DeepSeek opens app tools to developers as ChatGPT Wallet rolls out
DeepSeek opened web search, file input and speech tools to developers. ChatGPT Wallet began rolling out to eligible US Pro subscribers and requires approval for each purchase.
01
Products and tools
6 stories
2026-11-13DeepSeek
DeepSeek brings search, file input and speech to its API
DeepSeek added application tools to its Messages, Responses and Chat Completions APIs for V4.1-Flash. Once enabled, they let the model search the web and images, read links, and incorporate parsed PDF, Office and image files into a conversation. DeepSeek handles tool calls and context assembly on the server.
Search, URL fetching, optical character recognition, speech recognition and speech synthesis are also available as standalone calls. Extended tools incur their own charges alongside model token charges.
2026-11-13OpenAI
ChatGPT Wallet lets dots make purchases after individual approval
OpenAI began rolling out Wallet to eligible ChatGPT Pro subscribers in the US. Users aged 18 or older with an eligible US-issued card can have their dot complete a purchase. They review the items, total and selected card and approve each purchase with a security check such as facial or fingerprint recognition.
After approval, the payment network supplies a token in place of the actual card details. The chosen card is still charged, and the agent and merchant receive the token details. Checkout also depends on whether the merchant site and its payment systems accept these transactions.
2026-11-13Sierra
Sierra releases version 1.0 of the Poppy personal agent protocol
Sierra released Poppy 1.0, a protocol for personal agents to discover and use company websites, service agents and APIs. It builds on MCP, OpenAPI and UCP and uses OAuth so customers can control the scope of access.
Shopify stores now publish a Poppy file automatically. The commerce extension Shopify co-authored links Poppy to UCP, supports guest and signed-in shopping, and requires agents to ask customers before certain actions.
2026-11-13Ado Kukic
Claude Duet finishes inviting an eligible US waitlist cohort
Anthropic’s Ado Kukic said on November 13 that every US Claude Pro, Max and Ultra subscriber who was on the Claude Duet waitlist as of the previous day had been invited. The update applies to that specific group.
The waitlist remains open to new sign-ups. The October research preview introduced Duet’s background task capabilities. The latest invitations cover this group of existing waitlisted subscribers.
2026-11-13Apple Newsroom
Apple HomeView goes on sale in seven markets at $299 in the US
Apple put HomeView on sale on November 13 in the US, Canada, the UK, Australia, New Zealand, India and Singapore. The 7-inch device comes in tabletop and wall-mounted models, each priced at $299 in the US, and works with the Home app to manage accessories and automations.
Siri AI is in beta in HomeView software 27.2. Personal-context requests require a paired Apple Intelligence-enabled iPhone or iPad with Siri AI turned on and a supported language selected.
2026-11-13GitHub Changelog
GitHub Copilot app adds general availability for Enterprise Server
GitHub made Enterprise Server support generally available in the GitHub Copilot app. Developers can add an enterprise server account and work with its remote repositories, issues and pull requests. The integration was previously an opt-in preview.
The integration requires version 1.2.0 or later of the Copilot app. Cloud sessions are unavailable for Enterprise Server repositories.
02
Models and APIs
2 stories
2026-11-13Edge0 repository
Edge0 releases weights for its 145B and 120B models
Edge0 released weights for Edge0-145B and Edge0-120B. Both mixture-of-experts models derive from Qwen3.8-Flash-Next. The project repository documents running them on a single NVIDIA GPU under Linux or Windows.
One 145B benchmark caps an RTX 5090 D at 16 GiB of VRAM and limits system RAM to 20 GiB. Those are test limits for a particular quantization and hardware configuration. The project does not specify them as minimum requirements.
2026-11-13ggml-org
llama.cpp v1.0.0 removes deprecated APIs
The local inference project llama.cpp released v1.0.0, removing 35 deprecated functions, moving its NextN speculative-decoding interface into the public header and upgrading to ggml v1.0.0. Applications using the old interfaces need to check their calls.
