Anthropic reports Claude leading 26% of internal AI R&D under human supervision in August 2026. Prototype, self-reported measurement; no work classified as fully autonomous.
Standard input price, excluding the temporary launch discount. Output costs $0.75 per million tokens. This is a model-specific price, not an overall frontier minimum.
Largest context window available from a frontier model
10M tokens
Llama 4 Scout
Meta, Google, Anthropic2026-04-09
Secondary source
Context Window Frontier: History
2023-11128,000
2024-021,000,000
2024-052,000,000
2025-0410,000,000
METR Time Horizon
METR v1.1 estimates Mythos Preview (early) at 17.4 hours for 50% success, with an 8.5 to 55.1 hour confidence interval. Estimates above 16 hours are unreliable on this task suite.
Figure reports Helix 2.5 completing 56% of trials across three household tasks in 30 unseen homes, up from 9% without Index pretraining. No adaptation in evaluation homes; this is a controlled comparison, not a time series.
AlphaGenome Atlas provides predicted molecular effects for 9 billion possible single-letter DNA changes. Researchers can prioritize variants for experiments across the genome; these are predictions, not 9 billion validated treatments.