AI News

LLMs reach 1 million tokens of context

The context windows of large language models now reach 1 million tokens. A quiet revolution that opens up new use cases for document analysis and software development.

5 min read
LLMContexteInnovationAnthropicGoogle
⚡ The news in 30 seconds

1 million tokens: LLMs can now read entire libraries

Google Gemini 1.5 Pro and the latest models from Anthropic are pushing context windows beyond one million tokens. Concretely, an LLM can now process in a single request the equivalent of 10 books, a complete application source code, or 3 years of financial reports. This technical advance, long considered a distant horizon, transforms the possibilities for document analysis in business.

💡 1 million tokens is 750,000 words: enough to analyze an entire mid-market company's product catalog in a single request.

Opportunities for businesses

Extending context windows is not just an impressive number. It opens up unprecedented use cases that were technically impossible a year ago. For SMBs and mid-market companies, three applications stand out for their immediate impact.

📚

Massive document analysis

Inject an entire tender (200+ pages), internal documents and the regulatory context into a single prompt. The model produces a coherent answer by cross-referencing all the sources. No more manually splitting documents that are too long.

💻

Complete code audit

Developers can submit an entire codebase (50,000 to 200,000 lines) to obtain a security audit, automatic documentation or guided refactoring. Claude and Gemini understand the dependencies between files without losing the thread.

📈

Consolidated strategic intelligence

Aggregate 6 months of sector reports, articles and internal notes. Ask the model to identify emerging trends and competitive threats. The synthesis obtained in 2 minutes replaces 3 days of an analyst's work.

Risks and limits to know

Despite the legitimate enthusiasm, context windows of 1 million tokens have technical and economic limits that must be anticipated.

⚠️

Degradation on extreme contexts

"Needle in a haystack" benchmarks show that model accuracy drops by 5 to 15% beyond 200K tokens. Information located in the middle of a very long context is more likely to be "forgotten". For critical tasks, favor structured and segmented contexts.

🔒

Costs and data confidentiality

Sending 1 million tokens costs between 10 and 25 € depending on the model. With intensive use, the bill climbs quickly. Moreover, injecting all of your sensitive documents into a third-party API requires careful thought about data security and GDPR compliance.

Our recommendations

To take advantage of long contexts without falling into the traps, follow this three-step approach.

1

Identify your high-value use cases

Focus on tasks that genuinely require a wide context: analysis of document corpora, code audits, report consolidation. Don't overload the context out of habit: a well-targeted prompt of 10K tokens often remains more effective than a dump of 500K tokens.

2

Combine long context and RAG

The optimal approach pairs a vector database (RAG) for the initial search with long context for in-depth analysis. Retrieve the 50 most relevant passages via RAG, then submit them in a long context for a coherent synthesis.

3

Control costs from the start

Set up a prompt caching system (available from Anthropic and Google) to avoid paying for the same context again on every request. Monitor your consumption via the API dashboards and set monthly budget limits.

Key takeaways

Maximum capacity
1M tokens (Gemini)
Text equivalent
≈ 750,000 words
Cost per max request
10-25 €
Optimal reliability zone
< 200K tokens

Frequently asked questions

What is a token and why is context important?

A token is the unit of text processed by an AI model (about 0.75 word in French). The context window determines how much text the model can "read" at once. With 1 million tokens, an LLM can analyze about 750,000 words, the equivalent of 10 books or a complete application source code.

Are these models accessible to SMBs today?

Yes. Claude 3.5 Sonnet offers 200K tokens of context and Gemini 1.5 Pro reaches 1M tokens, both via paid pay-as-you-go API. The cost remains reasonable for targeted use cases: count 5 to 15 € to analyze a 500-page document.

Does quality degrade over long contexts?

Partially. Benchmarks show that performance remains stable up to 200K tokens, then can drop by 5 to 15% beyond that. For critical use cases, it is recommended to structure documents and combine long context with RAG techniques (retrieval-augmented generation).

Models and context capacities

Gemini 1.5 Pro

Google DeepMind

The first commercial model to reach 1 million tokens of context. Excellent performance in multimodal analysis (text, image, video). Available via the Google AI API and Vertex AI.

Claude 3.5 Sonnet

Anthropic

200K tokens of context with exceptional reliability on long documents. State-of-the-art performance in code analysis and reasoning. Supports prompt caching to reduce costs.

GPT-4o

OpenAI

128K tokens of context. A versatile model that performs well on synthesis and document analysis. Native integration into the Microsoft ecosystem for companies already equipped.

Pricing (per million tokens)

Gemini 1.5 Pro $3.50 / $10.50
Claude 3.5 Sonnet $3 / $15
GPT-4o $2.50 / $10

Comparison

CriterionGemini 1.5 ProClaude 3.5 SonnetGPT-4o
Max context1M tokens200K tokens128K tokens
Long-context reliability⭐⭐⭐⭐⭐⭐⭐⭐
Code analysis⭐⭐⭐⭐⭐⭐⭐⭐
Multimodal⭐⭐⭐⭐⭐⭐⭐⭐
Prompt cachingYesYesNo

Related articles