Three stories from today: Claude Opus 4.7, enterprise agent frameworks, and defensive cybersecurity

Claude Opus 4.7 Ships With Task Budgets for Agentic Loops

Anthropic released Claude Opus 4.7 with three new capabilities that matter for agent builders. High-resolution image support at 2576px (3.75MP), up from 1568px, which improves computer use and document understanding. A self-correcting planning loop that catches logical faults before execution, so you spend less time debugging hallucinated tool calls. And a task_budget beta header that gives the model an advisory token allowance for a full agentic loop, with a live countdown so it prioritizes and finishes gracefully within the cap. [1]

Pricing stays the same at $5 in, $25 out per million tokens. Context window is still 1M with 128k output. [1]

My Take

Runaway token spend is the main reason I have hesitated to let agents run unsupervised. A task budget fixes that. You set a target, minimum 20k tokens, and the model self-moderates across the entire loop, including thinking, tool calls, and output. This is different from max_tokens, which is a hard per-request ceiling the model never sees. Task budget is advisory and scope-aware. For anyone shipping autonomous agents, that is an upgrade to how agents behave, not just what they score on benchmarks.

One operational note: temperature, top_p, and top_k are gone. Any non-default value now returns a 400 error. [1] Reliability has to come from prompting, not sampling parameters.

Claude Design Launches for Pro and Team Subscribers

Anthropic Labs launched Claude Design on April 17, 2026, the same day Opus 4.7 went live. It is a visual workspace for creating designs, prototypes, slides, and one-pagers through conversation with Claude. You start with a text prompt and optional reference materials like a codebase, images, or documents. Claude generates a first draft. Refinement happens through an inline comment box, custom sliders for typography and color, or direct edits on the canvas. Exports cover standalone HTML, PDF, PPTX, Canva files, and internal share URLs. It is available to Claude Pro, Max, Team, and Enterprise subscribers, and it runs exclusively on Opus 4.7. [2] [3]

My Take

The most interesting flow is "upload your codebase, get a design system." Claude Design digests existing code and generates a cohesive design system with colors, typography, and components automatically. [3] For solo founders keeping a product visually consistent without a design partner, that is the missing piece. Export-to-HTML is the other practical win. The generated UI drops straight into your codebase without a translation layer. Pitch deck as PPTX from the same prompt flow removes another context switch between tools.

This puts Claude directly in the same lane as v0, Lovable, and Canva's AI tools. The difference is where it hooks in. Claude Design lives next to Claude Code and the rest of the Claude surface, so the transition from "design this" to "build this" happens in one workspace rather than across four.

OpenAI Opens a Verified Lane for Defensive Security Work

OpenAI released GPT-5.4-Cyber, a specialized defensive security model with binary reverse engineering baked in. The model analyzes compiled software for malware and vulnerabilities without needing original source code. Access is gated through the Trusted Access for Cyber program, which requires identity verification at chatgpt.com/cyber for individuals or enterprise channels for teams. [4]

My Take

If you build security tooling, refusal behavior has been the bottleneck. Ask a general-purpose model to analyze a binary and it often refuses, even for clearly legitimate defensive work. GPT-5.4-Cyber lowers that boundary once you pass verification. OpenAI cites over 3,000 critical and high-severity vulnerabilities patched through its Codex Security research preview earlier this year. [4] This model is the productized version of that capability.

For solo founders in the security space, this is real access that was previously reserved for large vendors with direct OpenAI relationships. The verification step is a small friction. The capability is not.

The pattern across all three releases is the same: capabilities that used to require a team are now one API call and a verification form away. That is the leverage a solo founder can build on.

Sources

${sourcesHtml}

Originally published on chento.io