Saturday, Nov 14

DeepSeek API Docs published this on Nov 12, 2026 (source: api-docs.deepseek.com).

DeepSeek made the complete DeepSeek app feature set available through its API: web search, file input, URL reading, model safety and voice (ASR/TTS), either by enabling built-in capabilities through a few parameters or by calling standalone OCR, ASR and TTS endpoints. An open-source frontend demo at cdn.deepseek.com/chat-demo lets developers reproduce the full DeepSeek app experience in their own products and import existing DeepSeek conversation history; the code ships as the deepseek-ai/chat-demo repository on GitHub, with a build-a-chatbot guide in the API docs.

Cursor Blog published this on Nov 13, 2026 (source: cursor.com).

Cursor published the findings of its two-week investigation into the October 29 security incident, in which a malicious unauthorized copy of Anthropic's Claude Code extension appeared in the Cursor VS Code marketplace. The report traces the intrusion to an access key the company had issued in connection with Fireworks infrastructure, says the vast majority of customers were not affected, and notes each affected customer was contacted directly with the details that apply to them. Cursor is remediating the impacts and hardening its security measures based on the investigation.

Sierra Blog published this on Nov 13, 2026 (source: sierra.ai).

Sierra released version 1.0 of Poppy, the Personal Agent Protocol, after five weeks of design-partner testing against real websites, APIs and agents. The protocol - built largely on existing standards like MCP, OpenAPI and UCP with OAuth-based customer control - defines how personal agents connect with implementing companies, including staged adoption, agent guidance and feedback channels, and extensions such as the commerce extension co-authored with Shopify. Poppy v1.0 is live in Meta's Muse AI assistant, available to every Shopify merchant, and backed by more than 100 design partners.

Google Blog published this on Nov 13, 2026 (source: blog.google).

Google added Google Wallet to the connected apps in Gemini's Personal Intelligence framework, rolling out to all Gemini app users in the United States. With Wallet connected, Gemini can generate dynamic charts tracking cash flow, asset allocation and budgets in chat, surface forgotten subscriptions and recurring charges while flagging price increases, and answer questions like setting a holiday budget from last year's spending. Users can optionally link credit-card and bank accounts to Wallet through Plaid for deeper insights, and Google says it added new privacy protections alongside the feature.

Devin published this on Nov 13, 2026 (source: devin.ai).

Cognition introduced Devin Apps, private and secure web applications that run directly on Devin's cloud computer at stable URLs. Users describe an internal tool in chat and Devin builds and hosts it, with enterprise-wide sharing gated by Devin sign-in authentication. Apps have access to a team's existing MCPs, integrations and codebase, support secrets management and usage limits, and are billed on machine awake time. Cognition positions Apps both as a way to power internal tools and preview deployments and as a surface for extending Devin itself.

Federal Reserve published this on Nov 13, 2026 (source: federalreserve.gov).

The Federal Reserve's November Financial Stability Report says elevated stock valuations are partly driven by enthusiasm around artificial intelligence and potential breakthroughs in superintelligence, and warns AI investment could rival the 1990s telecom boom. In the Fed's survey, 87% of market contacts cited AI as a financial stability risk, up from 50% in the spring, pointing to earnings disappointments, cyberattacks and complex financing structures. The report notes debt issuance tied to advanced computing reached $680 billion through October - five times the 2023 pace and 31% of nonfinancial investment-grade issuance - alongside more than $1 trillion in future lease commitments and over $2 trillion in take-or-pay contracts and residual value guarantees, and hedge fund leverage at its highest recorded level since 2013. The data cut off on October 28.

Bland published this on Nov 13, 2026 (source: bland.ai).

Bland, the AI phone-calls company, introduced Tofu-1, a speech-to-speech model designed for calls that happen in bad conditions - moving cars, hospital rooms, weak signal, several people talking at once. The model listens while it speaks and reasons while it talks, and can tell the caller apart from everything else on the line such as a TV or a crying baby. Tofu-1 is available through a research preview at bland.ai.

OpenAI Developers published this on Nov 13, 2026 (source: developers.openai.com).

OpenAI published its deployment approach for persistent agents: the company runs agents across the entire company, including hundreds of OpenClaw agents - the open-source persistent-agent framework - that work around the clock writing code and opening PRs. The writeup separates trusted deterministic components from untrusted LLM execution by running the OpenClaw Gateway in its own container and cluster apart from the agent harness, and enforces hard boundaries through sandboxing: organization-managed network allowlists with explicit denials for metadata and paste endpoints, workspace-scoped writes, codex Linux process isolation, and a credential broker that keeps secrets invisible to agents. Configuration forbids danger-full-access mode and routes web search through OpenAI's cached index; agent operations are auto-reviewed by a second model, and an OpenClaw Enterprise effort is piloting these controls as a general agent control plane.

@ArtificialAnlys on X published this on Nov 14, 2026 (source: x.com).

Artificial Analysis released Coding Agent Index v2.1, upgrading two of its three components: FrontierCode-AA 2.0 (an implementation of Cognition's FrontierCode 2.0 with 132 real coding tasks scored on whether repo maintainers would accept the agent's changes) replaces v1.1, and Terminal-Bench 4.1 keeps the same 66 tasks with improved verification and anti-gaming hardening. Claude Fable 5.2 and Claude Opus 5.5 tie at the top with 58 in Claude Code, with Opus 5.5 (max) also hitting 58 in Cognition's Devin Fusion CLI paired with GPT-6 Luna High. Scores dropped about five points on average under the new FrontierCode 2.0, with GPT-6.1 Terra the only configuration roughly holding its score; per-task costs for configurations scoring 48 or above range from $0.30 for Codex with GPT-6.1 Sol to $44.22 for Claude Code with Claude Fable 5.2.

MathArena published this on Nov 13, 2026 (source: matharena.ai).

MathArena, the independent math-benchmark tracker, added Gemini 4 Carbon and Claude Fable 5.2 results on November 13, with Gemini 4 Carbon taking the top spot on its leaderboard. The Decoder's writeup notes the win comes at a steep price: running Carbon costs roughly 15 times more per task than GPT-6.1 Sol, which remains the value pick on the board. The update adds two frontier reasoning models to MathArena's running comparison of competition-math performance across labs.