The Map: Who Makes What (2026)

A plain-language guide to the major AI labs - OpenAI, Anthropic, Google DeepMind, Meta, Mistral, xAI, and DeepSeek - what they actually ship, and how to think about the landscape without the hype.

TL;DR: Seven labs dominate AI in 2026 - OpenAI, Anthropic, Google DeepMind, Meta, Mistral, xAI, and DeepSeek. Each has a distinct strategy. Some keep their weights locked. Others give them away. The gap between "open" and "closed" is still real but narrower than it was a year ago. Here is what each lab ships and why it matters if you are building something.

Why the map matters

If you are building a product, a workflow, or even just trying to understand what is happening in the world, you need a working model of who makes what. The AI hype cycle has made this harder, not easier - every week brings announcements that feel world-ending but often just rename something that already existed.

This guide cuts through that. It covers the seven labs whose models actually end up in production in 2026, what they ship, and the one or two things that make each one distinct. No benchmarks for their own sake. No scoring debates. Just the map.

The first split: closed vs. open weights

Before the lab-by-lab breakdown, one concept does all the work: the distinction between closed-weight and open-weight models.

The practical implication: closed models tend to be more capable at the frontier right now, but open models give you control over latency, cost, and data privacy. Most serious builders end up using both - a closed API for tasks that need maximum quality, and an open-weight model fine-tuned on their domain for everything else.

The seven labs

OpenAI

OpenAI is the lab that made LLMs a household conversation. Their current flagship is GPT-5.5, released April 23, 2026 - it excels at agentic coding, knowledge work, and computer use, and uses significantly fewer tokens to finish the same tasks compared to earlier generations. Before that, o3 set the standard for reasoning-heavy work like math and competitive coding. The model lineup also includes GPT-4.1 (strong on code and instruction following) and a growing set of "mini" variants for cost-sensitive use cases.

The thing that sets OpenAI apart is depth of tooling. The API ships with file search, code execution, and computer use as native capabilities - not bolt-ons. If you want to build an agent that operates software the way a human would, OpenAI currently has the most polished path to doing that.

Anthropic

Anthropic was founded by former OpenAI researchers with an explicit focus on AI safety. Their model lineup spans four classes. The everyday tiers are Haiku (fast, economical), Sonnet (workhorse for coding and agentic work), and Opus (complex reasoning, long-running tasks) - with Claude Opus 4.8 (May 28, 2026) being the current Opus release, notable for honesty and reliability improvements including being far less likely to let code flaws pass without flagging them. Above all of these sits the new Mythos-class tier: Claude Fable 5 (June 9, 2026) is Anthropic's most capable widely available model, with adaptive thinking always on and a 1-million-token context window.

Anthropic's specific edge is instruction-following precision and long-context reliability. For tasks where the model needs to hold a lot of information in mind without losing the thread - long documents, big codebases, complex research - Claude models consistently perform well. Anthropic also introduced the Model Context Protocol, an open standard for connecting models to tools and data sources that is now governed by the Agentic AI Foundation (a Linux Foundation directed fund) and has seen broad industry adoption.

Google DeepMind

Google merged Google Brain and DeepMind into one unit in 2023. The result is a lab with both the research depth of DeepMind and the infrastructure of Google. Their model family is Gemini. The current lineup in mid-2026 includes Gemini 3.1 Pro (February 2026, strong on multimodal reasoning and 1M context), Gemini 3.5 Flash (announced at Google I/O, May 2026, faster with better benchmark performance than 3.1 Pro on nearly all tested dimensions), and Gemini Omni (also announced at Google I/O 2026), a model that takes text, audio, images, and video as input and generates video as output.

The distinguishing factor for Google is native multimodality at scale and integration into Google's product surface. Gemini models run inside Search, Android, Workspace, and Vertex AI. If your product touches Google's ecosystem - or you need serious video understanding and generation - DeepMind's work is the most relevant.

Meta

Meta's bet is simple and has not changed: give the weights away and win the ecosystem. Their Llama 4 family (released April 5, 2025, still current as of mid-2026) uses a Mixture-of-Experts architecture. Llama 4 Maverick runs 17 billion active parameters across 128 experts (400 billion total parameters). Llama 4 Scout is the smaller variant with an industry-leading 10-million-token context window. Both are downloadable from Hugging Face and the Llama site.

Meta ships into WhatsApp, Instagram, and Messenger as Meta AI. But for builders, the value is different: Llama 4 is the open-weight baseline that a huge number of fine-tunes and local deployments run on. If you are fine-tuning a model on your own data and want a permissive license with a strong starting point, Llama is the most common choice.

