New Manus 2.0 is Fully Autonomous, New Sonnet 5.5 Beats Opus, Tencent AI Companion & More AI News

7/10

AIAI RevolutionSeptember 29, 2026 at 10:59 PM14:54

Audio player

0:00 / 0:00

TL;DR

Manis has launched Manis 2.0 and the invite-only Q app to give AI agents their own communications, payments and computing resources, while Tencent tests an AI gaming companion and Anthropic introduces Sonnet 5.5, a cheaper coding-focused model with stronger safeguards.

KEY POINTS

Manis shifts from task tool to agent infrastructure

On September 28, Manis unveiled Manis 2.0 after announcing it would operate independently, alongside a standalone app called Q. The release frames AI less as a prompt-response tool and more as an always-on system that comes with compute, software environments and continuous runtime. The company’s leadership also took a direct swipe at rivals by arguing Manis had pursued a general-agent approach before products such as Codex or Claude collaboration tools emerged.

Cascade promises lower token use, faster runs and lower cost

The main technical change is a new execution base called Cascade, described as an in-house agent harness that keeps projects lightweight and calls in specialized capabilities only when needed. In one tested setup, Manis said the system used 23.2% fewer tokens, ran 28.2% faster and cut cost by 32% versus the previous version. Those figures were presented as a single benchmark rather than average production performance.

Studio expands into documents, code, video and games

The desktop app has been reworked into Manis Studio, a shared workspace for humans and agents that can handle local files and tools for documents, spreadsheets, PDFs, slides, websites, code, video and games. The system loads only the components a project needs, a design meant to keep workflows fast and cheap. The broader goal is to keep multiple formats and automations inside one connected project instead of splitting them across separate apps.

Video editing and game creation become flagship use cases

Two new pro environments headline the release: Video Editor and GameDev. Video Editor breaks projects into separate clips, images, subtitles, motion graphics and audio tracks so users can swap a song, insert footage or adjust animation without regenerating an entire ad. GameDev combines image, video and coding models, offers a playable template in the first session, and lets creators manually edit code, assets, movement and individual character details before publishing to the web.

Cloud computers underpin multiplayer and always-on work

Multiplayer game publishing is tied to Manis cloud computers, which keep servers and automations running after a user’s laptop is closed. The help documentation describes the service as a continuously running Ubuntu Linux virtual machine, while company statements also reference Linux, Mac and Windows instances, suggesting either an upcoming expansion or outdated documentation. The cloud setup is also central to remote computer use, allowing an agent to act inside an authorized session while the user watches live from a phone.

Automations now react to live events

Manis is extending automation beyond scheduled jobs to event-driven triggers. A new email, Slack message, calendar event, ad performance change or Notion update can now launch a workflow based on a single prompt that tells the system what to watch and what action to take. That pushes Manis further into operational software rather than a one-off assistant.

Q gives agents their own phone, email, wallet and machine

The new Q app applies the same infrastructure to personal AI agents. Each agent can receive its own phone number, email address, wallet and computer, enabling it to message people, make payments within a budget, complete tasks independently and even answer calls before sending a summary back to the user. Multiple agents can also work together in a group chat, splitting jobs such as venue research, shortlisting and deck drafting. Early access is invite-only across web, desktop and mobile.

Tencent tests an AI gaming companion with screen awareness

In China, Tencent is internally testing an AI companion called Gooseverse or Goose Dimension, built around digital human characters with voice chat, real-time screen recognition and in-game companionship. Users choose a male or female character and can speak with them normally or activate a gaming companion mode that talks aloud during play across mainstream online titles. The test is free, but interface elements already point to a metered model in which text chat, voice calls, gameplay companionship and image generation consume purchased energy credits.

Anthropic’s Sonnet 5.5 beats its own flagship on one coding benchmark

Anthropic released Sonnet 5.5 on September 28 as a mid-tier model for coding and office work. It is priced at $2 per million input tokens and $10 per million output tokens, with the company claiming it is more than 30% faster and up to 30% cheaper per task. On Terminal Bench 4.0, an agentic coding test, Sonnet 5.5 scored 70.6%, above Opus 5.5 at 66.4% and far ahead of Sonnet 5 at 10.3%, even though Opus remains positioned as stronger for open-ended tasks requiring sustained judgment.

Safeguards tighten as pricing raises questions

Sonnet 5.5 is the first Sonnet model to receive the same cyber safeguards used for Fable and Opus, with higher-risk cyber requests falling back to Sonnet 5. It also adds protections against reasoning extraction and ties preserved reasoning to the originating account, reflecting wider concerns after researchers reportedly extracted 315,320 reasoning blocks from public agent traces and uncovered 62 API keys, 33 passwords and seven private keys. A pricing inconsistency remains unresolved: Anthropic says Sonnet 5.5 matches Sonnet 5 pricing, but earlier reporting had said Sonnet 5’s lower rates were promotional and due to rise after August 31.

CONCLUSION

The latest releases show AI competition shifting from chatbots toward persistent agents that can act, pay, communicate and remain online. The next battleground is no longer just model quality, but who can turn AI into reliable infrastructure for work, play and continuous automation.

More from AI