Weekly roundup · 7–13 Nov 2026
A.I. for builders: week ending 13 Nov 2026
8 short notes on what changed that week, each with its primary source and one line on why it matters if you build things.
DeepSeek opens its full app features to the API (opens in new tab)
Why it matters: DeepSeek has put the features behind its own chat app into the DeepSeek API: web and image search with citations, deep thinking search, file input, URL reading and safety checks, run server-side on deepseek-flash. Search, Fetch, OCR, speech-to-text (ASR) and text-to-speech (TTS) are also sold as standalone APIs, and DeepSeek published an MIT-licensed demo chat client so you can see the whole flow working. The built-in tools work in all three protocols (Chat Completions, the OpenAI-style Responses API and the Anthropic-style Messages API) with the model name deepseek-flash (DeepSeek-V4.1-Flash-1017); you turn them on with the deepseek_web_search and deepseek_read_url tools, a safety_check parameter and an optional auto-generated conversation title. DeepSeek runs the searches and page reads itself, so your client only renders the returned blocks and passes them back unchanged next turn, which removes the search, scraping and file-parsing plumbing most chat apps build by hand. Tool fees come on top of token charges, per DeepSeek's pricing page: $0.75 per 1,000 searches or fetches, $0.225 per 1,000 pages for file parsing and OCR, and TTS at $0.225 per 10,000 characters and ASR at $0.058 per hour of audio, both shown as limited-time 50%-off prices. DeepSeek says public-facing apps must enable safety checks and that developers remain responsible for content moderation, model registration and A.I.-generated content labelling where those apply. The demo keeps your API key and chats in the browser and is marked as for demonstration only, so treat it as a reference, not a production client. If you are building a research assistant, document Q&A or voice bot on a budget, prototype it on deepseek-flash with these tools and compare cost per answered question against your current search-plus-scraper stack before committing.
Source: DeepSeek: The full DeepSeek App experience, now via API ↗DeepSeek API docs: Build a Chatbot ↗DeepSeek API docs: Models & Pricing ↗deepseek-ai/chat-demo on GitHub (MIT) ↗@deepseek_ai on X ↗
Read the full write-up about DeepSeek opens its full app features to the API
Google releases Gemini 4 Hydrogen, its fast Gemini 4 model (opens in new tab)
Why it matters: Google has released Gemini 4 Hydrogen, the fast, cheaper member of the Gemini 4 family, built for everyday coding and agent work and available to all tiers of Google AI Studio and the Gemini API from today. It costs $1.50 per million input tokens and $7.50 per million output tokens, flat across the full 1M-token context window, which makes it the default model to test for high-volume agent steps. You call it as gemini-4-hydrogen; the model docs list a 1M-token context, text, image, audio, video and PDF input, and up to 1M output tokens, and Google says it has four thinking levels (low, medium, high as the default, and a new xhigh). Google pitches it as a replacement for Gemini 3.8 Flash and as the worker model under Carbon or Argon: the new Orchestrator in Google Antigravity switches between Carbon and Hydrogen inside one session, and Google says that reached near-Carbon success on Terminal-Bench 4.0 with 24% less quota and 19% faster runs. The other performance claims are Google's own too, including 2.5x faster task completion than 3.8 Flash, 57% versus 18.2% on Terminal-Bench 4.1, and 2x to 2.8x speedups with 12% to 23% fewer tokens reported by early partners. Artificial Analysis independently scores it 54 on its Intelligence Index at the xhigh setting, and notes it is very verbose (it used 304M output tokens on the index), so check output-token spend, not only the price per token. In Antigravity it is limited to paid Google AI Pro and Ultra subscribers. If you run agents on 3.8 Flash or use a frontier model for routine edits, test runs, or file search, try Hydrogen on your own tasks at the default high level first, and compare cost per finished task before moving up to xhigh.
Source: Google: Introducing Gemini 4 Hydrogen ↗Gemini API docs: models (gemini-4-hydrogen) ↗Gemini API pricing ↗Google Antigravity: Gemini 4 Hydrogen and Orchestrator ↗Artificial Analysis: Gemini 4 Hydrogen ↗@GoogleDeepMind on X: Meet Gemini 4 Hydrogen ↗
Read the full write-up about Google releases Gemini 4 Hydrogen, its fast Gemini 4 model
Alibaba opens a Qwen4-Max preview for developers (opens in new tab)
Why it matters: Alibaba's Qwen team has opened a preview of Qwen4-Max, the flagship of its next-generation Qwen4 family, aimed at coding, front-end builds and agent work. Developers can call it today as qwen4-max-preview through a QwenCloud Token Plan subscription only (no pay-as-you-go), and the preview runs from November 11 to November 25, 2026, after which this model is retired. QwenCloud's docs list text, image and video input, a 1M-token context window, up to 393,216 output tokens shared between thinking and the answer, thinking on by default (reasoning_effort low, medium or xhigh), function calling, web search and JSON Schema outputs, and compatibility with the OpenAI Chat Completions, Responses and Anthropic Messages APIs, so tools such as Claude Code, Cursor, Cline, Codex and Qwen Code can point at the Token Plan base URL. During the preview, Token Plan charges it at 20% of the standard credit rate, with a further discount overnight (22:00 to 08:00 UTC+8). Qwen also offers early access in QwenWork, Qoder, the Qwen app and QwenStudio. Qwen says it brings deeper knowledge and stronger coding and agentic skills but has published no benchmark scores, parameter count, general-release date or open-weights plan for it. One API change to check: from this model on, max_tokens caps thinking plus answer together, so a limit sized for the answer alone may cut replies short. If you already use a Qwen Token Plan in your coding agent, this is a cheap two-week window to test it on your own tasks; don't build anything that depends on it past November 25.
