DeepSeek Opens API Access to App Capabilities, with Search and Voice Interfaces Available for Standalone Calls
Product LaunchAPI
DeepSeek is opening its app capabilities—including web search, file input, link reading, voice interaction, and safety review—to developers. They can be integrated into conversations through parameters, or developers can call the OCR, speech recognition, and speech synthesis APIs individually. Search and web scraping each cost $0.75 per 1,000 calls, while OCR costs $0.225 per 1,000 pages. Fees for these add-on tools are charged separately from model token fees. DeepSeek has also made available a frontend demo that supports importing conversation history.
OpenAI Launches ChatGPT Wallet, Enabling Agents to Shop with User Approval
Product LaunchAgentsPayments
OpenAI has launched ChatGPT Wallet, which lets agents complete purchases on a user's behalf after approval. The initial rollout is for ChatGPT Pro subscribers in the U.S. aged 18 and older. Each transaction requires security verification, and agents use payment tokens restricted to specific purposes without accessing users' actual card numbers. The feature was developed with American Express, Mastercard, and Visa. Eligible cardholders can retain their existing rewards and transaction dispute protections.
Cerebras Reports Third-Quarter Revenue of $230.9 Million and Cloud Partnership with Microsoft
ChipsComputeEarnings
Cerebras reported third-quarter revenue of $230.9 million, up 70% quarter over quarter and 55% year over year, with non-GAAP net income of $30 million. The company announced a cloud partnership with Microsoft, following earlier deals with AWS and Oracle, and plans to expand manufacturing capacity to more than 10 times its current level. The WSE-5 chip has completed tape-out, and mass production of the CS5 system featuring the chip is planned for early summer 2027, with a performance target twice that of CS4.
LangChain Raises $300 Million at a $3 Billion Valuation and Upgrades LangSmith
FundingAgentsDeveloper Tools
LangChain has completed a $300 million Series C round at a $3 billion valuation and launched a redesigned LangSmith platform spanning agent building, deployment, runtime tracing, evaluation, and governance. The platform supports deploying agents built with different frameworks and offers model fallback, cost controls, and sensitive-data protection. LangSmith Engine can detect failures and generate remediation and evaluation plans for human review. The company says more than 7,000 enterprises use LangSmith.
Local Inference Engine llama.cpp Releases Version 1.0, Supporting More Than 150 Model Architectures
Open SourceInferenceDeveloper Tools
The open-source local model inference engine llama.cpp has released version 1.0.0, alongside ggml 1.0.0. The release commits to keeping the public API stable, with breaking changes reserved for major versions. The engine currently supports more than 150 model architectures. Features include FP8 KV cache, MoE expert caching, hybrid CPU/GPU execution, and speculative decoding. It supports hardware from Apple, AMD, Intel, and NVIDIA.
Perplexity Releases Open Weights for Decider v1.3, Adding Precise Text Localization
Model ReleaseOpen Weights
Perplexity has released the open-weight Decider v1.3 and Decider Flash v1.3, built on Qwen3.8-27B and Qwen3.5-4B, respectively. Rather than generating text, the models return probabilities for true-or-false questions, multiple-choice questions, and scoring tasks. The new Locate feature returns source text spans and character positions for entity recognition, personal information detection, and extractive question answering. The models achieved an average accuracy of 91.6% across 63 query-classification tasks.
Tencent Hunyuan Releases Hy Realtime Speech Model with Tool Calling and Singing
Model ReleaseSpeech
Tencent Hunyuan has released Hy Realtime, a real-time speech model that adds tool calling and singing to its speech understanding and instruction-following capabilities. The model can determine from conversational context whether to search the web, and can handle ambiguous phrasing, unfinished instructions, and instructions changed mid-conversation. Its singing capability was designed jointly by Hunyuan and Tencent Music's Tianqin AI team. The model is already used for Yuanwen voice calls and Tencent Customer Service, with an API planned through Tencent Cloud.
Study Finds Initialization Signatures in Some Chinese Models Match Blackwell Chips
ChipsResearchExport Controls
An analysis of random-number patterns used to initialize model weights found that numerical characteristics in some Chinese models match those of NVIDIA Blackwell GB200 or GB300 chips. The models involved are from DeepSeek, Moonshot AI, MiniMax, and Ant Group. The method identifies chip signatures at the initialization stage; it does not prove that the same chips were used for the full training process, nor can it establish where the chips were located, how they were obtained, or whether export controls were violated.
Microsoft Uses AI to Scan Linux Kernel, Finding Eight Exploitable Privilege Escalation Vulnerabilities
AI SafetyCybersecurity
Microsoft used AI to scan the Linux kernel, identifying 59 upstream vulnerability reports, including 33 CVEs and eight exploitable privilege escalation vulnerabilities. The scan suggests that vulnerabilities are being found faster than they are being fixed.
10
Claude Managed Agents Opens Self-Hosted Private Beta
AgentsEnterprise Services
Anthropic's Claude Managed Agents has opened a private beta for self-hosted deployments. Enterprises can run agent sandboxes on their own infrastructure and manage credentials and conversation histories, giving them greater control over the runtime environment and data.
Don't Miss Tomorrow's Insights
Join thousands of professionals who start their day with AI Daily Brief