Google DeepMind announced Gemini 3.6 Flash today, claiming up to 65% lower token costs on long-horizon engineering tasks versus prior Gemini models.

• Pricing lands at $1.50 per million input tokens and $7.50 per million output tokens, with the cost reduction driven by architectural changes that reduce redundant computation on multi-step agent tasks.
• Two companion models ship alongside it: Gemini 3.5 Flash-Lite for cost-sensitive workloads and Gemini 3.5 Flash Cyber, a security-tuned variant.

Salesforce today expanded Agentforce with seven preconfigured, named AI agents covering sales, service, commerce, HR, and back-office functions, plus Slackforce Surfaces to render live Salesforce dashboards inside Slack.

• Hunter, the outbound sales agent, runs on a new long-horizon runtime that can maintain a pipeline for weeks; early customers report strong results, though the agent remains in pilot, and the acquired Fin agent already resolves 79% of Anthropic's support conversations autonomously.
• Slackforce Surfaces lets teams turn Salesforce and connected business data into interactive dashboards, reports, and presentations directly in Slack, fulfilling co-founder Parker Harris's prediction that users may never log in to Salesforce again.

Atlassian today announced new capabilities across Jira, Confluence, and its DX platform designed to govern AI agents across the full software development lifecycle.

• Code Context grounds agents in multi-repository codebases so they can reason across large projects without losing track of dependencies or business rules.
• Agentic loops in Jira automate execution of repetitive engineering tasks while keeping a human-reviewable audit trail, directly addressing compliance concerns enterprises raise about autonomous code changes.

Anthropic today announced Enterprise Frontier Safeguards, which stores monitoring data inside the customer's own cloud account, alongside a public beta of Claude Security scans for Claude Enterprise customers.

• Enterprise Frontier Safeguards was built with more than 100 enterprises and separates custody of monitoring data from Anthropic's infrastructure while keeping automated detection running on Anthropic's side; it is rolling out in phases this fall.
• Claude Security scans, powered by Mythos 5, connect to a GitHub repository, trace data flows across files, and return findings with CWE category, confidence, severity, and suggested patches — available now without a separate model add-on.

Perplexity is using GPT-6 Astra to write communications, change software, and monitor production systems, with human check-ins materially reduced compared to prior models.

• Johnny Ho, Cofounder and CSO, says every improvement in the model's code-writing ability directly compounds Perplexity's own product velocity, making model quality a first-order business variable rather than an infrastructure choice.
• The deployment spans end-to-end production workflows, not sandboxed experiments — a meaningful shift from the pattern of AI tools limited to drafting or code suggestion.