Source: QwenCloud docs: Latest model, Qwen4-Max-Preview ↗@Alibaba_Qwen on X: Qwen4 is coming ↗QwenCloud docs: Token Plan Personal (preview terms) ↗
Read the full write-up about Alibaba opens a Qwen4-Max preview for developers
Google DeepMind releases AlphaGenome 2 on its research API (opens in new tab)
Why it matters: Google DeepMind has released AlphaGenome 2, a DNA sequence model that predicts how genetic variants change gene activity in specific human cell types rather than across broad tissue averages. It is available today through the AlphaGenome API, free for non-commercial research, with commercial research access in private preview only. Unlike the first model, which predicted thousands of tissue tracks in parallel, AlphaGenome 2 is conditional: you pass a million-letter DNA sequence plus one target condition (a tissue, cell type or state, picked from a reference catalogue or built on the fly from your own single-cell data), and it adds DNA methylation to the readouts it predicts. DeepMind says it was trained on more than 107 million single-cell profiles from human and mouse samples and can condition on a new single-cell RNA-seq snapshot from a few dozen cells without retraining. The performance figures are DeepMind's own: on the TenK10K immune-cell dataset it says AlphaGenome 2 identified 30% of causal variants versus 16% for a fine-tuned AlphaGenome 1, and 43% versus 23% of directional gene effects. In the Python client it lives at alphagenome.models.v2; the docs say to use it for new analyses and keep AlphaGenome 1 to reproduce the original paper or score all tissues at once, note that AlphaGenome 2 returns raw scores only, with no quantile calibration, and list four input widths from 16 KB to 1 MB, with 1 MB recommended. The terms matter: API outputs are for non-commercial research only, must not be used to train other models and must not be used for clinical decisions; commercial work goes through Google Cloud Model Garden. DeepMind says AlphaGenome 2 Skills for Google Antigravity are coming soon but has not given a date. If you build bioinformatics pipelines or research agents, get a free API key and try the v2 client on a small set of variants; it is not a general-purpose model and not something to put behind a commercial product yet.
Source: Google DeepMind: AlphaGenome 2 ↗AlphaGenome API (get a key) ↗GitHub: google-deepmind/alphagenome ↗@GoogleDeepMind on X: AlphaGenome 2 ↗
Read the full write-up about Google DeepMind releases AlphaGenome 2 on its research API
Google launches Gemini Decision 1 and the Decisions API (opens in new tab)
Why it matters: Google has released Gemini Decision 1 (gemini-decision-1), a model that answers bounded questions with probabilities instead of generated text, and a new Decisions API endpoint to call it, available now through the Gemini API in Google AI Studio (v1beta). You send one input of text, images, audio, video or PDFs (up to 262,144 tokens) with up to 1,000 independent questions, mixing three types: yes/no, multiple choice (2 to 255 options) and ordered rubric scores, and each answer comes back with probabilities and a 0-to-1 confidence score your code can use to send unclear cases to a fallback model or a person. Pricing is $0.10 per million input tokens with no output-token charge, the shared input is billed once per request however many questions you ask, and there is a free tier in AI Studio for prototyping; the Batch API is not supported, and the model does no text generation, function calling or thinking. Google says it returns answers in 150 to 250 ms end to end and is nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API on its own benchmark suite, and that Gemini Notebook, Stitch and Search already use it for model routing and safety checks; those figures are Google's. Artificial Analysis says it scores 66 on its Decision Index v2.1. Google says integrations are rolling out in OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, with Google Cloud support coming later. If your agent spends generative calls on routing, tool selection, policy checks or grading outputs against a rubric, test this as a cheaper, faster classifier layer on your own labelled examples before swapping it in.
