What's new from vendors
Real product updates pulled from each vendor's own changelog page. Fresher than our editorial re-review cycle; grounded in what vendors are actually shipping.
Refer to this changelog for the full baseline of the recent Bifrost releases.
This prerelease features improvements such as ChatGPT passthrough routing, user agents tracking, and more.
An AI-powered tool that allows you to create, use, and share personalized micro apps ('sparks').
Be among the first to build Agent Apps for GitHub.
This update removes the boomerang feature and the related validation workflows.
Added features for serving project capabilities over MCP and fixed various bugs related to connection URL sanitization.
GPT-5.6 Sol and Luna are ahead of Terra at every point on the Intelligence vs Cost per Task chart.
Fixed `canUseTool` returning `{behavior: 'allow'}` without `updatedInput` being rejected as a deny.
Auto mode now available without opt-in; several bug fixes and improvements.
Added MDL layout versioning and enhanced the function list on specific SQL dialects.
Introduced a new core module with browser WASM support and several feature enhancements including BigQuery support.
Meta's Muse Spark 1.1 scores 51 on the Artificial Analysis Intelligence Index and is cost and token efficient compared to its peers.
Features include sharing the registered OAuth App client id, enabling interactive login.
Features support merged community benchmarks in the next release.
Added `command_lifecycle` frames to stream-json and SDK sessions.
Introduced the ability to contribute benchmarks as PRs without using gh CLI.
Introduced global --tui flag to force interactive TUI.
GPT-5.6 Sol comes close second to Claude Fable 5 in the Artificial Analysis Intelligence Index at one third of the cost.
Added directory path suggestions and improvements to background agent updates.
Interrupt control responses now include `still_queued` (UUIDs of queued async messages).
Bug fix for internal llmfit-core version requirement.
Fix for a bug with OpenAI Chat Completion endpoints where a tool call with empty arguments could result in a JSON error from some providers.
Added `terminal_reason` values `tool_deferred_unavailable` and `turn_setup_failed`.
SpaceXAI’s Grok 4.5 scores 54 to place fourth on the Artificial Analysis Intelligence Index.
Our independent leaderboard for ServiceNow’s EnterpriseOps-Gym.
New auto mode rules and numerous bug fixes, especially for Windows.
Fixed ties with hook events and improved background session management.
Added a `background_tasks_changed` system message with the full set of live background tasks.
Added warnings for expiring logins and fixed multiple bugs related to agent functionality.
Universal-3.5 Pro is now live for pre-recorded audio, built for real-world conversations, transcribing across 18 languages.
Enhanced memory management and added new data-reading capabilities.
Async STT transcript retention is now set to 30 days by default, clarifying previous documentation.
The singular language_code connect parameter is deprecated in favor of language_codes for better selection.
Universal-3.5 Pro Streaming now supports language_codes list for language steering while maintaining code-switching.
Streaming WebSocket now accepts compressed Opus audio, allowing more efficient data usage.
The 'Find Papers' sidebar has been introduced, allowing seamless research while writing.
Exciting new features including the addition of Claude Sonnet 5, folder export options, and improvements to document reordering.
Introduced v5 project layout and improved schema handling.
Users can now assign and edit contributor roles on library items, improving collaboration.
We have rebuilt the playground from scratch to be the main workspace for your AI applications, adding evaluators and the ability to load test sets.
Agenta now features a dark theme across the whole app, allowing users to switch between light, dark, and system themes.
Major updates including the launch of Ballad 1.1 for efficient credit use and an improved Plugin editor.
Introduction of Ballad 1.1 to enhance the Excellent Prose Mode's performance and capacity.
新增 Recall 上下文预算控制、L1 / L2 / L3 提示词按用户输入语言自适应等功能。
The addition of integrated, Story Bible-aware spelling and grammar checking within the editor.
New languages added to the language selector: Italian and Traditional Chinese.
Annotation queues help to manage traces by allowing users to build queues from them, score outputs, and export labeled test sets.
新增 Recall 上下文预算控制和语言自适应的新功能，修复兼容性问题。
修复 L2 抽取及兼容性相关问题，增加 Embedding 发送维度的可选关闭功能。
Introduction of Sudowrite Chat and Feedback, features designed to assist authors throughout the writing process.