The release adds FP8 key-value caches, a configurable host-memory expert cache and a text-to-speech endpoint. Its built-in web UI gains chat groups and a Markdown editor.
03
Research and practice
6 stories
2026-11-13Google Research
Google shows how Antigravity teams check mathematical proofs
Google Research showed how Antigravity’s /teamwork distributes long tasks among agents that coordinate, solve and verify. In the mathematics projects, researchers confirm that the formal statement matches the original problem, and Lean 4 checks the proof steps.
One formalization reported by Google contains about 1.8 million lines of Lean and involved 28,800 subagent invocations. Google also announced six months of access to its highest usage tier for 10,000 academic researchers.
2026-11-13OpenMOSS
MOSS-Live reports fewer false responses in full-duplex voice tests
OpenMOSS published a comparison of its real-time voice model MOSS-Live. In the team’s FDB v1.5 results, MOSS-Live responded incorrectly to irrelevant speech 2% of the time, compared with 68% for GPT-Live-1 and 75% for Doubao SeedDuplex.
A separate benchmark measures multi-turn tasks. In τ-Voice, MOSS-Live completed 57.93% of them and GPT-Live-1 completed 67.68%. The tests measure different outcomes: false responses during conversation and the completion of tasks across several turns.
2026-11-13Anthropic
Anthropic describes five building blocks for always-on agents
Anthropic published an architecture guide for always-on agents built around an event inbox, timers, memory, permissions and an outbox. A user reply, webhook or scheduled job can wake a worker, and follow-up messages can resume an existing session.
The guide keeps memories needed across tasks in durable storage and distinguishes actions that can proceed, should be blocked or need approval. An outbox can hold pending messages independently of the agent run so a dispatcher can recover after failures.
2026-11-13AWS Machine Learning Blog
AWS demonstrates stage-based model routing for coding agents
AWS demonstrated using NVIDIA NeMo Switchyard routing through a LiteLLM gateway to choose an Amazon Bedrock model based on a coding agent’s stage. In the example, ordinary requests go to an efficient model and a history of repeated failures can trigger an escalation to a more capable one.
The walkthrough uses a small number of runs across two coding tasks and records whether tests pass and estimated cost per task. The integration is experimental, and the demonstration does not establish cost or quality gains.
2026-11-13AWS Machine Learning Blog
AssemblyAI details how EC2 Spot lowered transcription compute costs
AWS described AssemblyAI’s speech transcription architecture, which uses Amazon ECS to mix Spot and On-Demand capacity across GPU families and queues work in Amazon SQS. AWS says AssemblyAI reduced inference compute costs by 70% compared with On-Demand pricing.
When an instance receives a two-minute interruption notice, its in-flight job returns to the queue so a healthy instance can take it. The reported 70% applies to the team’s inference compute costs.
2026-11-13AWS Machine Learning Blog
SageMaker HyperPod adds live inference data capture
AWS introduced Data Capture for SageMaker HyperPod. It offers three capture points along the request path and writes live inference data to Amazon S3 for error analysis or training data preparation.
A home-appliance support routing example uses Athena to identify misclassified tickets for a new product and then fine-tunes the model on labeled records. The walkthrough uses templated tickets and a separate held-out evaluation set.
04
Industry
1 story
2026-11-12Unemployment Labs
Researchers find Blackwell-compatible initialization traces in six open-weight models
Unemployment Labs analyzed weights released by DeepSeek, Moonshot AI, MiniMax and Ant Group and reported Blackwell-compatible initialization traces in six models. The researchers reproduced the matches for DeepSeek-V4-Flash and Kimi K3 on physical Blackwell chips.
The report says the matches reflect the environment that generated some initial weights and offer a clue about subsequent training hardware. The weights alone cannot establish which chips were used for the entire training run, where those chips were, or whether export controls were violated.