Volume 343, Issue 3 · 8 articles from 2 issues
Cohere research finds a strong lead agent with cheap workers nears frontier coding results at 70% lower cost.
Cohere research finds well-organized teams of coding agents outperform single agents by up to about 10 points on ProgramBench, but at three to seven times the cost.
Where agents are allocated matters more than how many are added, as delegating only exploration and planning matched broader delegation at half the cost.
A Sol lead coordinating 16 cheaper Luna planners reached 61.8 percent reward versus 63.2 percent for an all-Sol team, cutting cost from $510 to $155.
Cohere explains why enterprise AI ROI is hard to calculate and how to forecast and measure returns.
Cohere defines AI ROI as financial return relative to the total cost of deploying and operating an AI system, realized through cost savings, added profit or both.
It notes that AI costs vary with design, deployment and usage, and that operational gains such as speed, output and quality do not always translate cleanly into financial value.
The post explains these difficulties and how enterprises can account for them when forecasting and measuring returns.
Cohere explains agent sprawl in enterprises and proposes letting variant agents compete instead of restricting them.
As enterprise agent platforms make it easy for employees to build agents, organizations struggle to track what has been built, reuse existing agents and govern access to data and systems, an unmanageable result Cohere calls agent sprawl.
The post argues proliferation is not inherently bad, reviews how enterprises typically manage it, and proposes an alternative that lets variant agents compete.
It points to Cohere's North 2 platform for building and sharing reusable agents.
Cohere argues utilities must weigh where AI runs, what it can access and what it can do, not just capability.
Cohere says energy and utility companies use AI to strengthen cyber and infrastructure defenses, surface engineering knowledge and cut time spent gathering information.
For critical infrastructure operators, model capability is only part of the decision, since where AI runs, what information it can access, what actions it can take and how it is monitored determine which workloads can responsibly go into production.
The post urges utilities to apply the same discipline they use for other consequential technologies.
Cohere releases Rerank 5 in Pro and Fast tiers, adding Context Focus passage extraction and calibrated relevance scores.
Cohere has released Rerank 5, its most advanced enterprise retrieval models, in Pro and Fast tiers.
Cohere says Rerank 5 Pro leads major public benchmarks, especially on reasoning-intensive search and visually rich documents, while Rerank 5 Fast offers near-leading performance at lower cost.
A new Context Focus feature returns the most relevant passages within each document to cut LLM token costs, and calibrated scores let teams set a single threshold.
The models are available via the Cohere API, Model Vault, Microsoft Foundry, Amazon SageMaker, private VPC or on-premises deployments.
Cohere launches a public beta of Confidential Computing on Model Vault, protecting prompts and responses while inference runs.
Cohere announced the public beta of Confidential Computing on Cohere Model Vault, its dedicated managed inference service.
The feature extends protection beyond data at rest and in transit to prompts, documents, responses and model execution during inference.
Request payloads stay encrypted end-to-end and can only be decrypted inside an attested confidential environment that customers can verify cryptographically.
Cohere runs the service on Microsoft Azure using NVIDIA H100 GPU confidential computing, targeting workloads such as financial, clinical and privileged data.
Cohere explains the architecture behind confidential inference in Model Vault, including isolation and attestation.
Following its launch of Confidential Computing on Model Vault, Cohere goes deeper into how the system works.
The goal is to reduce the infrastructure customers must trust with plaintext data and give them cryptographic evidence before sensitive data is released.
Cohere says this requires protecting the computation, the request path and the mechanism customers use to decide whether to trust the destination.
The post explains how hardware isolation and attestation underpin the dedicated model deployment.
Cohere researchers introduce CALIBER, a method that trains reasoning models to estimate and update their chance of being right.
Cohere notes that producing a detailed reasoning trace does not guarantee a model can reliably assess its own uncertainty.
A useful confidence estimate should change as reasoning reveals the answer is likely right or wrong.
Cohere's CALIBER method trains a reasoning model to estimate its chance of success before it begins and then assess its specific answer afterward.
On Cohere's math benchmark, a calibrated 7B model substantially reduced calibration error.