{"company":{"name":"Cohere","slug":"cohere","website":"https://cohere.com","category":"ai-ml"},"question":"What has Cohere shipped recently?","answer":"In the last 30 days, Cohere shipped 24 tracked updates. The most recent was \"Cohere Labs - Catalyst Grants\" on 2026-11-20.","window":{"days":30,"updateCount":24,"returned":24},"generatedAt":"2026-11-22T11:44:30.342Z","updates":[{"title":"Cohere Labs - Catalyst Grants","summary":"Cohere Labs' Catalyst Grants give academic, civic and public-impact organizations access to Cohere models via the API for public-benefit or open-science work, focused on education, healthcare, culture and sustainability. The program is currently closed, with no next deadline listed, and visitors can sign up to be notified when it reopens. Grantees must disclose AI safety vulnerabilities and share research outputs at least 4 weeks before publication in that category, and credits are tracked in the Cohere dashboard.","date":"2026-11-20","dateIsEstimated":true,"signalType":"partnership","signalTypeLabel":"Partnership","sourceUrl":"https://cohere.com/research/grants","publisher":"cohere.com"},{"title":"Transformez votre expertise encompétences IA réutilisables","summary":"Cohere's North platform now offers Skills, which package instructions, templates, reference files and scripts so a team's working methods apply consistently across chats, agents and automations. Skills can be imported in the open Agent Skills format, shared through agents, deployed by administrators, and run in a sandbox inside the customer's own infrastructure, including on-premises or air-gapped environments.","date":"2026-11-18","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://cohere.com/fr/north/skills","publisher":"cohere.com"},{"title":"Le savoir de votre entreprise,une référence fiable pour vos agents","summary":"Cohere's North platform now offers Libraries, which centralize company files from tools such as SharePoint, OneDrive, Google Drive, email and direct uploads into a single source that multiple agents, conversations and automations can use. The feature includes automatic syncing, source-restricted answers, Compass-powered search, preserved source permissions, and deployment inside a customer's VPC, on-premises or isolated environment. Administrators control who can create and share agents, and every agent version is archived with testing history for risk and compliance teams.","date":"2026-11-18","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://cohere.com/fr/north/libraries","publisher":"cohere.com"},{"title":"Orchestrez des workflows agentiques avec un contrôle total","summary":"Cohere's French product page for North Automations describes a no-code way to build and deploy agentic workflows by describing goals in plain language, connecting tools via first-party integrations or MCP, and scheduling runs or triggering them through an API. The page highlights plan mode, model selection per step, versioning, sandbox testing, and governance features including approval checkpoints, usage monitoring, and token cost tracking. It presents these as part of the broader North agentic platform that can be deployed on-premises or in a cloud VPC. The page is undated and does not state a launch date.","date":"2026-11-18","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://cohere.com/fr/north/automations","publisher":"cohere.com"},{"title":"Unternehmenswissen,auf das Ihre KI-Agenten vertrauen können","summary":"Cohere's North platform now offers Libraries, a feature that lets organizations build a centrally managed, continuously synced knowledge base from sources such as SharePoint, OneDrive, Google Drive, email and direct uploads. Libraries can be attached to chats, AI agents and automation steps, answers are restricted to curated sources, and access permissions from source systems are preserved. Deployment is available in a VPC, on-premises or air-gapped environments. The page also states that admins control which users and groups can build and share agents, and that version documentation is available for risk and compliance teams.","date":"2026-11-18","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://cohere.com/de/north/libraries","publisher":"cohere.com"},{"title":"Steuern Sie agentische Workflows mit voller Kontrolle","summary":"Cohere's product page for North Automations describes a no-code way to build and schedule agentic workflows described in plain language, connected to first-party integrations or custom tools via MCP, and triggered through an API. It highlights a plan mode, model selection per step, version control, sandbox