The Open-Weights Race Just Got a 10M-Token Problem

For two years, the open-weights race was measured in benchmark points. This week it started being measured in tokens — ten million of them.

Meta's Llama 4 Scout landed with a context window large enough to hold roughly 7,500 pages of text in a single pass, and the rest of the ecosystem responded within 48 hours with their own long-context claims.

#Why context became the battleground

Reasoning benchmarks have saturated. The top open and closed models now sit within a few points of each other on math and code suites, so labs need a new axis to compete on. Context is that axis.

Long context unlocks workloads that short-context models simply cannot touch:

• Whole-repository code agents that read the codebase in one pass
• Multi-hour meeting and video summarization
• Legal and scientific document review across hundreds of files
• Persistent agent memory without external retrieval

#The catch nobody mentions

A 10M-token window is a capability claim, not a quality guarantee. Attention quality degrades unevenly across long contexts, and most public evaluations still test retrieval at 128K or less.

#What to watch next

Expect DeepSeek and Qwen to respond within the quarter, and expect inference providers to start pricing long-context requests differently. The economics of attention are about to become the economics of the whole stack.