Agnes 3.0 Flash

Agnes 3.0 Flash is a text model for agentic coding and tool-driven work from Agnes AI, the Singapore-based AI platform, released on 11 September 2026. Agnes AI's documentation describes it as focused on stable execution, reliable tool use, keeping hold of context over long tasks and trustworthy delivery, and serves it under the model id agnes-3.0-flash through Chat Completions, Responses and Messages-style endpoints, with text and image-URL input and up to 65,536 output tokens. The list price is $0.05 per million input tokens, $0.15 per million output tokens and $0.005 per million cached input tokens; the documentation currently shows all three at $0, with no end date given, and the list price is what is recorded here. Agnes AI's own pages disagree on the API model's context window: the documentation says 512K tokens, while the Hugging Face model card says the production/API model has a 1M-token window. So no context window is recorded. Alongside it, Agnes AI published an open-weights Preview checkpoint on Hugging Face under the Apache 2.0 licence: a 33-billion-parameter hybrid-attention model (54 recurrent delta-rule layers and 18 global-attention layers) with a 262,144-token context and text, image and video input. The model card says the Preview is a different checkpoint from the API model and that its benchmark table, including 85.05 on GPQA Diamond and 74.20 on IFBench, compiled across different harnesses, belongs to the Preview only, so no score is stored for Agnes 3.0 Flash. Agnes AI cites Artificial Analysis' Intelligence Index v4.3 score of 36 for the API model, which Artificial Analysis marked as an estimate pending independent evaluation.