Company, Model, Product and Agent

The easiest way to misunderstand AI is to compare names at different layers. ChatGPT is not a model in the same sense as GPT-5.6 Sol. Claude is both a consumer product name and the name of Anthropic’s model family. Gemini refers to Google’s model family and also appears across consumer products. Meta AI is a consumer assistant built on Meta models, while Llama refers to downloadable model families. A coding agent may use a general-purpose model but add shell access, repository indexing, patch application and persistent state.

The distinction matters because users often attribute product behavior entirely to the model. Search quality may come from a retrieval system. Coding success may depend on a harness that can run tests and inspect files. A model may appear to “remember” because the product stores account-level memory outside the model. An agent may complete a task because it can use a browser, shell or API rather than because the model internally knows the answer.

Dense Models, Mixture of Experts and Architecture

Traditional transformer language models are often described as dense: broadly speaking, each forward pass uses the model’s full parameter stack. Mixture-of-experts architectures route tokens through subsets of specialized feed-forward “experts,” allowing total parameter count to become much larger than the number of parameters active for each token. Llama 4 Maverick, for example, was disclosed by Meta as a 400-billion-total-parameter model with 17 billion active parameters and 128 routed experts. DeepSeek V4 Pro is disclosed as 1.6 trillion total parameters with 49 billion active. Qwen3.5-397B-A17B similarly exposes 397 billion total and 17 billion active parameters.

MoE changes the economics of comparison. A model can store enormous capacity while activating only part of it at inference. Total parameter count affects memory and storage; active parameter count is more directly connected to per-token compute. Routing quality, attention architecture, training data, post-training and inference implementation can matter as much as either number. This is why a single “how many parameters?” question is no longer enough.

Why Parameter Count Can Mislead

Parameter count was useful when the industry openly compared dense models of similar architecture. It becomes less reliable when some companies publish full technical reports and others disclose almost nothing. OpenAI does not publish the parameter count for GPT-5.6 Sol. Anthropic does not publicly disclose the parameter count for Claude Fable 5. Google does not provide a public parameter count for Gemini 3.5 Flash in its model documentation. SpaceXAI does not publicly disclose Grok 4.5’s parameter count in its launch material. The correct entry in a reference table is therefore not an estimate from social media or leaked speculation. It is NOT PUBLICLY DISCLOSED.

Even when parameters are disclosed, capability does not scale mechanically with them. Training data quality, reinforcement learning, synthetic data, inference-time reasoning, distillation, architecture and tool integration can allow a smaller system to outperform a much larger one on a specific task. Parameter count should therefore be treated as a systems specification rather than an intelligence score.

Context Windows and Context Utilization

Context window measures how much tokenized information a model can accept in one interaction. OpenAI’s current GPT-5.6 Sol API documentation lists a 1.05-million-token context window. Google’s Gemini 3.5 Flash supports 1,048,576 input tokens. DeepSeek V4 is offered with a one-million-token context. Qwen3.5 Plus similarly advertises a one-million-token hosted context window. Meta’s Llama 4 Scout was designed around an unusually large ten-million-token supported context.

Those numbers are not interchangeable with reliable reasoning across that entire length. Long-context performance can degrade because relevant details compete with irrelevant material, attention patterns become difficult, retrieval is uneven and the model may compress or prioritize information imperfectly. A million-token context window is a capacity limit. Context utilization asks whether the system can identify, retain and reason over the right information within that capacity.

For many workflows, retrieval-augmented generation remains preferable to dumping an entire archive into context. Search can narrow the evidence set. Summaries can preserve high-value structure. Agentic systems can fetch material only when needed. Long context is powerful, but intelligent information management often matters more than maximum input size.

Reasoning Models and Test-Time Compute

The model market increasingly exposes reasoning as a tunable resource. OpenAI’s GPT-5.6 family supports multiple reasoning-effort settings. Google Gemini 3.x exposes thinking levels. DeepSeek V4 provides thinking and non-thinking modes. Anthropic’s current Claude products expose effort controls. The underlying idea is that difficult tasks can justify more inference computation, while high-volume routine tasks should not pay the same latency and cost.

