—
Cohere announced a proposed agreement to acquire AISLE, which builds autonomous systems to discover and remediate previously unknown vulnerabilities in critical software. The proposed deal would add security capabilities to Cohere’s sovereign AI platform and deepen its presence in Europe.
Teams of coding agents unlock a new axis of inference-time scaling, achieving up to 10 percentage points improvement on ProgramBench compared to single agents. However, these gains come at 3–7× higher cost, and the key factor is where agents are allocated rather than how many are added. Delegating only exploration and planning can match the performance of full delegation at half the cost, showing that collaboration is most effective when agents target the task’s actual bottleneck.
Cohere’s analysis shows that the shape of embedding requests, rather than sheer volume, drives compute usage. Dedicated inference is more cost‑effective for steady workloads, while consumption‑based inference suits variable traffic. The economic crossover between shared and dedicated models depends on request size, token volume, and reranking candidate count, with latency and throughput also influencing the decision.
The article argues that truly multilingual AI should reason in the user’s language rather than just answer in it, noting that current models often think in English and lose nuance and cultural knowledge. It highlights the importance of moving from multilingual answers to in-language reasoning, termed “L2” reasoning, to make reasoning traces accessible to non‑English speakers. The authors propose that in‑language reasoning can be achieved through data mixing built on three pillars, with broader language coverage as the first step.
North 2 introduces a redesigned orchestration system that supports more sophisticated, reliable, multi-step automations and agents across the enterprise. It enhances intelligence by enabling smarter workflows and richer agent chat experiences through memory, libraries, skills, artifacts, and app creation. The platform is model-agnostic, allowing teams to use Cohere’s models or bring their own, while offering full control over AI deployment with flexible cloud, on‑premises, and air‑gapped options. Enterprise governance features include granular cost controls and rate limits.
PwC and Cohere have announced a global alliance to help organizations deploy secure enterprise AI and realize measurable business value. The partnership combines Cohere’s secure agentic AI platform and enterprise models with PwC’s industry, risk, regulatory, technology, and transformation expertise. Together they will identify high‑value use cases, securely connect AI to trusted enterprise data and systems, and establish governance and controls for responsible AI deployment at scale.
Cohere announced the Embed 5 series, designed for leading enterprise search performance. Embed 5 Pro achieved top scores on financial data, PDFs, and visual documents, while Embed 5 Fast offers low latency and cost efficiency. Both models share an embedding space and support multiple languages and long documents.
Cohere Inc. introduced RCP-nDCG@10, a new search evaluation method designed to address limitations of traditional nDCG. It evaluates documents against explicit relevance criteria to measure search system performance more accurately.
Cohere Inc. has started offering the cloud version of the retrieval platform Compass for AI applications utilizing enterprise data as a private beta. This allows developers to use retrieval functions tailored to data and workflows via API and SDK.
Change management traditionally involves new tools and human-led transformations. With AI's integration into organizations, questions arise about whether AI should be treated as software or as an active participant in workflows. This shift prompts a reevaluation of change management strategies.