This observation covers 7:51 am on 14 November to 7:51 am on 15 November 2026 in Australia/Sydney. It includes three technical changes that were opened or delivered and could be verified from direct sources. Some original releases occurred on 13 November, with independent reporting, version records or runtime measurements appearing within this observation window.
DeepSeek opens its full application API suite to developers
On 13 November, DeepSeek exposed the web search, file input, URL reading, safety and voice capabilities from its app through the API. This series previously recorded the release of the standalone Search and Fetch APIs on 17 October. The new state is that V4.1-Flash, served as deepseek-flash, can use these tools directly through the Messages, Responses and Chat Completions interfaces, with tool execution and context assembly handled by the service. Web search returns citations when thinking is disabled, while deep-thinking search can search and read pages during multi-step reasoning. File input supports PDFs, Office files and images. Search, Fetch, OCR, ASR and TTS are also available as standalone APIs, and DeepSeek released an open-source demonstration chat client.
Public-facing applications must enable safety checks, while developers remain responsible for content moderation, applicable model registration and labelling AI-generated content. Search and Fetch are billed only for successful requests that produce results, at US$0.75 per 1,000 requests for each service. Their account-level rate limits are 100 and 25 requests per second respectively. CASRAI independently checked the release scope and responsibility conditions on 14 November, but did not reproduce search quality, latency or file-parsing accuracy.
ChatGPT Wallet begins a limited rollout in the United States
OpenAI began gradually rolling out ChatGPT Wallet through dots on 13 November, enabling an agent to complete checkout on a merchant's website after the user authorises the purchase. After saving an eligible card, the user must review the merchant, items, amount and selected card before each purchase, then approve it using a security check such as Face ID or a fingerprint. The payment network then supplies a payment token in place of the actual card number, so neither the merchant nor the agent receives the real card details. OpenAI developed the feature with American Express, Mastercard and Visa, while existing card rewards and protections for eligible disputes continue to apply.
The present status is a limited rollout rather than general availability. It initially applies only to ChatGPT Pro users aged 18 or older in the United States who have an eligible US-issued card and US billing address. Some websites may block agent access, and user approval does not guarantee that a merchant has accepted an order. Reuters independently confirmed the launch, payment-token design and per-purchase security check on 13 November. Pre-authorising purchases by spending limit, product category or time period remains an announced later plan rather than a delivered capability in this release.
llama.cpp and ggml release stable 1.0 library versions
ggml-org released llama.cpp v1.0.0 and ggml v1.0.0 on 13 November. Compared with the 0.x releases, this change centres on version rules for public developer interfaces. An incompatible change to the public C API now increments the major version; a backward-compatible feature, backend or model-support update, or API addition increments the minor version; and a fix with no API change increments the patch version. Distribution channels include pre-release artefacts for five major operating systems, along with Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and Apple XCFramework packages.
The release does not claim a new model capability or a general inference speed increase. On 14 November, Local Model Watch confirmed the stable release and tested build b12359, which contains v1.0.0, on a three-thread Neoverse-N1 CPU. With the same small model, it measured prompt processing at 132.7±0.6 tokens per second and generation at 42.5±0.7 tokens per second; its v0.9.0 comparison measured 128.1±0.9 and 42.9±0.6 respectively. This test covers one CPU, one model and a short workload, and small differences may be run-to-run noise, so it cannot be generalised to GPUs or other models. Developers migrating to 1.0 still need to check removed legacy interfaces and dependency compatibility.
Sources
DeepSeek, The full DeepSeek App experience, now via API, 13 November 2026, https://api-docs.deepseek.com/news/news261113/
DeepSeek, Search & Fetch, accessed 14 November 2026, https://api-docs.deepseek.com/guides/search/
CASRAI, DeepSeek Opens App Features Through Its API, 14 November 2026, https://casrai.org/news/deepseek-app-capabilities-api-2026
OpenAI, Introducing ChatGPT Wallet, 13 November 2026, https://openai.com/index/introducing-chatgpt-wallet/
Reuters, OpenAI launches ChatGPT Wallet, letting AI agents pay for purchases, 13 November 2026, https://www.reuters.com/technology/openai-launches-chatgpt-wallet-letting-ai-agents-pay-purchases-2026-11-13/
ggml-org, llama.cpp v1.0.0: A stronger foundation for the local AI ecosystem, 13 November 2026, https://huggingface.co/blog/ggml-org/llama-cpp-v1
Local Model Watch, llama.cpp: Releases and Overview, updated 14 November 2026, https://localmodelwatch.tsuchitsuchi.com/en/engine-llama-cpp-en/
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.