This changes the meaning of a benchmark. A model tested at maximum reasoning may consume far more tokens and time than the same model at default settings. A cheaper model with more reasoning may compete with a larger model at lower reasoning. Comparisons therefore need to include cost, latency and inference configuration rather than reporting one score in isolation.

Multimodality

General-purpose models are increasingly trained or integrated across more than text. Google’s Gemini 3.5 Flash accepts text, image, video, audio and PDF input. Meta describes Llama 4 Scout and Maverick as natively multimodal through early fusion of text and vision. Qwen3.5 is a native vision-language model and Alibaba has expanded the family into omni-modal variants. OpenAI GPT-5.6 Sol accepts text and image input in the API while other OpenAI models and product layers handle audio, image generation and related modalities.

It is important to distinguish multimodal understanding from multimodal generation. A model may be able to interpret an image while outputting only text. A product may generate images through a separate image model. An “omni” system may unify more modalities natively. The user experience can look seamless even when the architecture underneath is a collection of specialized models.

Tools, MCP and Computer Use

Tool use is rapidly becoming as important as raw model knowledge. OpenAI’s Responses API exposes web search, file search, computer use and other tools to GPT-5.6 Sol. Google’s Gemini 3.5 Flash supports function calling, code execution, file search, search grounding and computer use in preview. Qwen3.5 emphasizes built-in tools and GUI interaction. DeepSeek V4 supports tool calls and is positioned for agentic coding. Anthropic’s Claude ecosystem centers heavily on tool-using workflows through Claude Code and related products.

The Model Context Protocol, function calling and similar standards let a model interact with external systems through defined schemas. This changes capability from “What does the model know?” to “What can the model access and do?” It also changes risk. A hallucination in plain text is a content error. A hallucinated tool call can create a real transaction or file change. That is why Articles 003 and 025 treat authorization and auditability as core agent infrastructure.

Open Weights, Local Deployment and Sovereignty

Open-weight models provide a different form of control. Organizations can download weights, run inference on their own hardware, inspect model files, quantize them, fine-tune them and deploy them without sending every prompt to a vendor API. Meta Llama, DeepSeek, Qwen and Mistral have all made major model weights available to developers, though licenses differ and should be read carefully.

Local deployment matters for privacy, latency, cost control, customization and national technological sovereignty. A company handling sensitive records may prefer a model inside its own environment. A military or government system may require operation without external connectivity. Edge devices may need low latency. Researchers may need direct access to weights. But self-hosting also transfers responsibility: infrastructure, patching, safety controls, evaluation and misuse prevention no longer come entirely from the upstream provider.

Current Major Model Families

The following table is a current reference snapshot rather than a permanent ranking. It prioritizes publicly documented specifications and deliberately avoids filling gaps with unofficial estimates. “Open weight” refers to weights available for download under the developer’s stated license; it does not automatically mean the entire training stack is open source.

OpenAI GPT: Frontier Model Plus Agentic Harness

OpenAI’s current GPT-5.6 family uses three capability tiers: Sol, Terra and Luna. Sol is the flagship model for complex professional work, while Terra and Luna trade capability for cost and speed. OpenAI’s API documentation lists GPT-5.6 Sol with a 1.05-million-token context window and up to 128,000 output tokens. The company does not publicly disclose Sol’s parameter count, so any numerical estimate outside official documentation should be treated as speculation.

OpenAI increasingly positions the model inside a tool-rich system. GPT-5.6 Sol supports function calling, web search, file search, code execution, hosted shell, computer use and MCP-related tool access through the Responses ecosystem. The company also exposes reasoning-effort settings and an “ultra” mode that can coordinate multiple sub-agents for complex work. This illustrates the model-versus-harness distinction clearly: practical capability comes from the model plus orchestration.

