Context Window

The memory capacity of artificial intelligence — and the bottleneck that determines whether your AI assistant remembers your conversation or forgets everything you said. In 2026, the AI industry is locked in a fierce competition to expand context windows — the amount of text an AI model can process at once. Google's Gemini 2.0 supports 2 million tokens (roughly 1.5 million words), whilst Anthropic's Claude 4 handles 200,000 tokens, and OpenAI's GPT-5 operates with 256,000 tokens. As of June 2026, Cohere has also announced an upgrade to its models, now supporting 256,000 tokens, aligning with OpenAI's offerings. For enterprises deploying AI, context window size directly determines what tasks are possible, what the costs will be, and where the technology's limitations lie. Understanding this single parameter is essential for making informed AI investment decisions.

What It Is

A context window is the maximum amount of text (or data) that an AI language model can consider at any one time when generating a response. Think of it like a whiteboard: the AI can only work with what fits on the board. Once the board is full, the oldest information must be erased to make room for new content. If you have a conversation with an AI that exceeds its context window, the model literally forgets the earliest parts of the exchange. The unit of measurement is the "token," which roughly corresponds to three-quarters of a word in English. A context window of 100,000 tokens is approximately 75,000 words — about the length of a full novel. This means a model with a 100,000-token context window could read and analyse an entire book in one go, maintaining awareness of every detail from beginning to end. The technical constraint arises from how transformer-based AI models work. The attention mechanism — the core innovation that makes large language models powerful — scales quadratically with input length. Doubling the context window roughly quadruples the computational cost. This is why expanding context windows is both a technical achievement and an expensive proposition. A query using a 1-million-token context window can cost 50 times more than a standard query, a factor that has significant implications for enterprise AI budgets.

Why It Matters for Business

Context window size determines the practical use cases for AI in your organisation. With a small context window (say, 8,000 tokens or roughly 6,000 words), an AI can handle emails, short documents, and brief conversations. With a large context window (200,000+ tokens), the same AI can analyse entire contracts, process complete financial reports, summarise hundreds of pages of research, or maintain coherent long-term customer interactions. For legal teams, this means the difference between an AI that can review a single clause and one that can analyse an entire 200-page merger agreement, cross-referencing definitions and identifying contradictions across the full document. Financial services firm Morgan Stanley reported that large context windows enabled its AI assistant to analyse complete client portfolios — including decades of transaction history — in a single query, reducing advisory preparation time by 70%. Cost is the other critical dimension. Processing a full context window of 200,000 tokens on GPT-4-class models costs approximately $6–12 per query, depending on the model. At enterprise scale — thousands of queries per day across hundreds of employees — context window usage can generate AI compute bills of $500,000 to $5 million annually. Smart context management, including chunking strategies and retrieval-augmented generation (RAG), can reduce these costs by 60–80% while delivering equivalent results for many use cases.

Enterprise Use Cases

Thomson Reuters integrated large-context-window AI into its legal research platform, enabling lawyers to upload entire case files — sometimes exceeding 500 pages — and receive comprehensive analyses that identify relevant precedents, flag potential risks, and suggest legal strategies. The system reduced average legal research time from 12 hours to 45 minutes per case, with accuracy rates comparable to senior associates. In pharmaceuticals, Pfizer uses large context windows to process complete clinical trial protocols and regulatory submissions in a single pass. Previously, these documents had to be split into sections for AI analysis, losing critical cross-references. With expanded context, the AI can identify inconsistencies between trial design and statistical analysis plans, accelerating regulatory review cycles by an estimated 30%. Customer service platform Zendesk uses context windows to maintain coherent, long-running support conversations. An AI agent with a large context window can remember a customer's entire interaction history — purchases, previous complaints, preferences — and provide personalised support without requiring the customer to repeat information. Companies using this approach report 25% improvements in customer satisfaction scores and 35% reductions in average resolution time.

Vendor Landscape

Google leads in raw context window size with Gemini 2.0's 2-million-token window, available through Google AI Studio and Vertex AI, with pricing at approximately $1.25 per million input tokens. Anthropic offers 200,000 tokens on Claude 4, with pricing around $3 per million input tokens. OpenAI provides 256,000 tokens on GPT-5 models, with costs ranging from $2.50 to $10 per million tokens depending on the model tier. Cohere, focused on enterprise deployments, has upgraded its offerings to support 256,000 tokens with competitive enterprise pricing. For organisations deploying open-source models, Meta's Llama 3 supports up to 128,000 tokens, and Mistral's models handle 128,000 tokens, both available for self-hosting with no per-token costs but requiring significant GPU infrastructure investment.

Strategic Considerations

Do not default to the largest available context window. Larger contexts increase costs, latency, and sometimes reduce output quality — models can become "distracted" by irrelevant information in a vast context. For most enterprise u... [truncated]