Mistral

Mistral is a French lab that punches above its headcount by focusing on efficiency. Their flagship open-weight model is Mistral Large 3 (released December 2025) - a sparse Mixture-of-Experts model with 41 billion active parameters out of 675 billion total. In 2026 they have expanded into audio (Voxtral TTS, released March 26, 2026 - their first text-to-speech model, supporting zero-shot voice cloning in 9 languages) and document understanding (Mistral OCR 4, released June 23, 2026, supporting 170 languages and running fully self-hosted).

Mistral's enterprise play is Forge, announced in March 2026 at NVIDIA GTC - a platform for training frontier-grade models on a company's own proprietary data. If data sovereignty matters (you are in Europe, or in a regulated industry), Mistral is the lab most explicitly built around self-hosted and on-premise deployment.

xAI

Elon Musk's lab. Their model is Grok, now at version 4.3 (released April 30, 2026 on the public API), with a 1-million-token context window and native video input. Grok has one structural advantage no other model has: real-time integration with X (formerly Twitter), which means access to live public conversation at a scale that other models cannot match through web search alone.

xAI also released Grok Build 0.1 (June 2026), a coding-focused model in public beta via the API, and an image-to-video model preview. The lab is moving fast and the roadmap includes Grok 5, reportedly training at a 10-trillion-parameter scale. For builders who need real-time social signal or who are building on the X platform, Grok is the natural starting point.

DeepSeek

DeepSeek is the Chinese lab that rattled the entire industry in early 2025 by releasing models that matched frontier performance at a fraction of the reported training cost. Their current release is DeepSeek V4 (preview released April 24, 2026), built on a new hybrid attention architecture - Compressed Sparse Attention interleaved with Heavily Compressed Attention - and natively supporting a 1-million-token context window. The reasoning-focused variant, R1, remains widely used for math, code, and scientific tasks, though V4's built-in thinking mode increasingly covers the same ground.

All of DeepSeek's research models are released as open weights under the MIT license. That means the architectural innovations - Mixture-of-Experts, the hybrid attention approach introduced in V4, efficient training pipelines - are publicly available for anyone to study and adapt. DeepSeek is the lab most worth following if you care about the underlying technical frontier rather than the product surface.

How to orient without getting lost

With seven labs and dozens of models, the instinct is to track every release. That instinct will exhaust you. A more useful frame:

  1. Pick your use case first, then the model. Closed frontier APIs (GPT-5.5, Claude Fable 5, Gemini 3.5 Flash) are strongest for complex reasoning, long context, and agentic tasks. Open-weight models (Llama 4, Mistral Large 3, DeepSeek V4) are strongest when you need control, privacy, or domain-specific fine-tuning.
  2. The gap between open and closed is real but narrowing. Closed models still lead at the hard end of reasoning. Open models are roughly 6 months behind the frontier - and closing. For many production tasks, an open model is already good enough.
  3. Multimodal is the direction everything is moving. Text-in-text-out is no longer the interesting axis. Every major lab now ships or is building toward audio, image, and video as first-class inputs and outputs.
  4. Infrastructure matters as much as the model. The API you call is only part of the stack. File search, code execution, tool use, memory, and agent orchestration are where the real capability differences show up in practice.
# A simple decision tree for picking a starting point:
#
# Do you need maximum quality on a hard reasoning task?
#   -> Closed API: GPT-5.5 / Claude Fable 5 / Gemini 3.5 Flash
#
# Do you need data privacy or self-hosted deployment?
#   -> Open weight: Llama 4 Maverick / Mistral Large 3 / DeepSeek V4
#
# Do you need real-time social/news data?
#   -> Grok 4.3 (xAI) with X integration
#
# Do you need video understanding or generation?
#   -> Gemini Omni (Google DeepMind)
#
# Do you need to fine-tune on your own domain data?
#   -> Start with Llama 4 or DeepSeek V4 as the base

Key takeaways

Try this next: once you know who makes what, the natural question is how to actually call these models - prompting patterns, API basics, and when to reach for which tier. See The Prompting Mental Model for the practical next step.

LearnlandscapeThe Map: Who Makes What (2026)
Guidelandscapeintro7 min read

The Map: Who Makes What (2026)

A plain-language guide to the major AI labs - OpenAI, Anthropic, Google DeepMind, Meta, Mistral, xAI, and DeepSeek - what they actually ship, and how to think about the landscape without the hype.

CREATORS OF TODAY.
@creatorsoftoday
321

TL;DR: Seven labs dominate AI in 2026 - OpenAI, Anthropic, Google DeepMind, Meta, Mistral, xAI, and DeepSeek. Each has a distinct strategy. Some keep their weights locked. Others give them away. The gap between "open" and "closed" is still real but narrower than it was a year ago. Here is what each lab ships and why it matters if you are building something.

