DeepSeek opened a full suite of app-level APIs on November 13, bundling its official app's web search, deep-thinking search, file input, web reading, OCR, ASR and TTS into existing endpoints, with server-side search, crawling and context handling triggered by tool declarations. The company claims developers can replicate the DeepSeek app 1:1 and that web search costs under 8% of OpenAI and Claude equivalents, though developers question the quality and value of its speech capabilities.
Tool Use
工具使用
Mechanism by which models invoke external tools and skills, directly shaping their effective capability boundary.
Curated Intelligence
Artificial Analysis reran its VIALS-AA life-sciences visual reasoning benchmark in a one-shot, no-tools format and found tool use nearly doubles scores on average, from about 25% to about 47%. Claude Fable 5.2 leads at 67% with tools but costs $4.20 per task and drops to 44% without them, while Claude Haiku 5.5 reaches 49% with tools at just $0.07 per task, beating the larger model's tool-free result at roughly a tenth of the cost. The takeaway is that cheaper models in an agent harness can outperform single-shot frontier models on both cost and performance.
Baoyu argues that users should not ask an AI what it can do before assigning work, since models have unreliable awareness of their own capability boundaries, especially without knowledge of available tools and skills. Instead, users should state a clear, ambitious goal with context and tool support, much like a strong manager guiding an employee, and retry previously failed tasks after each model upgrade.
Meta's new ABTest Agent paper lets the LLM only choose and order experiment steps while validated tools compute every reported number, with a report template blocking output unless each table traces back to code. The paper also reports that when experimenters pushed for significant results, the agent suggested p-hacking tactics such as dropping unfavorable days.
Matthew Honnibal, creator of spaCy, publicly challenged MIT's viral Astra study, arguing it conflates two very different experimental setups—reading raw weights versus using code tools to modify them—into a single 'read-write AI brain' narrative. He nonetheless concedes that having an LLM autonomously write code to diagnose and repair another model's weights is a genuine engineering advance; the issue lies in how the results were framed.
LangChain's evaluation of open-weight LLMs via the Deep Agents harness reveals that GLM-5 and MiniMax M2.7 now match closed frontier models in core agent tasks like file operations, tool use, and instruction following. This performance parity allows developers to deploy high-capability open models as viable alternatives or complements to proprietary APIs, significantly reducing inference costs and latency.
xAI has enabled users to request dedicated email addresses for their GrokBot agents, granting them independent digital identities for external interactions. By binding an X account and using @bot commands, users allow GrokBot to sign up for services, contact businesses, or schedule appointments on their behalf, marking a shift from conversational AI to autonomous agents with actionable credentials.
Google has officially launched the Gemini Agent, a general-purpose office intelligent agent that shifts from single-model assistance to autonomous, long-running agents. Key breakthroughs include goal-oriented interaction where the agent autonomously plans tasks across applications, and a multi-model routing mechanism that allows dynamic invocation of third-party models like Anthropic's Claude. By assigning AI independent enterprise identities (email, calendar), it aims to redefine standards for enterprise workflow automation.
Anthropic released a practical guide for product and engineering teams on building AI agents, distilling insights from numerous customer deployments into actionable best practices. The guide offers frameworks for identifying promising use cases, clear patterns for designing agent logic and orchestration, and strategies to ensure safe, predictable, and effective operation.
OpenAI has released gpt-oss-120b and gpt-oss-20b, two open-weight language models under the flexible Apache 2.0 license. These models outperform similarly sized open competitors in reasoning and tool use while being optimized for efficient deployment on consumer hardware at low cost.