testing, built-in approvals, usage analytics, and token-level cost monitoring. Governance controls are positioned for admins managing automations across teams. The page is marketing copy and gives no launch date or availability details.","date":"2026-11-18","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://cohere.com/de/north/automations","publisher":"cohere.com"},{"title":"Shaping the frontier ofML research","summary":"Cohere's Cohere Labs research page presents the lab's mission of advancing fundamental ML research while broadening who shapes it. It highlights ongoing work including the Future(s) of Work research program, the Aya multilingual model initiative (described as covering 101 languages, including more than 50 previously underserved ones), the Tiny Aya model family, and the Global MMLU benchmark. It also lists recent papers dated October 2026 on topics such as multilingual safety, speculative decoding for sparse mixture-of-experts models, and agentic AI. Programs described on the page include a Research Scholars program, an Open Science Community of 4,500 members from 150 countries, and Catalyst Grants that give academic and civic partners free access to the Cohere API. The page is primarily descriptive and does not announce a new product, pricing change, or organizational event.","date":"2026-11-17","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://cohere.com/research","publisher":"cohere.com"},{"title":"Explore the Agentic Task Ecosystem","summary":"Cohere published a research page for its Agentic Task Ecosystem (ATE), a dataset of 694,000+ executable tools drawn from 123,000 public MCP servers across 7 directories. The page frames it as a way to see how automation is affecting work and what parts of jobs may change. The page links to a companion blog post and the dataset itself.","date":"2026-11-17","dateIsEstimated":true,"signalType":"content_marketing","signalTypeLabel":"Content","sourceUrl":"https://cohere.com/research/agentic-task-ecosystem","publisher":"cohere.com"},{"title":"Human-verified multilingual evaluation at scale","summary":"Cohere's Global MMLU research page highlights a new release, Global MMLU Lite V3, which adds six languages (Oriya, Hungarian, Tajik, Slovak, Czech and Italian) for 23 languages total and uses a 6,000-sample human-verified dataset. The full Global MMLU benchmark covers 42 languages with 589,764 samples, and the page reports it is available on Kaggle, was featured in the 2024 Stanford HAI Index, and was presented at the 2024 ACL main track.","date":"2026-11-17","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://cohere.com/research/globalmmlu","publisher":"cohere.com"},{"title":"Deliver the AI advantage with Cohere","summary":"Cohere's partner program page outlines three partnership paths (strategic, consulting and implementation, and cloud and distribution, plus technology and community/research tracks) and three tiers: Collaborator, Accelerator, and Pioneer. It highlights featured partners including Fujitsu, LG CNS, Dell, AMD, Bell Canada, Oracle and McKinsey. The page also notes a Download guide and Apply now call to action, but does not state a launch date.","date":"2026-11-17","dateIsEstimated":true,"signalType":"partnership","signalTypeLabel":"Partnership","sourceUrl":"https://cohere.com/partners","publisher":"cohere.com"},{"title":"On-Demand Events","summary":"Cohere's on-demand events page lists recorded sessions from industry events. The featured video is a RAISE Summit in Paris panel with Cohere and Adobe on transforming enterprise search with AI. The panel covers the Adobe-Cohere partnership, multilingual AI applications, and how better search gives AI agents the context they need to handle complex, ROI-driving tasks. The video runs about 18 minutes. Two other recordings, \"Architects of AI: pivotal moments\" (about 46 minutes) and \"Enterprise workflow automation with genAI\" (about 44 minutes), are listed with titles only and no descriptions. The page contains no new product, pricing, or dated announcement. Its most material detail is the reference to an existing Adobe-Cohere partnership, which the panel explains but the page neither dates nor details.","date":"2026-11-17","dateIsEstimated":true,"signalType":"event","signalTypeLabel":"Event","sourceUrl":"https://events.cohere.com/on-demand/","publisher":"events.cohere.com"},{"title":"Schnell entwickeln. Die Kontrolle behalten.","summary":"Cohere's Model Vault page describes a fully managed, dedicated SaaS inference platform for Cohere models and leading open-weight models from third parties. Cohere states