Why the map matters

If you are building a product, a workflow, or even just trying to understand what is happening in the world, you need a working model of who makes what. The AI hype cycle has made this harder, not easier - every week brings announcements that feel world-ending but often just rename something that already existed.

This guide cuts through that. It covers the seven labs whose models actually end up in production in 2026, what they ship, and the one or two things that make each one distinct. No benchmarks for their own sake. No scoring debates. Just the map.

The first split: closed vs. open weights

Before the lab-by-lab breakdown, one concept does all the work: the distinction between closed-weight and open-weight models.

  • Closed-weight - you call the model through an API. You never see the underlying parameters. The lab controls pricing, access, and what you can do with it. OpenAI, Anthropic, Google, and xAI all operate this way for their frontier work.
  • Open-weight - the model file is downloadable. You can run it on your own hardware, fine-tune it on your own data, and deploy it without per-token API costs. Meta and Mistral lead here. DeepSeek publishes open weights too, and has become one of the most downloaded families on Hugging Face.

The practical implication: closed models tend to be more capable at the frontier right now, but open models give you control over latency, cost, and data privacy. Most serious builders end up using both - a closed API for tasks that need maximum quality, and an open-weight model fine-tuned on their domain for everything else.

The seven labs

OpenAI

OpenAI is the lab that made LLMs a household conversation. Their current flagship is GPT-5.5, released April 23, 2026 - it excels at agentic coding, knowledge work, and computer use, and uses significantly fewer tokens to finish the same tasks compared to earlier generations. Before that, o3 set the standard for reasoning-heavy work like math and competitive coding. The model lineup also includes GPT-4.1 (strong on code and instruction following) and a growing set of "mini" variants for cost-sensitive use cases.

The thing that sets OpenAI apart is depth of tooling. The API ships with file search, code execution, and computer use as native capabilities - not bolt-ons. If you want to build an agent that operates software the way a human would, OpenAI currently has the most polished path to doing that.

Anthropic

Anthropic was founded by former OpenAI researchers with an explicit focus on AI safety. Their model lineup spans four classes. The everyday tiers are Haiku (fast, economical), Sonnet (workhorse for coding and agentic work), and Opus (complex reasoning, long-running tasks) - with Claude Opus 4.8 (May 28, 2026) being the current Opus release, notable for honesty and reliability improvements including being far less likely to let code flaws pass without flagging them. Above all of these sits the new Mythos-class tier: Claude Fable 5 (June 9, 2026) is Anthropic's most capable widely available model, with adaptive thinking always on and a 1-million-token context window.

Anthropic's specific edge is instruction-following precision and long-context reliability. For tasks where the model needs to hold a lot of information in mind without losing the thread - long documents, big codebases, complex research - Claude models consistently perform well. Anthropic also introduced the Model Context Protocol, an open standard for connecting models to tools and data sources that is now governed by the Agentic AI Foundation (a Linux Foundation directed fund) and has seen broad industry adoption.

Google DeepMind

Google merged Google Brain and DeepMind into one unit in 2023. The result is a lab with both the research depth of DeepMind and the infrastructure of Google. Their model family is Gemini. The current lineup in mid-2026 includes Gemini 3.1 Pro (February 2026, strong on multimodal reasoning and 1M context), Gemini 3.5 Flash (announced at Google I/O, May 2026, faster with better benchmark performance than 3.1 Pro on nearly all tested dimensions), and Gemini Omni (also announced at Google I/O 2026), a model that takes text, audio, images, and video as input and generates video as output.

The distinguishing factor for Google is native multimodality at scale and integration into Google's product surface. Gemini models run inside Search, Android, Workspace, and Vertex AI. If your product touches Google's ecosystem - or you need serious video understanding and generation - DeepMind's work is the most relevant.

Meta

Meta's bet is simple and has not changed: give the weights away and win the ecosystem. Their Llama 4 family (released April 5, 2025, still current as of mid-2026) uses a Mixture-of-Experts architecture. Llama 4 Maverick runs 17 billion active parameters across 128 experts (400 billion total parameters). Llama 4 Scout is the smaller variant with an industry-leading 10-million-token context window. Both are downloadable from Hugging Face and the Llama site.

Meta ships into WhatsApp, Instagram, and Messenger as Meta AI. But for builders, the value is different: Llama 4 is the open-weight baseline that a huge number of fine-tunes and local deployments run on. If you are fine-tuning a model on your own data and want a permissive license with a strong starting point, Llama is the most common choice.

