What is a large language model (LLM)?
A large language model (LLM) is a neural network, usually a transformer, trained on very large amounts of text to predict the next token, which lets it generate text, follow instructions and call tools when prompted.
Most current LLMs descend from the transformer architecture in "Attention Is All You Need" (Vaswani and colleagues, 2017). Scale turned a text predictor into a general tool: OpenAI's GPT-3 paper (May 2020) showed a 175-billion-parameter model doing new tasks from a few examples in the prompt, with no retraining.
For a team building a product, the model is a component with unusual properties. It is stateless between calls, priced per token, probabilistic in its output, and replaced every few months by a newer version with different behaviour.
How an LLM is built
Pretraining
The model learns to predict the next token across a very large text corpus. This is where its general knowledge and language ability come from.
Post-training
Instruction tuning and reinforcement learning from human feedback, described in OpenAI's InstructGPT paper (2022), turn a text predictor into an assistant that follows requests.
Reasoning training
Since OpenAI's o1 (September 2024) and DeepSeek-R1 (January 2025), many models are also trained with reinforcement learning to reason step by step before they answer.
Tool use
Models are trained to emit structured tool calls, which is what lets them act inside an agent rather than only write text.
What it means in production
No memory between calls
Everything the model needs for a task has to be in the context window on every request: instructions, history, retrieved data and tool results.
Priced per token
Input and output are billed separately. As of November 2026, Anthropic lists Claude Opus 5.5 at $4 input and $20 output per million tokens, and Claude Haiku 5.5 from $0.10 and $0.50.
Output varies
The same prompt can give different answers, so quality is measured across many runs with evals, not judged from one transcript.
Models retire
Anthropic's deprecation page lists Claude Sonnet 4.5 retiring on 30 November 2026, with Sonnet 5.5 as the replacement. A model swap is a change that needs a regression run.
Choosing a model
The major vendors now ship tiers: a flagship, a workhorse, a mid tier and a small, cheap model. OpenAI's guide to building agents gives a practical order: prototype on the most capable model, set an eval baseline, then try smaller models and keep the cheapest one that holds the baseline.
Open-weight models such as the DeepSeek V4 family can run on a team's own infrastructure, which matters for data residency and cost at volume. They bring the work of serving, scaling and updating them in-house.
Where LLMs fail
Hallucination
Fluent, confident output that is false or unsupported.
Long-context decay
Recall drops as the input grows. Chroma's Context Rot study (July 2025) found this across 18 models.
Instruction confusion
The model cannot reliably tell trusted instructions from text inside a document it is reading, which is what prompt injection exploits.
How Motiontic's engineers apply it
Motiontic is vendor-neutral. Engineers pick models per task from eval results on the client's own cases, keep the model ID in configuration rather than code, and run the regression suite before any model change ships.
An LLM, an augmented LLM and an AI agent
| LLM | Augmented LLM | AI agent | |
|---|---|---|---|
| What it is | A model that maps input tokens to output tokens | A model with retrieval, tools and memory attached | An augmented LLM running in a loop toward a goal |
| Who decides the next step | The calling code | The calling code, with the model choosing tools inside one call | The model, turn by turn |
| Knows live data | No, only what is in the prompt | Yes, through retrieval and tools | Yes, and acts on it |
| Typical use | Drafting, summarising, classifying | Answering from company documents | Finishing a multi-step task |
Common questions
What does LLM stand for?
Large language model: a neural network with billions of parameters, trained on large text corpora to predict the next token, and then tuned to follow instructions.
Is ChatGPT an LLM?
ChatGPT is a product built on OpenAI's LLMs. The product adds a chat interface, memory, tools such as search, and safety layers around the underlying models.
What is the difference between an LLM and an AI agent?
An LLM produces output for one request. An AI agent wraps an LLM in a loop with tools, so the model can take an action, read the result and decide the next step until a task is done.
How do companies choose which LLM to use?
By running their own tasks through candidate models and comparing quality, cost and latency. Vendor benchmarks are run on each vendor's own harness, so they rarely settle the question for a specific product.
Related
Augmented LLM
A model plus retrieval, tools and memory: the unit every workflow and agent is built from.
Token (LLM)
The unit models read, write and bill in: a word or a piece of one.
Context window
The model's working memory for one request, measured in tokens, and finite.
Claude vs ChatGPT
Plans, API prices, context limits and agent tooling for Anthropic's Claude and OpenAI's GPT models.
AI product development
New products and major AI features, built from scratch alongside the client's team.
Sources
- Vaswani et al., Attention Is All You Need (arXiv 1706.03762, 2017)
- Brown et al., Language Models are Few-Shot Learners (arXiv 2005.14165, 2020-05-28)
- Ouyang et al., Training language models to follow instructions with human feedback (arXiv 2203.02155, 2022)
- Anthropic docs, Pricing
- Anthropic docs, Model and parameter deprecations
- OpenAI, A practical guide to building agents (2025)
- Chroma, Context Rot (2025-07-14)
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.