The Model Landscape

This content was generated with the assistance of AI and should be verified against the official documentation of each tool and framework before being relied on in production.

This section’s bibliography lists the reference material consulted while preparing these pages.

This page surveys the model families in active use as of 2026-09-28, checked against Anthropic’s models overview, OpenAI’s models page, Google’s Gemini models page and the Hugging Face Hub. Examples target Python 3.12 with the current anthropic and openai Python SDKs, and Ollama 0.34’s local REST API. Names, tiers and links change often — this page deliberately omits prices and benchmark scores, which move even faster than model names do and are covered on Cost, Latency & Throughput and Benchmarks & Leaderboards instead.

Proprietary families

A proprietary (closed-weight) model is served only through its vendor’s own API or cloud marketplace listing — the weights themselves are never published, so the only way to run one is to call it. The three families a production stack is most likely to reach for:

  • Anthropic Claude — the current lineup (Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5, Claude Haiku 4.5) spans a slow, deep-reasoning tier down to a fast, near-frontier one; every current model accepts text and image input with tool use, and claude-sonnet-5 is the model these docs write their own examples against. See Anthropic — Models overview.

  • OpenAI GPT — the current flagship generation (GPT-6, in Astra / Sol / Luna variants) sits alongside the still-current GPT-5.6 family and specialised models for cybersecurity, life sciences, realtime voice, image generation and transcription. See OpenAI — Models.

  • Google Gemini — the current Gemini 3 generation is organised the same way as its predecessors: a Pro tier for the deepest reasoning and agentic work, a Flash tier for general-purpose speed, and a Flash-Lite tier for the highest-throughput, lowest-cost workloads, alongside the still-listed Gemini 2.5 generation. See Google — Gemini models.

Open-weight families

An open-weight model publishes its trained parameters (usually on Hugging Face) so anyone can download, fine-tune, quantise and self-host it, even though the training data and code that produced it are typically not released alongside it. Families a production stack commonly encounters:

  • Llama (Meta) — general-purpose chat and instruction-tuned models across a range of parameter counts.

  • Qwen (Alibaba) — a broad family spanning dense and mixture-of-experts variants, with strong coding and multilingual variants.

  • Gemma (Google) — Google’s open-weight counterpart to Gemini, aimed at running well on a single GPU or on-device.

  • Mistral (Mistral AI) — dense and mixture-of-experts models, including small models tuned for local and edge deployment.

  • DeepSeek (DeepSeek) — large mixture-of-experts models with strong coding and reasoning results relative to their active-parameter count.

  • gpt-oss (OpenAI) — OpenAI’s own open-weight release, distinct from its closed-weight GPT-6/GPT-5.6 API models above.

  • Phi (Microsoft) — small models trained on curated, textbook-quality data, aimed squarely at the on-device and small-tier end of the spectrum below.

Every one of these families is versioned and re-released frequently; the authoritative source for exact current checkpoints, licenses and parameter counts is each family’s own listing on the Hugging Face Hub, not a snapshot on this page.

Tiers

Vendors group their own models into tiers along the same rough axis — more reasoning depth and cost versus more speed and lower cost — even though the tier names differ:

  • Frontier — the vendor’s most capable model at the time, tuned for the hardest reasoning, coding and long-horizon agentic tasks, at the highest per-token cost and latency (Claude Opus/Fable, GPT-6, Gemini Pro).

  • Fast — a "best combination of speed and intelligence" tier meant as the default choice for most production workloads (Claude Sonnet, Gemini Flash).

  • Small — a cheaper, lower-latency tier for high-volume or latency-sensitive calls that don’t need frontier reasoning (Claude Haiku, Gemini Flash-Lite, the smaller members of the open-weight families above).

  • On-device — models small enough to run locally on a laptop, phone or edge device without a hosted API call at all — the smallest Llama/Gemma/Phi/Mistral checkpoints, run through Ollama or llama.cpp, or through a device’s own on-device stack such as Apple’s Foundation Models framework.

Modalities

"Model" no longer means "text in, text out" by default. Most current frontier and fast-tier models accept text plus image input and produce text output with tool use, as covered on Multimodal Models; some vendors additionally offer distinct model families for audio/realtime voice, image generation, video generation and embeddings (for example OpenAI’s realtime and image-generation lines above), which are called out separately in each vendor’s own models listing rather than folded into the same table.

A dated snapshot (as of 2026-09-28)

Family Vendor Open / closed Typical tiers Official models page

Claude

Anthropic

Closed

Frontier (Fable, Opus) · Fast (Sonnet) · Small (Haiku)

Models overview

GPT

OpenAI

Closed

Frontier (GPT-6) · Fast/specialised (GPT-5.6 family)

Models

Gemini

Google

Closed

Frontier (Pro) · Fast (Flash) · Small (Flash-Lite)

Gemini models

Llama

Meta

Open

Fast · Small · On-device

Meta Llama on Hugging Face

Qwen

Alibaba

Open

Frontier · Fast · Small · On-device

Qwen on Hugging Face

Gemma

Google

Open

Small · On-device

Google on Hugging Face

Mistral

Mistral AI

Open

Fast · Small · On-device

Mistral AI on Hugging Face

DeepSeek

DeepSeek

Open

Frontier · Fast

DeepSeek AI on Hugging Face

gpt-oss

OpenAI

Open

Fast · Small

OpenAI on Hugging Face

Phi

Microsoft

Open

Small · On-device

Microsoft on Hugging Face

No price or benchmark-score column is included above on purpose: both move faster than this table could stay accurate, and both are covered in depth, with their own caveats, on Cost, Latency & Throughput and Benchmarks & Leaderboards.

For hosted catalogs that resell several of these families behind one cloud subscription instead of a direct vendor API key, see the Azure AI, data & IoT overview — Azure AI Foundry’s model catalog is one example of a single hosted surface spanning both proprietary and open-weight families.

Example: listing the models an SDK or install can actually see

A models table on a page is a snapshot; the models an account, API key or local install can actually use right now are best read back programmatically. All three ecosystems below expose that as a simple list call.

Anthropic: client.models.list()

import anthropic

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY from the environment

for model in client.models.list():
    print(model.id, "--", model.display_name)

The response is paginated (has_more, first_id, last_id) and each entry also carries created_at, max_input_tokens, max_tokens and a capabilities object (whether the model supports batching, thinking, structured outputs, image/PDF input, and which effort levels it accepts) — useful for picking a model by capability instead of hard-coding an ID. See Anthropic — List models.

OpenAI: client.models.list()

from openai import OpenAI

client = OpenAI()  # reads OPENAI_API_KEY from the environment

for model in client.models.list():
    print(model.id, "--", model.owned_by)

Each entry reports the model’s id, its Unix created timestamp and the owned_by organisation. See OpenAI — Models.

Ollama: ollama list and GET /api/tags

Listing what is available locally is a different question from listing what a hosted vendor offers — Ollama answers it from the models already pulled onto the machine, with no network call to a vendor at all. From the CLI:

ollama list

Or the same information from Ollama’s own local REST API, which any HTTP client (or the OpenAI-compatible client used elsewhere on this site) can call directly:

curl http://localhost:11434/api/tags

The response is a JSON object with a models array, each entry giving the model’s name, size, digest and last-modified time. See Ollama — List local models.

Related pages

References