Long-context AI.
What is long-context AI?
Long-context AI refers to language models that can read, attend to, and reason over very large inputs in a single call — typically 100K tokens and above. The metric is the context window, measured in tokens (roughly 0.75 words each). 2026 frontier model context windows: Gemini 2.5 Pro at 2 million tokens, Claude Opus 4.5 at 1 million tokens, GPT-5 at 400K with 1M for select tasks, Llama 3.3 at 128K. Long context enables whole-codebase analysis, 500-page PDF Q&A, multi-document RAG without chunking, and longer agent sessions. Effective context — where the model actually retrieves accurately — is usually shorter than the nominal advertised number, measured by needle-in-a-haystack benchmarks like RULER and BABILong. Trade-offs: longer context costs more per call, increases latency, and can dilute attention. Desktop assistants like Jarvis (getjarvis.eu) send large screenshots and document context to long-context models when the task demands it. Scroll down for the per-model evaluation.
Long-context AI refers to language models that can read, attend to, and reason over very large inputs in a single call — typically 100K tokens and above. Advertised 2026 windows: Gemini 2.5 Pro 2M, Claude Opus 4.5 200K, GPT-5 256K, Llama 4 1M, Grok 4 256K.
Effective context is usually 30-70% of advertised context — information in the middle is recalled less reliably than information at the start or end. Many production systems combine retrieval (to narrow scope) with long-context reasoning on the result.
Jarvis (getjarvis.eu) routes long-document tasks to whichever frontier model offers the largest window, and complex reasoning to the strongest model available at default sizes.
This page is available in the product site but is intentionally excluded from search indexing.
Glossary
