AnyLearn
All lessons
AIintermediate

The LLM Landscape in 2026: Who Makes What

A practitioner's map of the major frontier-model labs as of early 2026 — who they are, what they ship, which models are open vs closed, and how the field has split into general-purpose chat and reasoning families.

llmopenaianthropicgooglemetamistraldeepseekqwenopen-weights
Part of:🧭LLM Models in Practice
Not signed in — your progress and quiz score won't be saved.
Lesson progress1 / 10

Why the map matters

Three years ago you could pick a model by typing one name. In 2026 there are roughly a dozen labs shipping serious frontier or near-frontier systems, split across continents, business models, and licensing regimes. Knowing who makes what is now part of the basic engineering literacy: it decides whether your data leaves the building, what you pay per million tokens, whether you can fine-tune, and which capability gaps you'll hit first.

This lesson is a map, not a ranking. Benchmarks shift every quarter; the structural facts — who runs the lab, where the weights live, how it's licensed — move much more slowly and are what you actually plan against.

Full lesson text

All 10 steps on one page — for reading, reference, and search.

Show

1. Why the map matters

Three years ago you could pick a model by typing one name. In 2026 there are roughly a dozen labs shipping serious frontier or near-frontier systems, split across continents, business models, and licensing regimes. Knowing who makes what is now part of the basic engineering literacy: it decides whether your data leaves the building, what you pay per million tokens, whether you can fine-tune, and which capability gaps you'll hit first.

This lesson is a map, not a ranking. Benchmarks shift every quarter; the structural facts — who runs the lab, where the weights live, how it's licensed — move much more slowly and are what you actually plan against.

2. OpenAI: GPT-5 and the o-series

OpenAI ships two parallel families. The GPT line (GPT-5, GPT-4.1, GPT-4o for cheaper multimodal) is tuned for general chat, agents, and tool use. The o-series (o1, o3, and successors) is reasoning-first: the model spends hidden "thinking" tokens before answering, trading latency and cost for scores in the high range on AIME and GPQA Diamond.

Both are closed-weights, US-based, API-only through OpenAI or Microsoft Azure. No fine-tuning of the top reasoning models for end customers; supervised fine-tuning is exposed for a subset of GPT-class checkpoints. Expect aggressive prompt-caching, batch-API discounts, and a Responses API that consolidates chat, tools, and streaming.

3. Anthropic: Claude 4.x

Anthropic ships the Claude 4 family — Opus, Sonnet, Haiku, in roughly descending size and cost. Opus 4.x is the flagship reasoning + coding model, Sonnet is the daily-driver workhorse, Haiku targets latency-sensitive and cheap-volume use. Claude is known for strong performance on SWE-bench Verified and long-form writing, plus a serious safety/constitutional-AI bent.

Closed-weights, US-based, available via Anthropic's API, AWS Bedrock, and Google Cloud Vertex. Prompt caching is first-class (you mark which prefix to cache). Context windows reach into the hundreds of thousands of tokens; a 1M context variant is generally available for select models.

4. Google DeepMind: Gemini 2.x

Google DeepMind ships Gemini 2.x in Pro, Flash, and Nano tiers. Gemini's calling card is native multimodality (text, image, audio, video all in one model) and very long context — multi-million-token windows for Pro variants. Strong on multilingual and on video understanding specifically.

Closed-weights, accessed via Google AI Studio, Vertex AI, or the Gemini app. Google also publishes the Gemma open-weights family — smaller models (a few B to ~30B parameters) released under a custom permissive license, intended for on-device and self-hosted use. Gemma is not the same as Gemini; treat them as separate product lines that share research lineage.

5. Open-weights: Meta, Mistral, DeepSeek, Qwen

The open-weights frontier in 2026 is dominated by four players. Meta's Llama (3.x, 4) ships in sizes from ~8B to several-hundred-B parameters under a community license that's permissive for most commercial use. Mistral (French) publishes both open-weights (Mistral Small, Mixtral MoE) and closed (Mistral Large) variants. DeepSeek (Chinese) shocked the field with V3 (general) and R1 (reasoning) — MIT-licensed weights that close most of the gap to closed frontier on math and code at a fraction of training cost. Alibaba's Qwen 2.5 ships in many sizes and languages, with strong code and multilingual variants under Apache-2.0 or similar.

These are the models you can actually download, run on your own hardware, fine-tune, and ship in air-gapped products.

6. The 2026 lab map

Labs by region and weights model.

flowchart TD
A["Frontier LLM Labs (2026)"]
A --> B["US closed-weights"]
A --> C["US open-weights"]
A --> D["European"]
A --> E["Chinese open-weights"]
B --> B1["OpenAI: GPT-5, o3"]
B --> B2["Anthropic: Claude 4.x"]
B --> B3["Google DeepMind: Gemini 2.x"]
C --> C1["Meta: Llama 3.x / 4"]
C --> C2["Google: Gemma"]
D --> D1["Mistral: Large (closed), Small (open)"]
E --> E1["DeepSeek: V3, R1"]
E --> E2["Alibaba: Qwen 2.5"]
E --> E3["Moonshot / Zhipu / others"]

7. Chat models vs reasoning models

Around 2024 the field split into two model archetypes that now exist side by side in every major lab:

  • General-purpose chat models (GPT-5, Claude Sonnet, Gemini 2.x Pro, Llama 4) — answer immediately, optimized for latency, tool use, multimodality, and broad helpfulness.
  • Reasoning models (o1, o3, Claude Opus with extended thinking, Gemini 2.x Thinking, DeepSeek R1) — emit hidden chain-of-thought before the visible answer. They dominate on AIME, GPQA Diamond, ARC-AGI, and SWE-bench, but cost 3-30x more per task and add seconds-to-minutes of latency.