Mistral

Mistral is a French lab that punches above its headcount by focusing on efficiency. Their flagship open-weight model is Mistral Large 3 (released December 2025) - a sparse Mixture-of-Experts model with 41 billion active parameters out of 675 billion total. In 2026 they have expanded into audio (Voxtral TTS, released March 26, 2026 - their first text-to-speech model, supporting zero-shot voice cloning in 9 languages) and document understanding (Mistral OCR 4, released June 23, 2026, supporting 170 languages and running fully self-hosted).

Mistral's enterprise play is Forge, announced in March 2026 at NVIDIA GTC - a platform for training frontier-grade models on a company's own proprietary data. If data sovereignty matters (you are in Europe, or in a regulated industry), Mistral is the lab most explicitly built around self-hosted and on-premise deployment.

xAI

Elon Musk's lab. Their model is Grok, now at version 4.3 (released April 30, 2026 on the public API), with a 1-million-token context window and native video input. Grok has one structural advantage no other model has: real-time integration with X (formerly Twitter), which means access to live public conversation at a scale that other models cannot match through web search alone.

xAI also released Grok Build 0.1 (June 2026), a coding-focused model in public beta via the API, and an image-to-video model preview. The lab is moving fast and the roadmap includes Grok 5, reportedly training at a 10-trillion-parameter scale. For builders who need real-time social signal or who are building on the X platform, Grok is the natural starting point.

DeepSeek

DeepSeek is the Chinese lab that rattled the entire industry in early 2025 by releasing models that matched frontier performance at a fraction of the reported training cost. Their current release is DeepSeek V4 (preview released April 24, 2026), built on a new hybrid attention architecture - Compressed Sparse Attention interleaved with Heavily Compressed Attention - and natively supporting a 1-million-token context window. The reasoning-focused variant, R1, remains widely used for math, code, and scientific tasks, though V4's built-in thinking mode increasingly covers the same ground.

All of DeepSeek's research models are released as open weights under the MIT license. That means the architectural innovations - Mixture-of-Experts, the hybrid attention approach introduced in V4, efficient training pipelines - are publicly available for anyone to study and adapt. DeepSeek is the lab most worth following if you care about the underlying technical frontier rather than the product surface.

How to orient without getting lost

With seven labs and dozens of models, the instinct is to track every release. That instinct will exhaust you. A more useful frame:

  1. Pick your use case first, then the model. Closed frontier APIs (GPT-5.5, Claude Fable 5, Gemini 3.5 Flash) are strongest for complex reasoning, long context, and agentic tasks. Open-weight models (Llama 4, Mistral Large 3, DeepSeek V4) are strongest when you need control, privacy, or domain-specific fine-tuning.
  2. The gap between open and closed is real but narrowing. Closed models still lead at the hard end of reasoning. Open models are roughly 6 months behind the frontier - and closing. For many production tasks, an open model is already good enough.
  3. Multimodal is the direction everything is moving. Text-in-text-out is no longer the interesting axis. Every major lab now ships or is building toward audio, image, and video as first-class inputs and outputs.
  4. Infrastructure matters as much as the model. The API you call is only part of the stack. File search, code execution, tool use, memory, and agent orchestration are where the real capability differences show up in practice.
# A simple decision tree for picking a starting point:
#
# Do you need maximum quality on a hard reasoning task?
#   -> Closed API: GPT-5.5 / Claude Fable 5 / Gemini 3.5 Flash
#
# Do you need data privacy or self-hosted deployment?
#   -> Open weight: Llama 4 Maverick / Mistral Large 3 / DeepSeek V4
#
# Do you need real-time social/news data?
#   -> Grok 4.3 (xAI) with X integration
#
# Do you need video understanding or generation?
#   -> Gemini Omni (Google DeepMind)
#
# Do you need to fine-tune on your own domain data?
#   -> Start with Llama 4 or DeepSeek V4 as the base

Key takeaways

  • Seven labs dominate: OpenAI and Anthropic lead on closed-weight quality; Meta and Mistral lead on open-weight reach; Google has the broadest multimodal surface; xAI has live social data; DeepSeek drives technical innovation from outside the US.
  • Closed vs. open weight is the most important structural split - it determines cost, control, and what you can legally do with the output.
  • The frontier is genuinely moving fast. Model versions from six months ago are often already retired or deprecated.
  • Multimodal (text + audio + image + video) is the direction every lab is building toward - text-only is the legacy case.
  • Pick your use case first, then pick the model. The map is a tool for decision-making, not a scoreboard.

Try this next: once you know who makes what, the natural question is how to actually call these models - prompting patterns, API basics, and when to reach for which tier. See The Prompting Mental Model for the practical next step.

References & sources

Reviews

Only verified humans can leave reviews. It keeps every rating real.

Verify to review

No reviews yet. Be the first to share your take.