Anthropic Claude: Long-Horizon Professional and Agentic Work

Anthropic’s current family includes Claude Fable 5, Sonnet 5 and Opus 4.8 among its major public offerings. Fable 5 was launched in June 2026 as Anthropic’s highest-capability generally available model, with special safety routing for some cyber, biology and chemistry requests. Anthropic does not publicly disclose Fable 5’s parameter count, so the correct technical entry remains NOT PUBLICLY DISCLOSED.

Anthropic has differentiated itself strongly through Claude Code and long-running agentic workflows. Its current products emphasize coding, research, tool use and professional work, while safety classifiers can route some higher-risk requests to different models. That is an important reminder that “which model answered?” can itself become dynamic inside modern safety architectures.

Google Gemini: Multimodality, Long Context and Action

Google’s Gemini family is among the clearest examples of multimodal AI becoming a general-purpose platform. Gemini 3.5 Flash is documented with a 1,048,576-token input limit, support for text, image, video, audio and PDF input, and built-in capabilities including code execution, file search, search grounding, function calling and computer use. Google has also introduced Gemini 3.6 Flash as a newer production model balancing speed and intelligence.

Google does not publish a parameter count for these proprietary models in its API documentation. The strategic advantage instead comes from vertical integration: Gemini can connect to Search, Google Cloud, Android, Workspace, developer tooling and multimodal products. Google’s model family therefore competes not only on weights but on access to a giant software and information ecosystem.

SpaceXAI Grok: Coding, Knowledge Work and the X Ecosystem

Grok 4.5 launched in July 2026 as SpaceXAI’s newest flagship model for coding, agentic tasks and knowledge work. The company emphasizes engineering benchmarks and integration into Grok Build, its coding-agent environment. SpaceXAI does not publicly disclose a parameter count in the launch material, so parameter estimates should not be presented as technical facts.

Grok’s differentiating environment includes the X platform and SpaceXAI’s own product stack. That can provide product-level advantages distinct from the model itself, especially around real-time information, coding workflows and consumer distribution. As with other closed systems, model capability and product capability should be evaluated separately.

Meta Llama and Muse: Open Weights Meet Personal Superintelligence

Meta’s Llama 4 family remains important because it provides unusually detailed public architecture. Maverick uses 400 billion total parameters, 17 billion active parameters and 128 routed experts. Scout uses 109 billion total parameters, 17 billion active parameters and 16 experts, with support for an unusually large ten-million-token context. Both are natively multimodal text-and-vision models and are available as downloadable weights under Meta’s license.

Meta has also begun describing a newer internal direction around Muse Spark and personal superintelligence. That development should not be conflated automatically with Llama releases. Llama remains the open-weight developer family; Muse represents a newer frontier direction whose deployment and openness may follow a different path. The field guide should therefore track Meta’s model strategy by family rather than assuming every Meta model is simply “Llama.”

DeepSeek: Open Weights, Efficiency and Sparse Attention

DeepSeek became strategically important by demonstrating that frontier-class capability and low-cost inference could coexist with open-weight distribution. DeepSeek V4, released in April 2026, extends that thesis with two models. V4 Pro is disclosed as a 1.6-trillion-total-parameter model with 49 billion active parameters, while V4 Flash uses 284 billion total and 13 billion active. Both support a one-million-token context in official services and thinking/non-thinking modes.

DeepSeek’s technical materials emphasize sparse attention and agentic optimization. The company’s transparency center and API documentation also make model versions easier to track than many proprietary competitors. Because the weights are available, researchers can inspect, quantize and self-host the systems. The tradeoff is that deployment organizations must take more responsibility for security, safety and infrastructure themselves.

Alibaba Qwen: Native Multimodal Agents

Qwen3.5 illustrates how quickly open-weight Chinese model families have moved into the frontier conversation. Alibaba disclosed Qwen3.5-397B-A17B as a 397-billion-total-parameter mixture-of-experts model with 17 billion active parameters. The architecture combines sparse MoE with a hybrid attention system, and the model is trained for text, image and video understanding. Alibaba’s hosted Qwen3.5 Plus offers a one-million-token context window and built-in tool use.

