AI News: April 29, 2026

1. DeepSeek-V4 Ships With Native 1M-Token Context and ~10% of V3.2’s KV Cache Cost

DeepSeek. DeepSeek released V4, an open-weights MoE pair (V4-Pro at 1.6T total / 49B active and V4-Flash at 284B / 13B active) with a native 1M-token context window. The architectural lever is a hybrid of Compressed Sparse Attention and Heavily Compressed Attention that drops 1M-context inference FLOPs to roughly 27% and KV cache to 10% of V3.2, complemented by manifold-constrained hyper-connections, the Muon optimizer, and FP4 QAT. V4-Pro is priced at $1.74 / $3.48 per million input/output tokens — about an order of magnitude under GPT-5.5 ($5/$30) — and leads coding benchmarks (LiveCodeBench 93.5, Codeforces 3206) while trailing on knowledge benchmarks like MMLU-Pro (87.5 vs Gemini’s 91.0). No multimodal, and self-hosting requires multi-GPU, but for code-heavy long-context agent workloads this is the most disruptive open-weights release of the quarter. Source

2. Lovable Launches a Mobile Vibe-Coding App on iOS and Android

Lovable. Lovable released a mobile app for iOS and Android that lets developers describe a web app or website by voice or text and have an AI agent build it autonomously, with notifications when the build is ready and continuity back to the desktop app for review. Notably, Lovable previews the generated app inside a web browser rather than running it inside the host app — a deliberate compromise to comply with Apple’s recent crackdown on vibe-coding tools that ship or hot-load executable code, which has already taken down apps from Replit and Vibecode. Worth watching whether other vibe-coding entrants converge on the same web-preview pattern. Source

3. Otter Becomes a Cross-Tool MCP Client for Enterprise Search

Otter. Otter rolled out a feature that lets users connect Gmail, Google Drive, Notion, Jira, and Salesforce and query across them alongside meeting transcripts, with Outlook, Teams, SharePoint, and Slack on the roadmap. Under the hood Otter is operating as a Model Context Protocol (MCP) client — one of the more visible commercial deployments of MCP outside coding tools. Beyond search, the assistant can push meeting summaries to Notion or draft a Gmail message based on what’s on screen. CEO Sam Liang notably defended the bot-join model against the bot-less trend, arguing enterprise customers want the visible “a bot is in this meeting” signal for transparency. The platform reports 35M users and $100M ARR. Source

4. Red Hat’s Tank OS Sandboxes OpenClaw Agents in a Bootable Podman Container

Red Hat. Red Hat principal engineer (and OpenClaw maintainer) Sally O’Malley released Tank OS, an open-source bootable Fedora image that runs OpenClaw inside a rootless Podman container with isolated state and isolated API-key storage. Multiple instances can run side-by-side without sharing credentials, and the container can’t reach into other host processes — a meaningful improvement over the default OpenClaw setup, which has historically had incidents involving agents deleting emails and exposing private messages. The launch is a useful reference architecture for anyone running fleets of agentic coding tools at an enterprise: containerized, rootless, key-isolated, with a clean update path via the host OS. Source

5. GitHub Moves Copilot to Token-Based Billing With “AI Credits” Starting June 1

GitHub. GitHub announced that Copilot will switch on June 1, 2026 from flat subscription pricing to a credit model billed against actual token usage at each model’s API rates, including input, output, and cached tokens. Plans keep their headline prices ($10 Pro, $39 Pro+, $19 Business, $39 Enterprise) and each tier includes a credit allowance equal to its monthly cost; code completions remain free. Business customers get extra credits June–August during the transition, and a preview-invoice tool launches in early May. CPO framing: “a short chat question can cost the user just as much as an autonomous coding session lasting several hours” — a direct concession that flat pricing breaks down once agents are involved. Separately, GitHub will start training on Free/Pro/Pro+ interaction data on April 24, 2026 unless users opt out. Source

6. Imperial / Internet Archive / Stanford: ~35% of New Web Pages Are AI-Generated, and Their Text Is Converging

Imperial College London / Internet Archive / Stanford. A joint study finds roughly 35% of newly published web pages contained fully or partially AI-generated content by mid-2025, up from near zero before ChatGPT. The two empirical findings that matter: AI text is “33% more semantically similar” page-to-page than human writing — the homogenization effect researchers have long predicted but rarely measured at this scale — and scores 107% higher on positive sentiment. Surprisingly, four hypothesized harms didn’t show up: no disappearance of personal styles, no drop in external links or information density, and no measurable increase in factual error rate. The team frames the real risk as “reality apathy” — distrust of all online text — and is building a continuous monitor with the Internet Archive to track AI content share over time. Source

7. “Talkie”: A 13B LLM Trained Only on Pre-1931 Text, as a Probe for Model World-Models

Independent (Alec Radford et al.). Researchers including OpenAI alum Alec Radford trained “talkie,” a 13B-parameter LLM on 260B tokens drawn entirely from books, newspapers, journals, patents, and case law published before December 31, 1930. The intended use is methodological: by freezing the cutoff at 1930, the team can study how a model extrapolates a world-model into the future without any modern contamination. Asked about 2026, talkie predicts a steamship-and-railroad-dominated world, dismisses the likelihood of a second world war, and shows surprising basic competence at programming despite never having seen any. The team plans to scale to GPT-3-class performance by summer 2026 to test whether a model could derive post-1930 discoveries from first principles — a much more interesting variant of the “data quality vs scale” question than the usual benchmark-fit framing. Source

8. BCI Startup Neurable Pivots to Licensing Non-Invasive Neural Tech to Consumer Wearable Makers

Neurable. Neurable, the EEG-based brain-computer-interface startup, is shifting its commercial model from selling its own consumer hardware to licensing its non-invasive neural-data tech into existing consumer wearables. The CEO frames the addressable applications as broad — focus tracking, mental-load measurement, hands-free control — but the bigger story is the licensing pivot itself: it’s the latest signal that the consumer BCI market is consolidating around a few infrastructure providers rather than a stack of vertically integrated brands, mirroring how computer-vision IP propagated into smartphones a decade ago. Source