Source: Google: Gemini Decisions API ↗Gemini API docs: Gemini Decision model ↗Gemini API pricing: Gemini Decision ↗@ArtificialAnlys on X: Gemini Decision 1 on the Decision Index ↗
Read the full write-up about Google launches Gemini Decision 1 and the Decisions API
Gemini 4 Carbon is now open to developers on the Gemini API (opens in new tab)
Why it matters: Google has opened Gemini 4 Carbon, its strongest frontier model, to all paying developers and enterprises, two weeks after a cyber-defenders-only release. You can call it today as gemini-4-carbon on the paid tier of the Gemini API in Google AI Studio, Android Studio and Gemini for business, at an introductory $2 per million input tokens and $10 per million output tokens until November 30, 2026, then $4 and $20. The model card lists text, image, audio and video input, text output and a shared 1M-token input-plus-output window. Google positions Argon as the default for most coding and knowledge work and Carbon for the hardest multi-step coding, debugging and research jobs, noting it spends more tokens than Argon to reason and verify. Google is also previewing /swarm in Google Antigravity, which turns one prompt into a coordinated team of agents, with Carbon in the planning, coordination and audit roles and lighter models doing parallel work. The benchmark claims are Google's own: it says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2x lower cost per task than the next best outside model, and the cyber, safety and prompt-injection figures are Google-reported too. The strongest cyber version stays behind the Fairwind Program for verified defenders. If you run long coding agents, test Carbon as the planner or reviewer over Argon or a cheaper model on your own tasks, and watch token use, since the price doubles after November 30.
Source: Google: Gemini 4 Carbon is now available ↗Google DeepMind: Gemini 4 Carbon model card ↗Google for Developers: Gemini 4 Carbon for developer agents ↗@GoogleAI on X: Gemini 4 Carbon is now available ↗
Read the full write-up about Gemini 4 Carbon is now open to developers on the Gemini API
Anthropic releases Claude Fable 5.2 and Mythos 5.2 (opens in new tab)
Why it matters: Anthropic has released Claude Fable 5.2, available today to all users and on the Claude Platform as claude-fable-5-2, alongside Claude Mythos 5.2, the same underlying model without Fable's cyber and biology safeguards and open only to vetted organisations through Anthropic's Cyber and Life Sciences Verification Programs. Pricing is lower than the 5.1 models at $8 per million input tokens and $40 per million output tokens, with an optional fast mode (up to 2.5 times faster generation) at $16 and $80. Fable 5.2 is also on Amazon Web Services, Google Cloud and Microsoft Azure, and fast mode works in Claude Code, the Claude Platform, Claude on Google Cloud and Microsoft Foundry. Anthropic still recommends starting with Claude Opus 5.5 for most coding and knowledge work and moving up to Fable for the hardest open-ended jobs, or using Fable as an Advisor that smaller models consult or an Orchestrator that hands simpler tasks to them. The benchmark claims are Anthropic's own: it says Fable 5.2 at High effort matches Opus 5.5's best FrontierCode 2.0 score (45.3% vs 45.5%) for about half the cost per task, reaches 49.1% at Xhigh effort, and generates output up to 30% faster than Fable 5.1. Artificial Analysis independently ranks Fable 5.2 (Max) first on its Intelligence Index with a score of 65, but also flags it as slow and verbose, at about 46 minutes and $26.54 per index task, with a 1M-token context window. Fable's safeguards still block most cybersecurity work such as penetration testing and exploit generation (blocked queries are answered by an older Opus model instead), and preserved thinking now applies to all API accounts, so check Anthropic's docs if your integration edits earlier assistant turns. If you run long multi-agent coding or research jobs, test Fable 5.2 as the orchestrator or advisor over cheaper models rather than as your default, and watch per-task cost at high effort levels.
Source: Anthropic: Claude Fable 5.2 and Mythos 5.2 ↗Artificial Analysis: Claude Fable 5.2 ↗@claudeai on X: Claude Fable 5.2 and Mythos 5.2 ↗
Read the full write-up about Anthropic releases Claude Fable 5.2 and Mythos 5.2
Meta releases Muse Ember, a low-cost model for agent subtasks (opens in new tab)
Why it matters: Meta has released Muse Ember, its smallest and most efficient model, built for the repeat work inside agents: tool calls, coding subtasks, routing, extraction and long-document parsing. It is available today on the Meta Model API and in Muse Code as muse-ember-1.0, at $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. It has a 1M-token context window, up to 512K output tokens, supports the Responses API and takes text, images, audio and video as input. The comparisons are Meta's own, citing Artificial Analysis Intelligence Index v5 results published the same day: Meta says Ember matches Claude Haiku 5.5's best score at about a third of the cost and scores 36 on the index against 37 for Muse Spark 1.3, at roughly $0.14 versus $2.39 per index task; Meta's own table also shows Haiku 5.5 ahead on FrontierCode 2.0. Meta suggests pairing it with Muse Wonder, which plans while Ember handles the volume. If you run a coding agent or workflow harness with lots of small subagent calls, Ember is worth a cost test as the subagent model; check results on your own tasks before switching, since Artificial Analysis has not yet posted its own write-up.
Source: Meta: Introducing Muse Ember ↗@AIatMeta on X: Muse Ember ↗@AIatMeta on X: pricing and API availability ↗
Read the full write-up about Meta releases Muse Ember, a low-cost model for agent subtasks
Newsletter
Get the best new A.I. picks in your inbox
The best new agent workflows and tools we’ve vetted, with a clear verdict on each. Short, occasional, useful.