# Agent Platform Research Briefing — July 25, 2026
Welcome to the agent platform research briefing for Saturday, July 25th, 2026.
Anthropic launched Claude Opus 5 on July 24th, the fourth model in the Claude 5 family, positioned as a cost-efficient alternative to the restricted Fable 5. API pricing is $5 per million input tokens and $25 per million output — identical to Opus 4.8 and exactly half the price of Fable 5 at ten dollars in and fifty out.
The benchmarks are striking. On Frontier-Bench v0.1, Opus 5 achieves 43.3%, beating Fable 5's 33.7% at half the cost. On ARC-AGI-3, it hits 30.2%, while the next-best model manages 7.8%. Anthropic published full cost-versus-score effort curves, showing five levels from minimum to maximum reasoning depth. Developers can now tune per-task: low effort for fast cheap responses, high for complex reasoning. The model has a 1 million token context window and up to 128,000 tokens of synchronous output. Importantly, Opus 5 remains behind Mythos 5 on cybersecurity tasks — a deliberate differential design choice.
Alongside the Opus 5 launch, Anthropic upgraded Claude Voice Mode on July 24th. Paid subscribers can now switch between Haiku, Sonnet, and Opus during voice conversations. Previously voice mode was limited to Haiku only — the fastest but simplest model. Voice mode now defaults to whichever model you were last using for text, and inherits the full conversation context from your text session. Tech coverage frames Anthropic as betting on voice-driven agent action rather than conversational naturalness — while ChatGPT Voice focuses on natural conversation, Claude Voice now emphasizes action: it can act in Gmail, Slack, and your calendar. This positions Claude Voice as an agent interface, not just a chat interface.
OpenAI pushed ChatGPT Voice into the desktop app on July 23rd for Mac and Windows, available to Plus, Pro, Business, Edu, and Enterprise plans. Powered by GPT-Live, the full-duplex voice model that speaks and listens simultaneously, it can now control your computer and coordinate multiple agents running in ChatGPT Work or Codex by spoken command. You can start tasks, check progress, interrupt or redirect work, and receive updates while tasks continue in the background. This is a significant evolution from GPT-Live's initial launch on July 8th — it's not just voice chat anymore, it's voice as the orchestrator for agentic work. VentureBeat calls it "hands-free agentic coding."
On July 23rd, OpenAI rolled out ChatGPT Health to all U.S. users over 18 across all plans. Users can securely connect Apple Health data and medical records from Epic and other EHR providers, enabling contextual health insights, trend tracking, and personalized health conversations directly in ChatGPT. But privacy experts are raising alarm: once health records are connected to ChatGPT, they permanently lose HIPAA protection. OpenAI positions this as voluntary user choice, but the tradeoff means sensitive medical data becomes subject to OpenAI's privacy policy rather than healthcare regulations. The feature also faces legal pressure from a Florida challenge.
Two significant MCP vulnerabilities emerged this week. First, Manifold Security disclosed a critical flaw in Microsoft's Azure DevOps MCP server where hidden HTML comments in pull request descriptions can hijack AI code review agents and exfiltrate data from projects the attacker cannot directly access. There is no fixed release yet — the latest version, 2.8.0, remains vulnerable. Second, AWS Labs published CVE-2026-16584 for the AWS API MCP Server — a high-severity security policy bypass triggered by startup initialization failure. These join a growing pattern of MCP server vulnerabilities affecting major tech companies' official implementations.
That's the briefing for today. Have a good one.