it handles deployment, maintenance, updates and scaling, offers self-service setup within minutes, real-time monitoring of request rates, latency and token usage, and full network isolation without rate limits from shared resources. Data is said to be encrypted in the chosen region, not used for training, and covered by a Zero Data Retention policy. An optional Confidential Computing mode (Model Vault Encrypted) is described, and Cohere positions Model Vault as the inference layer for North deployments, with the North control plane running on CPUs only.","date":"2026-11-17","dateIsEstimated":true,"signalType":"product_launch","signalTypeLabel":"Launch","sourceUrl":"https://cohere.com/de/model-vault","publisher":"cohere.com"},{"title":"THE NARRATIVE GAP: Can LLMs Help Us Navigate Diverse Narratives Across Languages?","summary":"Cohere published a research paper introducing NarrativeBench, a benchmark evaluating how well LLMs help users seek information and navigate diverse narratives across languages, with the aim of reducing echo chambers. The authors listed are Nikhil Sharma, Kelly Marchisio, Kenton Murray, and Ziang Xiao. The page also lists related Cohere research on low-resource Southeast Asian language agent evaluation, African-language safety refusal, and cultural riddles, but the content provided does not include the abstract text or results.","date":"2026-11-09","dateIsEstimated":false,"signalType":"content_marketing","signalTypeLabel":"Content","sourceUrl":"https://cohere.com/research/papers/the-narrative-gap-can-llms-help-us-navigate-diverse-narratives-across-languages-2026-11-09","publisher":"cohere.com"},{"title":"Turn possibility into production","summary":"Cohere introduced Rerank 5 with Context Focus for sharper search and fewer tokens, alongside the new Embed 5 family of frontier embedding models. These updates enhance retrieval optimization, semantic representation, and enterprise-grade AI capabilities within its North platform.","date":"2026-11-08","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://cohere.com/products","publisher":"cohere.com"},{"title":"Cohere models now available on OpenRouter with Embed 5 and Rerank 5 support","summary":"Cohere announced its Command, Embed, Rerank, and North models are accessible via OpenRouter, including the new Embed 5 and Rerank 5 variants. Users can leverage a single API key for Cohere and other providers, with Cohere serving as the backend for its models on OpenRouter.","date":"2026-11-04","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://docs.cohere.com/v2/docs/cohere-on-openrouter","publisher":"docs.cohere.com"},{"title":"Cohere Embed 5 and Rerank 5 models now available on OpenRouter","summary":"Cohere’s Embed 5 and Rerank 5 models (Pro and Fast variants) are now accessible via OpenRouter’s Embeddings and Rerank APIs. Developers using OpenRouter can integrate Cohere’s models for search and RAG pipelines using their existing OpenRouter API key.","date":"2026-11-04","dateIsEstimated":true,"signalType":"integration","signalTypeLabel":"Integration","sourceUrl":"https://docs.cohere.com/v2/changelog/embed-rerank-v5-openrouter","publisher":"docs.cohere.com"},{"title":"Cohere adds Encrypted Vaults with end-to-end confidential computing to Model Vault","summary":"Cohere introduced Encrypted Vaults, adding end-to-end encryption and confidential computing to its Model Vault service. Prompts, responses, and intermediate data remain encrypted even from Cohere and cloud operators, running inside hardware-backed TEEs with verifiable attestation. The feature supports models like Gemma, GPT-OSS, Phi, and Qwen, and is available in the U.S. with EU and Canada rollouts planned.","date":"2026-11-03","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://docs.cohere.com/v2/docs/model-vault/encrypted","publisher":"docs.cohere.com"},{"title":"Cohere Model Vault Encrypted expands supported models and confidential GPUs","summary":"Cohere’s Model Vault Encrypted now supports additional Cohere models (Embed, Rerank, Parse, Compass, North Mini Code, Transcribe, Customized) and open models (Gemma, GPT-OSS, Phi, Qwen, Ministral, Devstral) running on NVIDIA H100 80GB GPUs in confidential-computing mode. Regional expansion to EU and Canada is planned, with pricing based on GPU-hour usage.","date":"2026-11-03","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://docs.cohere.com/v2/docs/model-vault/encrypted/supported-models","publisher":"docs.cohere.com"},{"title":"Cohere Model Vault adds open models and regional serving with Zero Data Retention","summary":"Cohere’s Model Vault now runs leading