The Qwen team explicitly frames the family around native multimodal agents, including GUI interaction and computer-oriented workflows. This is strategically important because it turns open weights from a language-model story into an agent-platform story. Organizations can use the model as a local or cloud foundation for systems that see interfaces and act through software rather than merely answer questions.

Mistral: Efficiency, Permissive Licensing and Enterprise Customization

Mistral has built its identity around efficient models, European technological independence and permissive deployment. Mistral Large 3, released in late 2025, is a sparse mixture-of-experts model with 675 billion total parameters and 41 billion active. Mistral Small 4, released in March 2026, combines reasoning, multimodal understanding and agentic coding in a smaller model and is released under Apache 2.0.

The company is also pushing beyond generic model APIs through Forge, which allows enterprises to build models grounded in proprietary organizational knowledge. That reflects another major direction in the model market: the future may include fewer “one model for everyone” deployments and more organization-specific models built or adapted around proprietary data and workflows.

Other Important Families

The field is too broad to reduce to eight companies. Moonshot’s Kimi family is competitive in long-context, coding and agentic work. Cohere continues to focus heavily on enterprise retrieval and secure deployment. AI21, MiniMax, GLM/Zhipu, Baidu, ByteDance and other Chinese and international developers remain important. Specialized coding, video, image and scientific models can outperform general-purpose systems within narrower domains.

The correct editorial rule is therefore inclusion by strategic relevance rather than an attempt to catalog every model release. A field guide should track the families that materially influence architecture, pricing, openness, agent design or deployment practice. It should not become a weekly leaderboard.

The Model-Versus-Agent Divide

A language model produces an output from an input. An agent maintains state across steps, selects tools, observes results and decides what to do next. The distinction is becoming increasingly important because agents can outperform a raw model on complex work by externalizing memory and action. A coding agent can inspect a repository, run tests, make a patch, observe failure and try again. A research agent can search, open sources, compare evidence and write a report. The underlying model may be unchanged, but the system becomes much more capable.

That also means benchmarks should report the harness. A model with computer access and repeated tool calls is not directly comparable to the same model answering in one pass. OpenAI, Anthropic, Google, SpaceXAI, DeepSeek and Qwen all increasingly optimize models for agentic environments, making this one of the most important structural changes in the field.

Safety and Capability Risk

As model capability rises, safety evaluation is shifting from generic harmful-content tests toward capability-specific risk. Frontier developers increasingly evaluate cybersecurity, biological assistance, autonomous tool use, persuasion and other domains where the question is not simply whether a model says something offensive but whether it materially increases a user’s ability to carry out harmful or high-consequence work.

Anthropic’s Fable 5 provides a particularly visible example: the company routes some categories of cyber, biology and chemistry requests away from its most capable model through additional classifiers. OpenAI’s GPT-5.6 launch similarly emphasizes stronger safeguards for high-risk cyber and scientific capabilities. These systems show that safety can be part of the serving architecture rather than only embedded in the model weights.

Open-weight models create a different governance environment because downstream users may remove or replace safety layers. That increases research freedom and local control, but it also means safety becomes decentralized. The relevant comparison is not “open equals unsafe” or “closed equals safe.” It is who controls the safeguards, who can modify them and what accountability exists around deployment.

How to Choose a Model

The best model is workload-specific. A research organization may value long context, citation-quality tool use and scientific reasoning. A coding team may value repository-scale agent reliability. A call-center deployment may prioritize latency and price. A government environment may require local weights and network isolation. A creative studio may prioritize image, audio and video understanding. A regulated enterprise may value auditability, contractual controls and data residency more than a small benchmark advantage.

This is why model selection should be performed as an engineering procurement decision rather than a popularity contest. The organization should define the task, identify failure costs, test multiple models under the same harness, measure latency and total cost, evaluate data-handling terms and verify that required tools and deployment modes exist. Benchmark leadership can help identify candidates, but production evidence should decide the winner.

