It’s Topical
← The AI Builder Roundup archive
The AI Builder Roundup

GPT-6 gets interactive as agent costs fall

Chat interfaces just got more hands-on, while agent economics and control moved too. Here’s what changed for interactive answers, high-volume workflows, retrieval, containment, and AI-media provenance.

In today’s edition

  • GPT-6 adds interactive answers inside ChatGPT.
  • Claude Haiku 5.5 cuts small-model costs.
  • Perplexity releases multimodal PDF retrieval models.
  • Windows makes agent containment generally available.
  • Google opens SynthID checks to everyone.

GPT-6 brings interactive UI into ChatGPT

ChatGPT can now answer with interactive components, not just text. GPT-6 can combine visuals, charts, buttons, forms, and task-specific tools inside a conversation, with paid tiers receiving the rollout first.

  • OpenAI says Plus, Pro, Business, and Enterprise users receive GPT-6 with Intelligent UI globally today.
  • Free and Go users begin receiving the update tomorrow.
  • Responses can include tappable controls, forms, graphics, charts, and interactive experiences built around the question.
  • The change applies to ChatGPT’s Chat tab only. OpenAI says the models powering Work and Codex are not changing in this rollout.

What it means Watch how users adopt interactive AI answers, because ChatGPT is making task-specific interfaces a standard part of consumer conversations rather than a separate app experience.

Source: OpenAI

Claude Haiku 5.5 makes subagents much cheaper

Anthropic has released Claude Haiku 5.5 for high-volume tasks such as classification, summaries, compaction, database queries, and coding subagents. It says the model costs about 75% less to run on average than Haiku 4.5.

  • Requests with prompts up to 100,000 tokens are priced 90% below Haiku 4.5, while longer requests are 50% cheaper.
  • Haiku 5.5 adds adjustable effort, letting you trade cost for more intelligence task by task.
  • Sonnet 5.5 cache reads now cost $0.10 per million tokens, down from $0.20, which Anthropic says cuts most agentic work by about 20%.
  • Anthropic also updated its Python and TypeScript SDKs with computer-use and browser-use support in beta.

What it means Reprice repetitive agent steps and subagent jobs: cheaper short-context calls and lower Sonnet cache-read costs can change which tasks are practical to run at volume.

Source: Anthropic

Perplexity releases embeddings that search PDFs as pages

Perplexity has released two late-interaction embedding models for retrieving text, images, and rendered document pages in one shared embedding space. The models are publicly available on Hugging Face.

  • Instead of reducing each document to one vector, the models retain a 128-dimensional vector per token and score query-to-document matches with MaxSim.
  • They embed rendered PDF pages directly, so retrieval can preserve tables, figures, and layout without OCR.
  • A 9B model can index a corpus while a 0.6B model handles queries against that same index.
  • Perplexity reports 92.4% on MADQA, a 500-question benchmark over 800 PDFs.

What it means Use this approach when document layout carries meaning, because rendered-page retrieval can search scans, slides, tables, and figures without first flattening them into text.

Source: Perplexity

Windows ships runtime containment for AI agents

Microsoft Execution Containers, or MXC, are now generally available on Windows 11. Organizations can define which files and network destinations an agent may access, with policies enforced while it runs.

  • MXC is built for containment needs that can change by prompt and tool call, rather than a fixed traditional sandbox.
  • Microsoft lists Codex, GitHub Copilot, Replit, LM Studio, NVIDIA OpenShell, and Unsloth as already supporting MXC.
  • Anthropic Claude Code, Box, Manus, Perplexity, Raycast, and others are listed as planning support.
  • The platform also ties agent identity and activity into management tools such as Agent 365 and Intune.

What it means For Windows-based agent deployments, runtime file and network policies give you a concrete containment layer to require before agents receive broader local access.

Source: Microsoft Windows

Google opens SynthID watermark checks to everyone

Anyone can now upload image, video, or audio files to Google’s SynthID Detector to check for a SynthID watermark. The portal covers media generated by Google and participating partners including OpenAI, NVIDIA, and Kakao, with Apple described as forthcoming.

  • The check is for SynthID watermarks, not a universal detector for all AI-generated media.
  • Google says it has watermarked 180 billion images and videos, plus more than 240,000 years of audio.
  • The company says its existing verification systems handle one million requests per day.
  • SynthID checks were already available in Google Search, Gemini, and Chrome.

What it means Treat AI-media verification as an ecosystem-specific provenance signal, not proof of origin, while public checks make watermark support more visible to users and customers.

Source: Google AI

More from The AI Builder Roundup →