third-party open models (e.g., DeepSeek, Mistral AI, GPT-OSS) alongside its own models, with Zero Data Retention by default and regional serving in the US, EU, and Canada. Encrypted Vault adds confidential computing for select models like Gemma and Phi.","date":"2026-11-03","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://docs.cohere.com/v2/changelog/model-vault-open-models","publisher":"docs.cohere.com"},{"title":"Release Recap: Scaling Inference-Time Compute with Teams of Coding Agents","summary":"Cohere published research showing that strategic delegation in multi-agent systems can reduce costs by 70% while maintaining near-optimal performance, using a 17-agent heterogeneous team that matches a 5-agent homogeneous team’s reward within 1.4 points. The findings provide guidelines for optimizing agent budgets under task, cost, and latency constraints.","date":"2026-10-30","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://cohere.com/events/release-recap-scaling-inference-time-compute-with-teams-of-coding-agents-2026-10-30t16-00-00-000z","publisher":"cohere.com"},{"title":"Why Speculative Decoding Works Better Than Predicted on Sparse MoE Models","summary":"Cohere published a research paper, dated October 28, 2026, examining why speculative decoding, a technique that speeds up autoregressive inference by verifying multiple draft tokens in a single forward pass, performs better than predicted on sparse mixture-of-experts (MoE) models. The page states that prior work predicted a concave relationship between speedup and batch size, with gains peaking at intermediate batch sizes due to expert saturation. The paper's findings are not shown in the captured text, which contains only the introduction, the author list (Ekagra Ranjan, Komal K. Teru, Bharat Venkitesh, Acyr Locatelli), and links to three related MoE training papers. The page does not state a product, availability, or commercial change.","date":"2026-10-28","dateIsEstimated":false,"signalType":"content_marketing","signalTypeLabel":"Content","sourceUrl":"https://cohere.com/research/papers/why-speculative-decoding-works-better-than-predicted-on-sparse-moe-models-2026-10-28","publisher":"cohere.com"},{"title":"Research papers","summary":"Cohere's research papers page lists its recent publications, running from May to late October 2026. Recent titles cover speculative decoding on sparse MoE models, safety and refusal for African languages, multilingual speech understanding, agentic task evaluation, and a North Small Translate model branded as Cohere CAT+. The page itself is a listing with no dated announcement beyond the individual paper dates.","date":"2026-10-28","dateIsEstimated":false,"signalType":"content_marketing","signalTypeLabel":"Content","sourceUrl":"https://cohere.com/research/papers","publisher":"cohere.com"},{"title":"Ada finds the right ingredients for faster, more efficient AI agents","summary":"Cohere published a customer story on Ada, a customer service AI company that evaluated four embedding approaches across 15,000 queries. Ada uses Cohere Model Vault for offline knowledge-base indexing and self-hosts Cohere Embed on Amazon SageMaker AI for online retrieval. The article reports 90.7% top-10 recall, 77 ms median retrieval latency, about 2.7x better throughput-per-dollar on Model Vault than Ada's prior self-hosted setup, and a 4x vector storage reduction from Embed's native compression with no measurable retrieval degradation.","date":"2026-10-27","dateIsEstimated":false,"signalType":"case_study","signalTypeLabel":"Case Study","sourceUrl":"https://cohere.com/customer-stories/ada","publisher":"cohere.com"},{"title":"AFRIGUARD: Safety and Robust Refusal For African Languages","summary":"Cohere's research section published AFRIGUARD, a paper on safety and robust refusal behavior in large language models for African languages. The paper argues that safety risks are expressed and interpreted differently across linguistic and cultural contexts, making appropriate responses harder to ensure. The provided page contains only the abstract introduction, author list, and links to related research, so the full findings and methods are not visible.","date":"2026-10-26","dateIsEstimated":false,"signalType":"content_marketing","signalTypeLabel":"Content","sourceUrl":"https://cohere.com/research/papers/afriguard-safety-and-robust-refusal-for-african-languages-2026-10-26","publisher":"cohere.com"}],"attribution":{"source":"Spyingbee","url":"https://spyingbee.com/updates/cohere","citation":"Spyingbee, \"Cohere updates\", https://spyingbee.com/updates/cohere (retrieved 2026-11-22)"}}