In this newsletter:

• Claude Opus 4.7 checks its own work before responding
• Perplexity personal computer automates your to-do list
• Canva AI assistant creates designs using tools and layers

Plus, you’ll find new AI tools and this week’s top AI news headlines!

​

​

⚡ Anthropic's Claude Opus 4.7 Checks Its Own Work Before Responding

Anthropic released Claude Opus 4.7, its most powerful publicly available model, outperforming OpenAI's GPT-5.4 and Google's Gemini 3.1 Pro on key benchmarks, including agentic coding, tool use, computer use, and financial analysis.

What's new:

• Available today across Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry
• API pricing unchanged at $5/$25 per million tokens
• Processes images up to 2,576 pixels on the longest edge (roughly 3.75 megapixels), a threefold increase from previous versions
• New "effort" parameter with xhigh (extra high) setting between high and max for granular control over depth of reasoning
• Task budgets in public beta let developers set a hard ceiling on token spend for autonomous agents
• Updated tokenizer improves text processing efficiency but can increase token count of certain inputs by 1.0–1.35x
• Claude Code gets a new /ultrareview command that flags subtle design flaws and logic gaps like a senior human reviewer
• Auto mode extended to Max plan users for autonomous decisions without constant permission prompts

Performance highlights:

• GDPVal-AA knowledge work: Elo score of 1753, beating GPT-5.4 (1674) and Gemini 3.1 Pro (1314)
• SWE-bench Pro agentic coding: 64.3% task resolution vs 53.4% for Opus 4.6
• GPQA Diamond graduate-level reasoning: 94.2%
• arXiv visual reasoning with tools: 91.0% vs 84.7% for Opus 4.6
• XBOW visual-acuity tests: 98.5% vs 54.5% previously
• BigLaw Bench: 90.9%
• CyberGym vulnerability reproduction: 73.1%
• GPT-5.4 still leads in agentic search (89.3% vs 79.3%), multilingual Q&A, and raw terminal-based coding

How it works differently:

• Devises own verification steps before reporting task complete
• Internal test example: built a Rust-based text-to-speech engine from scratch, then independently fed generated audio through a separate speech recognizer to verify output against a Python reference
• Follows instructions literally, requiring legacy prompt libraries to be re-tuned
• Older models "read between the lines," Opus 4.7 executes the exact text provided

Enterprise feedback:

• Intuit VP of Technology: the ability to "catch its own logical faults during planning phase" is a game-changer for velocity
• Replit President: higher quality at lower cost for log analysis and bug hunting, "feels like a better coworker."
• Cognition CEO: can work coherently "for hours" and pushes through difficult problems that cause models to stall
• Notion AI Lead: 14% improvement in multi-step workflows, 66% reduction in tool-calling errors, feels like "true teammate"
• Factory Droids: carries work through to validation steps rather than "stopping halfway."
• Harvey, Head of Applied Research: "noticeably smarter handling of ambiguous document editing tasks."

Cybersecurity controls:

• Systems designed to detect and block requests suggesting high-risk cyberattacks like automated vulnerability exploitation
• New Cyber Verification Program allows vulnerability researchers, penetration testers, and red-teamers to apply for access for defensive purposes
• More powerful Mythos model remains restricted to a small number of external enterprise partners for cybersecurity testing through Project Glasswing

Why it matters:

The razor-thin margins between Opus 4.7, GPT-5.4, and Gemini 3.1 Pro signal that the frontier model race has plateaued, with improvements measured in single percentage points, forcing companies to compete on reliability rather than raw intelligence.

Self-verification addresses the hallucination problem that kept enterprises from trusting AI for autonomous work, shifting AI from a creative assistant to a tool that can run unsupervised.