Suivez les actualités des modèles DeepSeek, de l’API, des tarifs et des outils pour développeurs, ainsi que les nouvelles de l’IA. DeepSeek V4 Hub est une publication indépendante.
Dsh-guizang-design is a DeepSeek Harness plugin that turns one article into a web page, image-text cards, or a horizontal web deck. It needs dsh 0.2.1-rc.2 or newer and below 0.3.0, and installs from the plugin page or one CLI command. All three outputs share one theme, and the de-AI skill only handles Chinese.
Lire l’articleDsh-media-desk is a free, MIT-licensed community plugin that adds a Media Desk page to the DeepSeek Harness sidebar. It captures one-line ideas, drafts a spoken script, title, cover, images and a six-frame storyboard, then puts the finished piece on a calendar. Text runs on the model you already use in DSH, image generation needs your own OpenAI-compatible image key, and optional account import costs under a cent to about two cents per account. It does not publish for you and it keeps all content local.
Lire l’articleKimi K3.1 Pro is available now with a 1,048,576-token context window and $2 per million input tokens / $10 per million output tokens. GPT-6.2 is not officially documented as of November 6, 2026. Google Cloud quota entries name gpt-6.2-astra and gpt-6.2-sol , and The Information reported a possible Friday launch. Kimi has published token prices; reported per-task figures for GPT-6.2 do not establish how it will be billed. If you need a model today, Kimi is the one you can call. If you already use the GPT tool stack, wait for the official price sheet.
Lire l’articleAu 4 novembre 2026, DeepSeek Harness comptait 1,54 million d’utilisateurs actifs quotidiens et consommait environ 92 000 milliards de tokens par jour. Son écosystème comprend près de 35 000 plugins tiers et un programme de soutien aux auteurs est en préparation.
Lire l’articleChatGPT Dots, les agents persistants de ChatGPT, peuvent désormais tenir une conversation en tête-à-tête sur RCS dans Google Messages sur Android. Sur iPhone, le même pont passe par iMessage. Il s'agit d'un aperçu limité pour les abonnés ChatGPT Pro aux États-Unis. L'option Texting doit apparaître sur le Dot et le numéro de téléphone doit être vérifié. Il n'y a pas de solution de secours SMS, pas de discussion de groupe, pas d'appel et pas de cryptage de bout en bout.
Lire l’articleMoonshot a publié les poids de Kimi K3.1 Pro, un modèle à mélange d’experts de 2 800 milliards de paramètres doté d’une fenêtre de contexte d’un million de tokens. Il obtient 65,7 % sur Terminal-Bench 4.0, contre 66,4 % pour Claude Opus 5.5 et 58,2 % pour GPT-6 Astra. Avec des outils, il atteint 92,7 sur BenchCAD (Opus 5.5 : 96,2) et 64,4 sur OfficeQA Pro v2. Le taux d’hallucination par affirmation passe de 23,4 % pour K3 à 8,3 %. Des poids ouverts ne rendent pas l’inférence bon marché : à cette échelle, la plupart des équipes utiliseront un service hébergé.
Lire l’articleDeepSeek Bot is an open-source desktop app and plugin built on DeepSeek Harness (DSH). You install it, point each bot at a model you have already configured, and give the team roles and memory. The bot files stay on your machine. ChatGPT Dots is OpenAI's hosted agent feature, introduced on September 29, 2026. A Dot gets its own cloud computer, connects to ChatGPT apps, and can pick up Codex work or edit Work automations. It is limited to Personal Pro, Business Premium, and Enterprise plans, not Free, Go, or Plus. In practice, DeepSeek Bot is a toolkit you assemble and own. Dots is a managed agent you configure inside ChatGPT.
Lire l’articleDeepSeek V4.1 Flash is available now through its public API. It offers a 1M-token context, image input, tool calls, MIT-licensed weights, and off-peak prices of USD 0.15 input and USD 0.60 output per million tokens. Google describes Gemini 4 Carbon as its strongest model for long-horizon engineering and security work. It plans introductory pricing of USD 2 input and USD 10 output per million tokens. As of November 2, 2026, public access, a model ID, and a context window have not been announced; initial access is limited to selected cyber defenders. DeepSeek exposes Search and Fetch as separate APIs. Google’s Gemini API supports search grounding and URL reading inside a request, but Carbon-specific tool availability has not been confirmed.
Lire l’articleDeepSeek Harness (DSH) plugins run with your permissions, so an unattended run can delete a file, leak a key or burn through your token budget. Scan plugins before install, keep a recycle-bin guard and a secret scanner on, review the output with a second model, and set a monthly budget. The official notice says DSH has not had a security audit, so keep backups.
Lire l’articleStart with GLM-5.3 for self-hosting, Claude Fable 5 for complex coding and knowledge work, GPT-6.1 Terra for visual reasoning and computer use, and Gemini 4 Argon for long-context, multimodal work. These are starting points, not a definitive ranking: the four models have not been compared here under identical conditions.
Lire l’articleStart with DeepSeek V4.1 Flash if API cost, a large context window, and open weights matter most. Test Grok 4.7 if your workflow relies on Cursor, integrated search tools, or complex knowledge work. Vendor-reported benchmarks use different setups, so they do not establish a head-to-head winner.
Lire l’articleKimi K3.1 Pro stood out for its SVG design and detailed interactions in a four-model review. Claude Opus 5.5 earned the reviewer's strongest praise for visual quality overall. DeepSeek V4.1 Flash 1017 had the lowest reported bill. The two cases below show why those judgments need some context.
Lire l’article