Enterprise agents land as Claude Opus 5 reshapes the cost frontier

Agents enter production

The defining story of the week is OpenAI's launch of Presence, a deployed enterprise agent product that pairs model reasoning with policies, guardrails, and escalation rules for voice and chat agents. This isn't a research preview — Presence is already powering OpenAI's own phone support, resolving 75% of inbound issues without human assistance, and has been proven with BBVA, SoftBank, and IAG. The big idea here is that enterprises no longer need to assemble agent infrastructure from scratch; Presence packages the necessary systems, evaluations, and deployment expertise into a single product.

Meanwhile, OpenAI Codex added ChatGPT Voice — powered by GPT-Live — for talking through work in the ChatGPT desktop app, plus support for multi-folder local projects with a primary folder for Git operations and automatic discovery. On the Claude Code side, Anthropic's v2.1.219 introduced nested subagents up to depth 3 (up from 1), a sandbox network strict allowlist, and the new Claude Opus 5 as the default Opus model with 1M context. The subagent depth increase alone opens up new patterns for hierarchical task decomposition.

The model that changes the math

Anthropic's Claude Opus 5 is the most consequential model release of the week. It matches the frontier intelligence of Claude Fable 5 on coding and knowledge work evaluations like FrontierBench and GDPval-AA, but at half the cost per task. Pricing lands at $5/M input and $25/M output tokens — same as Opus 4.8 — making it the new default on Claude Max and the strongest model on Claude Pro. It's designed for everyday use, not just benchmark chasing, though it still trails Mythos 5 on cybersecurity tasks. For developers, this means the cost-benefit calculus for routing complex reasoning to Opus just shifted dramatically.

Infrastructure and SDK hardening

Across the SDKs, this week was about compatibility and cost control. OpenAI's Go SDK v3.45.0 drops Go 1.22-1.24, requiring Go 1.25+ and upgrading golang.org/x/net to address security findings — a firm nudge toward modern toolchains. The Python SDK v2.47.0 adds experimental HTTPX2 client support, while both the Go SDK v3.46.0 and Node SDK v6.49.0 add spend limit admin APIs, letting platform teams enforce cost boundaries programmatically. The Node SDK also picked up Zod helpers and standard schema support.

Claude Code received several quality-of-life fixes: v2.1.216 fixed a quadratic slowdown in long sessions and OAuth token expiry race, v2.1.217 addressed an MCP memory leak and session isolation, and v2.1.218 fixed Windows Unicode path corruption and a data-loss arrow key bug. Together, these patches make the tool more reliable for extended sessions.

Beyond the code

Anthropic published the research agenda for its $200 million Economic Futures Research Fund, focusing on interventions to prepare society for AI's economic impacts. OpenAI launched Health in ChatGPT for eligible U.S. users, connecting medical records and Apple Health for personalized health insights — a meaningful expansion of ChatGPT's utility beyond coding and productivity into regulated domains.