g. H100), whereas gpt-oss-20b is the consumer/on-device variant

Model Comparison Factors

Capability Benchmarks

See Evaluation & Benchmarking for details.

  • Humanity’s Last Exam (HLE) — Frontier cross-disciplinary reasoning and expert knowledge
  • GPQA Diamond — Graduate-level science
  • SWE-bench Pro — Complex real-world software engineering
  • ARC-AGI-2 / ARC-AGI-3 — Novel abstract reasoning and generalisation
  • Arena — Crowdsourced human preference rankings (replacing saturated legacy benchmarks like MMLU, HumanEval, and GSM8K)

Practical Considerations

FactorConsiderations
LatencyTime to first token, tokens/second
CostPer-token pricing, volume discounts
ContextHow much text can be processed
ReliabilityUptime, consistency
PrivacyData handling, compliance
EcosystemSDKs, documentation, support

Licence Types

  • Proprietary API — No access to weights (GPT-6.2, Claude 5.5 / Fable 5.2, Gemini 4, Qwen4-Max, Grok 4.7)
  • Gated open — Weights available with restrictions (Llama 4)
  • Permissive open — Few restrictions (Mistral Large 4 / Small 4, Qwen 3.8, DeepSeek-V4, Gemma 4, GPT-OSS)
  • Fully open — Weights, code, and training data (OLMo 3 / 3.1)

API Providers

Model Providers

Direct from the source:

Aggregators / Routers

Access multiple models through one API:

Cloud Platforms

See Model Serving for self-hosted inference and deployment options.

Choosing a Model

Decision Framework

  1. Task requirements — What capability is most important?
  2. Latency needs — Real-time vs batch processing
  3. Cost constraints — Budget per million tokens
  4. Privacy requirements — Can data leave your environment?
  5. Context needs — How much text per request?
  6. Compliance — Regulatory requirements

Rules of Thumb

  • Start with a capable model (Claude Sonnet 5.5 / Opus 5.5, GPT-6.2 Astra / Sol, Gemini 4 Argon / Neon)
  • Optimise for cost/speed once it works (Haiku 5.5, GPT-6.2 Sol, Gemini 4 Hydrogen, DeepSeek-V4-Flash)
  • Open models for privacy-sensitive use cases (DeepSeek-V4, Mistral Large 4 / Small 4, Qwen 3.8, Gemma 4, Llama 4)
  • Frontier API pricing continues to drop rapidly per unit of capability — re-evaluate cost assumptions regularly
  • Smaller models for high-volume, simple tasks

Staying Current

The landscape changes rapidly. Track developments:

Resources