OpenAI just pushed Codex into a very different category. It is no longer just a coding agent next to your terminal. With background computer use, an in-app browser, memory, automations, image generation, and a broader plugin layer, Codex is being positioned as the place where work gets assigned, tracked, and carried forward across tools.

That is the real shift. This is no longer just “AI that helps you code.” It is “AI that helps run the workflow.” Once an agent can use apps on your Mac, remember how you work, continue long-running tasks, and take on repeatable work, the contest is no longer only about which model writes better code. It becomes a fight over runtime, context, and the daily working surface itself.

→ So what: The builder AI race is moving from model quality to operating-layer ownership.

Anthropic has officially released Claude Opus 4.7. The core story is not just another benchmark jump. The more important signal is that Anthropic is pushing toward a model that can handle harder, longer tasks with less supervision. The company is emphasizing stronger instruction following, self-verification, software engineering performance, and multimodal understanding, while keeping pricing flat versus 4.6.

That makes the contrast with OpenAI clearer. OpenAI is expanding Codex into a broader work surface. Anthropic is still sharpening the brain inside the agent: making the model more reliable, more self-checking, and more capable of staying coherent through long runs without falling apart.

→ So what: Builder tools are starting to split in two directions: who owns the workflow, and who builds the most trustworthy model to hand work to.

OpenAI has also introduced GPT-Rosalind for life sciences research, aimed at workflows like target discovery, literature synthesis, genomics interpretation, and hypothesis generation. This is not a broad consumer launch. It is a domain-specific model line for eligible enterprise teams in a high-value, high-stakes industry.

That matters because it suggests frontier AI is starting to split vertically. The flagship general-purpose model still matters, but once you get into areas like cyber or biology, the real edge may come from models tuned for a specific type of work, user, and decision environment. GPT-Rosalind, landing just days after GPT-5.4-Cyber, is a strong signal that OpenAI is testing exactly that strategy.

→ So what: “One model for everything” may not be the end state. AI is moving toward specialized models for the most expensive workflows.

Perplexity just launched Personal Computer, integrating its Mac app with local files, native apps, and the browser. The signal is clean: answer engines no longer want to stop at giving answers. They want to become the layer that orchestrates work.

Canva’s AI assistant can now pull context from Gmail, Slack, Drive, Calendar, and Zoom, call tools, and even schedule recurring tasks. Design software no longer wants to be just an app with AI inside it. It wants to become the place where AI coordinates the workflow itself.

Hightouch says it added $70 million in ARR in 20 months after launching its AI product, bringing total ARR to $100 million. That is a useful commercial signal: AI workflow tooling is not just generating attention. It is generating real revenue inside business-critical verticals.

Jack Dorsey only posted a short line, “goose is all you need,” but the broader signal is worth tracking. Open-source, local-first agent stacks are increasingly being seen as a credible alternative to closed AI work surfaces.

If you use Claude Code, Anthropic’s new best-practices guide is worth reading not because it markets Opus 4.7, but because it spells out a very practical change in how the model works best. The core idea is simple: Opus 4.7 performs better when you treat it like a capable engineer you can delegate to, not a pair programmer you need to steer line by line.

The most effective setup is to batch the task in the first turn. Give it the full brief up front: intent, constraints, acceptance criteria, and relevant file paths. Anthropic also recommends reducing unnecessary user turns, because each additional turn adds reasoning overhead. For long, clearly bounded tasks, auto mode is the cleaner way to reduce cycle time instead of constantly checking in.

The effort setting matters too. The key takeaway is that xhigh is now the default and recommended setting for most agentic coding work. high is a better tradeoff if you want lower cost or more concurrent sessions. max should be used deliberately for genuinely difficult work, because it is more prone to overthinking and token blowups.

If you are moving from Opus 4.6 to 4.7, the practical lesson is this: do not just port over your old habits. Write prompts like you are assigning work to a senior engineer, reduce the number of turns, and choose effort based on the job instead of leaving it on autopilot.

Real impact: A small change in how you hand off tasks can improve output quality while cutting a surprising amount of wasted token spend.

Most AI newsletters are written for engineers. This one isn't.

The AI Report is read by 400,000+ executives, operators, and business leaders who want to know what's happening in AI — without wading through code, jargon, or hype.

Every weekday, we break down the AI stories that matter to your business: what's being deployed, what's actually working, and what it means for your team.

Free. 5 minutes. Straight to the point.

Join 400,000+ business leaders staying ahead of AI — without the technical overwhelm.

Read it free today

Ai2 makes a timely point: everyone is talking about science agents, but most of the claims are outrunning the benchmarks and evidence. The gap between a model that “knows” science and an agent that can actually do science in an open, multi-step environment is still large.

→ Implication: As labs roll out domain-specific models like GPT-Rosalind, the key question is not just how capable they look, but who has a credible way to measure them

IBM’s VAKRA focuses on evaluating agent reasoning across multiple data sources and APIs, with an emphasis on full execution traces rather than just the final output. That matters because it evaluates the agent as a working system, not just as a chatbot with good answers.

→ Implication: If 2025 was the year of agent demos, 2026 may be the year benchmarks and reliability start deciding which agents are actually production-ready