Gemini 4 Argon starts with cyber defenders
A frontier model is arriving behind a controlled-access gate. Meanwhile, two retrieval releases give RAG builders fresh options for better context and lower-latency search.
In today’s edition
- Gemini 4 Argon begins with cyber defenders.
- Perplexity previews contextual document embeddings.
- Cohere ships Embed 5 Pro and Fast.
Gemini 4 Argon starts behind a cyber-defense gate
Google has introduced Gemini 4 Argon, a frontier model rolling out first to trusted cyber defenders through its Fairwind Program rather than to developers broadly.
- Google positions Argon for long-horizon work in software engineering, enterprise knowledge tasks, and cybersecurity defense.
- The model has a 1 million-token output limit for long, multi-step tasks.
- Google says feedback from early testers will inform a later rollout to developers, enterprises, and consumers.
- An introductory price announcement lists Argon at $2 input and $10 output.
What it means Keep using deployable models for production today, but track Argon closely because its 1 million-token output ceiling raises the bar for long-running agent tasks.
Source: GoogleDeepMind
Perplexity trains embeddings to retrieve supporting evidence
Perplexity has released a preview contextual embedding model that encodes document chunks with the full document in view, aiming to retrieve supporting evidence alongside answer passages.
- The preview model, `pplx-embed-v2-context-9b-preview`, is listed on Hugging Face.
- Its training method uses query-aware context compression to score document tokens, then turns those scores into chunk-level training targets.
- Perplexity says the model leads the average score on ConTEB, though not every task.
- The company says its 1024-dimensional int8 vectors use 1 KB each, versus 8 KB for Voyage Context 4’s 2048-dimensional float32 vectors.
What it means Test contextual embeddings on document-heavy RAG, where retrieving evidence as well as answer chunks may improve grounded answers without inflating vector storage.
Source: perplexity_ai
Cohere splits retrieval work between Embed 5 Pro and Fast
Cohere has launched Embed 5 Pro and Embed 5 Fast, two embedding models that share one vector space so teams can use different models for indexing and retrieval.
- Embed 5 is available through Cohere’s API, Microsoft Foundry, Amazon SageMaker, and Cohere Model Vault for single-tenant deployments.
- Cohere positions Pro for complex retrieval across visually rich documents, financial filings, parsed PDFs, code, and multilingual data.
- Because the models share an embedding space, Cohere says you can index with one and retrieve with the other while maintaining performance.
- Cohere says Fast beats other fast-tier models it tested by six or more points while costing a third less than Pro.
What it means Test indexing durable corpora with Pro and serving searches through Fast before maintaining separate embedding indexes for quality and low-latency retrieval.
Source: cohere