The SURVXCOM Model Selection Test

The model market is too dynamic for a permanent winner, but organizations can use a durable evaluation framework. SURVXCOM therefore evaluates a model and its deployment environment across twelve dimensions.

What to Watch Next

Model families becoming durable while snapshots accelerate. OpenAI’s Sol/Terra/Luna structure and similar tiering strategies may make family names more stable even as underlying generations change. Watch whether vendors make model selection simpler or continue proliferating overlapping names.

Parameter disclosure declining in importance. Proprietary vendors are increasingly withholding architecture details while open-weight developers publish them. Watch whether independent evaluation and systems-level cost metrics replace parameter count as the dominant comparison language.

Agent benchmarks. Coding and research are increasingly evaluated through multi-step environments rather than static question answering. Watch benchmark designs carefully for tool access, time budget and harness differences.

Context beyond one million tokens. Large context windows are now common among major model families. The next competitive question is reliable retrieval and reasoning across those contexts, not simply accepting a larger input.

Open-weight frontier competition. DeepSeek, Qwen, Llama and Mistral continue narrowing the practical gap with proprietary systems. Watch licensing, local deployment cost and agent tooling as much as raw benchmark performance.

Multimodal agents. Models increasingly understand screens, video and documents while using computer interfaces. This moves AI from language assistance into general software operation.

Reasoning cost curves. Models are becoming configurable across intelligence and cost tiers. The practical winner may be the family that delivers enough reasoning at the lowest total task cost rather than the absolute highest score.

Safety routing. Anthropic and OpenAI are showing that the serving system can dynamically restrict or reroute high-risk capability. Watch whether safety becomes a standardized multi-model orchestration layer.

Local agents. As open weights become more capable and hardware improves, organizations may increasingly run private agents on-premises. That will shift security and governance responsibility away from model vendors and toward deployers.

Model sovereignty. Article 030 will examine the national version of this question: whether countries can truly control AI if they depend on foreign chips, clouds, power systems, models and network infrastructure.

The Intelligence Layer

Language models began as software components that predicted the next token. They are becoming something broader: an intelligence layer through which humans and software increasingly interact with information, code, tools and machines. That does not mean the model itself has become an operating system, but it increasingly sits where operating systems once sat in the hierarchy of control. Users state intent in natural language. Models interpret that intent. Agents translate it into actions. Tools execute those actions in the underlying digital or physical environment.

This is why model selection is becoming strategic. Choosing an AI model can determine where data travels, which company mediates access to knowledge, how much inference costs, which safety rules apply, what tools can be used and whether the organization can move to another provider later. A model is no longer merely a text generator. It can become part of the institutional architecture through which work is performed.

The market will remain volatile. GPT, Claude, Gemini, Grok, Llama, DeepSeek, Qwen, Mistral and newer families will continue leapfrogging one another. Some current leaders will lose ground. Some model names will disappear. Pricing will fall. Context windows will grow. Agents will become more capable. Open models will improve. Proprietary systems will gain tighter integration with enormous product ecosystems.

The durable skill is therefore not memorizing the leaderboard. It is learning how to read the system. Is the model dense or mixture-of-experts? Are parameters disclosed? What context can it accept, and how well does it use that context? Does it reason at configurable effort? What modalities does it understand? Can it use tools? Can it act on a computer? Are the weights available? What license applies? What data leaves the organization? Which safety layer controls the output? What happens when the provider changes the underlying model?

Those questions reveal why the field guide belongs inside the initial Critical Technology tree. Artificial intelligence is not one technology. It is becoming a stack of competing intelligence platforms whose architectural choices determine cost, control, capability and risk across nearly every other technology in the stack.

The winner of the AI model race will not necessarily be one permanent model. The deeper contest is over which model ecosystems become the default intelligence layer through which people, companies and autonomous systems understand information and act on the world.