Collaborative Rooms, frontier models, and platform updates

This update covers model choices, ChatGPT subscription connection, platform changes, and the current guide to working in Rooms.

Current guide to Rooms

The current Rooms guide covers shared conversations with colleagues and agents. Use it to understand room participation, @mentions, and ongoing work.

Learn more in Rooms.

New frontier and coding models

OpenAI's GPT-6.1 Sol (openai/gpt-6.1-sol) is now available out of the box in the model picker. Delivering frontier-tier coding, computer use, and complex agentic reasoning approaching GPT-6 Astra, GPT-6.1 Sol features a 950,000-token context window and 128,000 max output tokens.

Anthropic's Claude Opus 5.5 (claude-opus-5-5) is now selectable across all threads, persistent agents, and schedules. With a 1,000,000-token context window and adaptive reasoning, Opus 5.5 is designed for complex agentic analysis and technical synthesis. In addition, xAI Grok 4.7 (370K context, 128K max output) and Meta Muse Spark 1.3 (920K context) expand the catalog of high-capacity coding models.

Learn more in Models and compute.

Sign in with ChatGPT

You can now sign in to Hyperagent directly using your ChatGPT account across both the web app and the iOS app. If you hold an active ChatGPT Plus or Team subscription, Hyperagent automatically connects your plan through token sharing so your OpenAI model requests run directly against your subscription instead of consuming Hyperagent platform credits.

The model picker displays a clear badge indicating when an OpenAI model runs on your plan. If your ChatGPT usage allowance runs out mid-task, opted-in paid accounts automatically continue to their Hyperagent plan or switch models with a single click.

Learn more in Models and compute.

Auto profiles and manual selection

The model picker has been redesigned around three clear Auto profiles: Quality (gem icon) for frontier coding and complex reasoning, Balanced (scale icon) for everyday research and analysis, and Low cost (coins icon) for high-volume automated tasks.

To keep the menu compact, the full catalog of specific pinned models is organized into a clean Manual drawer grouped by provider. Scheduled invocations and webhooks now feature a dedicated Default section at the top of the menu to inherit an agent's standing model cleanly.

Learn more in Models and compute.

Working notes retain access

The Documents setting controls access to the document library. When it is off, agents can still read and update the current run's own working notes so ongoing work can retain its decisions and progress.

Learn more in Tools.

Updated default speech model

The default flash tier of the GenerateAudio tool has been upgraded to Google's Gemini 3.8 Flash-Lite TTS. Audio generation latency is cut by roughly 40%, and native language coverage expands to more than 100 languages. All existing agents and automated voice delivery jobs benefit automatically with zero prompt changes.

Learn more in Media generation.

Additional improvements

• Inbound room webhooks. Collaborative rooms include a dedicated Webhooks management panel from the room actions menu to post external events and route payloads to eligible room agents.
• Sent attachment previews and downloads. File chips in chat messages for video, audio, PDFs, and text open the full output viewer or download directly from the chip with one click.
• Microsoft grant transparency. The Microsoft integration manage dialog audits actual account consent alongside Hyperagent's configured runtime tier, clarifying when write consent remains stored on Microsoft's servers.
• Marketplace homepage redesign. The Marketplace opens on a refreshed "Hire your team of agents" hero layout featuring portrait-led agent cards with four-line capability summaries and expandable descriptions.
• Configurable reasoning effort. Effort controls cover Google models such as Gemini 3.8 Flash as well as open-weight and specialist models. Gemini is not an open-weight model; the available effort levels depend on the model and provider.
• Self-delegation approval. A permitted self-delegation starts without a separate handoff approval. The child inherits the calling run's action-approval posture; tool permissions, write restrictions, and depth limits still apply.
• Chat transcript tool folding. Sequences of repetitive tool calls and iterative document revisions fold into a single compact, titled pill beneath the assistant's explanation.
• Mobile fullscreen viewer. Tapping document cards and file attachments in chat on iOS and mobile web opens them directly into the full mobile viewer.
• Kimi multi-page PDF vision. Moonshot Kimi models process uploaded PDF documents as rendered page images, enabling accurate chart and layout analysis.
• Publishing dialog clarity. The public visitor comments toggle is elevated to the main dialog card, and an explicit Publish button confirms public link generation.
• Library ID and link lookup. Paste an artifact ID, document ID, thread ID, or thread URL directly into the Library search box to locate matching files immediately.
• Unified Slack bot removal. Conflicting bot menu items are consolidated into a single Remove dialog with clear impact descriptions ordered by consequence.
• Automatic tool enablement. Connecting an integration account to an agent automatically enables the corresponding tool capabilities.
• In-place MCP re-authentication. Custom MCP server connections in Settings feature a Sign in again button to refresh expired tokens without wiping agent bindings.
• Curated tool error messages. Integration errors across Slack, Google Slides, and MCP servers replace raw stack traces with human-readable diagnostic summaries and recovery steps.
• Advertising toolkit updates. Google Ads adds ad-group location and language targeting for Demand Gen campaigns, while Meta Ads adds full update controls for ad sets and ads.