Cohere Command A+ hits the API: 192K context at $0.30/M input

The flagship does native tool calls and JSON at budget prices, while Safari 27 hands agents a live browser and Gemini keeps API keys out of the sandbox.

Nowline SEP 24 12:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Command A+ is live on the API

    Cohere's new flagship pairs a 192K context with native tool calling, strict JSON schemas, optional reasoning, and text-plus-image input, and it's priced at $0.30/M in and $1.50/M out with up to 64K output. That undercuts most mid-tier frontier models, making it a cheap, agent-ready default you can call today on Cohere or OpenRouter.

  • Safari 27 gives your agent a live browser

    Safari 27 ships a native, local MCP server: connect Claude Code, Codex, or any MCP client and it can read the DOM, take screenshots, watch network requests, tail the console, and audit accessibility, with no install and no outbound calls. Enable it under Safari > Settings > Developer, though reports note Apple gives enterprises no way to switch it off.

  • Gemini agents stop leaking your keys

    Google's managed-agent update adds a Credentials API so agents can call GitHub, Slack, and other services without the token ever entering the sandbox, plus a Files API to move data in and out. The new antigravity-preview-09-2026 harness brings Antigravity's tools into AI Studio on Gemini 3.8 Flash, with a reported ~22% efficiency gain.

  • AutoGen moves to maintenance mode

    Microsoft has retired AutoGen in favor of its unified Agent Framework, freezing the popular multi-agent library at maintenance-only while new features land in the successor. If you ship on AutoGen, start planning a move — Microsoft has published a from-AutoGen migration guide.

  • OpenAI's Agents API opens to everyone

    OpenAI's Codex-harness Agents API is now in public beta for all developers, with durable sessions that auto-compact past the context limit, tool search, parallel tool calls, MCP and custom functions, and subagents you can fan out in parallel. Run it on OpenAI's sandboxes or self-host through Cloudflare, Vercel, E2B, or Modal, with no fees beyond tokens and tool costs.