For most product workloads you still want a chat model. Reasoning models earn their keep on hard math, complex code, scientific QA, and multi-step agent planning.

8. Specialty and second-tier players

The map has a long tail worth knowing:

  • xAI's Grok — closed-weights, integrated with X, competitive on some general benchmarks.
  • Cohere — enterprise-focused (Command R+), strong retrieval-augmented generation tooling.
  • AI21, Databricks (DBRX), Snowflake (Arctic) — enterprise stacks bundling models with data platforms.
  • Apple Intelligence — small on-device models plus opt-in cloud (Private Cloud Compute) to OpenAI/Anthropic.
  • Microsoft Phi — small (<10B), open-weights, surprisingly strong reasoning-per-parameter.

None of these are the cheapest API or the top of MMLU, but each owns a specific niche — enterprise RAG, on-device, regulated industries — where the frontier labs aren't a fit.

9. Regulatory and regional reality

Where the lab is matters for procurement and compliance, not just nationalism. EU customers under the AI Act distinguish general-purpose AI models with systemic risk (the frontier closed models) from smaller models with lighter obligations. US providers default to US data residency unless you pay for EU regions; Chinese-trained weights like DeepSeek and Qwen are MIT/Apache licensed but raise procurement questions in some regulated sectors and on US federal contracts.

A practical rule of thumb: closed US APIs for raw capability, EU-hosted variants (Mistral, Azure EU, Bedrock EU) for European data-residency requirements, and open weights when you need to run inside a customer's VPC or on-prem.

10. How to use this map

When you start a new LLM project in 2026, walk this list once before you write a line of code:

  1. Does my data have to stay on-prem or in a specific region? If yes, open weights or region-specific API only.
  2. Is the task latency-sensitive (chat UX, voice)? Skip reasoning models.
  3. Is it a hard reasoning / code / math task? Start with a reasoning model and downgrade only if the eval allows.
  4. Do I need multimodal in/out? Gemini 2.x and GPT-4o-class tend to lead; check the exact modality.
  5. What's my budget per million tokens at production scale? Open weights on your own GPUs win past a certain volume; closed APIs win below it.

The lab map decides options 1, 3, and 4 before you ever look at a benchmark.

Check your understanding

The lesson ends with a 5-question quiz. Take it in the player above to see your score.

  1. Which of these model families is released with open weights?
    • OpenAI GPT-5
    • Anthropic Claude Opus 4
    • Meta Llama 4
    • Google Gemini 2.x Pro
  2. What is the main practical tradeoff of a reasoning model like o3 or DeepSeek-R1 vs a general chat model?
    • It uses less memory at inference time
    • It returns answers faster
    • It costs more and adds latency in exchange for better hard-reasoning scores
    • It always has a larger context window
  3. DeepSeek and Qwen are notable in the 2026 landscape primarily because they:
    • Are the only models with native video understanding
    • Ship competitive open-weights models out of Chinese labs under permissive licenses
    • Are the cheapest closed APIs available globally
    • Are exclusively distributed through AWS Bedrock
  4. Gemma is best described as:
    • Google's flagship closed multimodal model
    • An open-weights family from Google, separate from but related to Gemini
    • A fine-tuned variant of Llama 3 maintained by Google
    • Google's reasoning-specialized model
  5. You need an LLM that can run inside a customer's on-prem VPC with no outbound network. Which family is the most natural fit?
    • OpenAI o3 via the Responses API
    • Anthropic Claude Opus 4 via Bedrock
    • Meta Llama 4 or DeepSeek V3 self-hosted
    • Google Gemini 2.x Pro via Vertex AI

Related lessons

AI
advanced

RL in Reasoning Models: How o1, DeepSeek-R1, and Friends Think

A deep look at how reinforcement learning on chains-of-thought powers o1, DeepSeek-R1, Claude reasoning, and Gemini Thinking — covering GRPO, MCTS-style search, test-time compute scaling, and distillation into smaller models.

13 steps·~20 min·audio
AI
advanced

Loop engineering: verification, orchestration, and anti-patterns

Make loops trustworthy. Adversarial verification panels, sub-agent orchestration, loop-until-dry for unbounded discovery, loop-until-budget for paid depth, multi-modal sweeps with diverse prompts, eval-driven cap selection, and the five anti-patterns — silent caps, infinite plans, drift, premature termination, thrashing — that ship to prod more than they should.

9 steps·~14 min
AI
intermediate

Loop engineering in production: state, errors, and budgets

Once an agent loop runs past a few iterations the failure surface shifts: state bloats, tool calls fail in three distinct ways, retries multiply if you stack them wrong, and unbounded budgets rack up four-figure bills overnight. The state, errors, and budgets you need to make the loop survive contact with reality.

8 steps·~12 min
AI
beginner

Loop engineering basics: the agent control loop

How LLM agents actually run: the iterative prompt-action-observation loop, the ReAct shape, the smallest tool-calling loop in twelve lines, why you always stack three termination layers, what the model sees on iteration N, and when a single prompt is the better answer.

8 steps·~12 min