Meta Prices a 1M-Token Agent at $4.25 as Tencent Circles the $2B Manus Deal

In partnership with

1,000+ Proven ChatGPT Prompts That Help You Work 10X Faster

ChatGPT is insanely powerful.

But most people waste 90% of its potential by using it like Google.

These 1,000+ proven ChatGPT prompts fix that and help you work 10X faster.

Sign up for Superhuman AI and get:

• 1,000+ ready-to-use prompts to solve problems in minutes instead of hours—tested & used by 1M+ professionals

• Superhuman AI newsletter (3 min daily) so you keep learning new AI tools & tutorials to stay ahead in your career—the prompts are just the beginning

Claim your free prompts

Microsoft Brings Multi-Agent Orchestration to Go

Microsoft released the public preview of Microsoft Agent Framework for Go, extending its agent-development stack beyond Python and .NET. The SDK supports Microsoft Foundry, Azure OpenAI and other OpenAI-compatible models, Anthropic, Gemini, and the Agent2Agent protocol. Developers can add MCP tools, automatic tool calling, middleware, approvals, context providers, and OpenTelemetry tracing. It also supports multi-agent routing, checkpoints, streaming, background workflows, and human review. Strategically, Microsoft is positioning its framework as a model-neutral foundation for agents running inside cloud-native services, command-line tools, and backend infrastructure written in Go.

Meta Prices Its 1-Million-Token Agent Model at $4.25

Meta opened public access to Muse Spark 1.1 through the new Meta Model API, giving developers a multimodal model designed specifically for agentic tasks. The model can manage a 1-million-token context window, operate computers, use MCP servers and unfamiliar tools, and delegate work across parallel subagents. Meta priced the public preview at $1.25 per million input tokens and $4.25 per million output tokens, with $20 in introductory credits. Muse Spark can alternate between scripting and direct interface interaction while maintaining context across changing multi-application workflows. Strategically, Meta is combining low inference pricing with long-context orchestration to challenge OpenAI, Anthropic, and Google for high-volume enterprise agent workloads.

Tencent Eyes Manus After Meta’s $2 Billion Deal Unravels

Tencent is in talks to become Manus’s largest shareholder after Beijing ordered Meta to unwind its $2 billion acquisition of the agent startup. Tencent, ZhenFund, and HSG reportedly plan to buy Manus back from Meta for at least $2 billion. Tencent would likely hold the largest single stake but remain a minority shareholder, and the discussions are not final. Manus, which develops general-purpose agents that complete tasks with limited human input, moved its operations from China to Singapore last year. The strategic signal is bigger than one deal: frontier agent companies are becoming geopolitical assets whose ownership can be reshaped by national investment controls.

Mercor Buys Agent Simulator Deeptune While Chasing $20 Billion

Mercor acquired Deeptune, a startup that builds simulated software environments where AI agents can practice workplace tasks before deployment. Financial terms were not disclosed, although Deeptune raised $43 million in March in a round led by Andreessen Horowitz. Mercor CEO Brendan Foody personally invested in that round and later acknowledged that a potential acquisition influenced the investment, creating a governance question around the deal. Mercor also says its annualized gross billings have surpassed $2 billion and doubled in four months, while it explores financing at a $20 billion valuation. Strategically, Mercor is trying to own the full agent-training stack—simulation environments, expert graders, training data, and the benchmarks used to measure agent performance.

Oracle Gives Enterprise Agents Memory With Deletion Controls

Oracle introduced AI Agent Memory 26.6, an enterprise memory layer designed to prevent agents from retrieving the wrong user’s information or retaining data that should have been deleted. The system combines vector search, keyword search, and scoped retrieval across users, agents, conversations, metadata, and record types. It runs inside Oracle AI Database with in-database embeddings, HNSW indexing, hybrid search, and fewer external network hops. The release adds full create, read, update, and delete controls for threads, messages, memories, and profiles, along with cascading deletion and time-to-live policies. Strategically, Oracle is positioning governed agent memory as a database-management problem rather than an optional feature attached to a chatbot.

Cursor Is Reportedly Building a Workplace Agent Called Sand

Cursor is reportedly developing a general-purpose workplace agent internally called Sand, expanding beyond its original focus on AI coding. The proposed agent could respond to emails and messages, organize spreadsheets, and complete engineering work. Cursor declined to comment, and the company has not made a final decision about launching the product. The roadmap could also change as SpaceXAI moves toward completing its planned acquisition of Cursor. Strategically, the project suggests that coding-agent companies increasingly view general enterprise work as the larger market and are preparing to compete directly with ChatGPT Work and Claude Cowork.

The toughest room in advertising. Ad Studio just cleared it.

Major brands just took over a Times Square billboard, and every ad was built in Hightouch Ad Studio. Sweetgreen, HelloFresh, and Tripadvisor cleared the toughest room in advertising. Create, edit, and launch on-brand ads in a few clicks.

See what you can build

🗺️ The Map of AI Agents (Live & Growing)

We’re charting the entire AI agent ecosystem — thousands of options across categories.
Your next agent is already on the map.

👉 Explore the Landscape Map →