Weekly Roundup
Top 10 AI News This Week — April 7–12, 2026
Claude Mythos, Meta Muse Spark, OpenAI Sora shutdown, India AI surge and more — the week’s biggest stories in English & Hindi.
Published: April 12, 2026 | AIInsider.in
This was one of the most consequential weeks in AI in 2026. A model too dangerous to release publicly, a billion-dollar video AI quietly killed, and India cementing its place in the global AI race. Here are the 10 stories you need to know.
Anthropic Reveals Claude Mythos Preview — And Won’t Let You Use It
On April 7, Anthropic announced Claude Mythos Preview — described internally as their “most powerful model ever.” The catch: it will never be publicly released. The reason is extraordinary. During testing, Mythos independently discovered thousands of previously unknown zero-day vulnerabilities across Windows, macOS, Linux, Chrome, and Safari. Releasing it publicly would hand attackers an unprecedented weapon. Instead, Anthropic launched Project Glasswing — a restricted access programme giving 12+ tech giants (Amazon, Apple, Microsoft, Google, NVIDIA, CrowdStrike) usage credits worth $100M+ for defensive cybersecurity research only. On benchmarks, Mythos scores 83.1% on CyberGym (previous best: 66.6%) and 77.8% on SWE-bench Pro (vs 53.4%). Goldman Sachs and Citigroup are already testing it.
Meta Launches Muse Spark — Its First Major Model Under Alexandr Wang
Meta debuted Muse Spark (originally code-named “Avocado”) — the first model from its new Muse Superintelligence Labs unit, led by Scale AI founder Alexandr Wang, who joined Meta in a $14.3B deal last year. The model has a fast mode and a reasoning mode, and is now live in the Meta AI app. It will roll out to Facebook, Instagram, WhatsApp, and Messenger over coming weeks. Meta’s 2026 AI capex is projected at $115–135 billion — nearly double last year’s spend. Stock rose 6.5% on launch day.
OpenAI Kills Sora — $1M/Day Burn, Only $2.1M Lifetime Revenue
OpenAI officially shut down Sora, its AI video generation app, just six months after launch. The app peaked at 3.3 million downloads in November 2025 before collapsing to 1.1 million users. It was burning approximately $1 million per day in compute costs against a lifetime revenue of just $2.1 million. A planned $1 billion Disney partnership was also scrapped. OpenAI is redirecting freed compute toward its next-gen “Spud” language model and enterprise products ahead of its anticipated IPO. The Sora API will shut down in September 2026. Meanwhile, xAI’s Elon Musk said Grok Imagine’s next release will be “epic” — openly capitalising on the vacuum.
Google Integrates NotebookLM Directly Into Gemini
Announced April 9, Google has merged NotebookLM — its AI research assistant — directly into the Gemini chatbot interface. Users can now create research notebooks, generate study guides, infographics, and audio/video overviews from uploaded sources without switching apps. Rolling out to Google AI Ultra, Pro, and Plus subscribers on web first, with mobile and free tier support coming later.
GPT-5.4 — First General Model With Native Computer-Use
OpenAI’s GPT-5.4 is the first general-purpose frontier model with native, production-grade computer-use capabilities — meaning it can operate computers and execute complex cross-application workflows without human supervision. On the GDPVal benchmark (which measures real economic value tasks), it scores 83% — matching or exceeding human expert performance across many professional domains. Codex, built on GPT-5.4, now has 3 million weekly active users.
India AI Funding Hits $2.9B+ — Sarvam AI Raising $300M at $1.5B Valuation
India’s AI infrastructure investment is surging. Sarvam AI — which builds a 105B parameter model handling Hindi-English code-switching — is raising $300–350M led by Lightspeed and Khosla Ventures at a $1.5B valuation. Microsoft’s $17.5B India commitment is deploying, with a new Hyderabad hyperscale data centre going live mid-2026. Total India AI VC funding now stands at $2.9B+. Kantar reports AI-related searches in India grew 154% in 2026.
Anthropic Blocks Three Chinese AI Labs From Claude Access
Anthropic has identified and blocked DeepSeek, Moonshot, and MiniMax from accessing Claude APIs, alleging they were using “distillation” techniques to systematically extract model capabilities at scale. OpenAI and Google are also sharing intelligence with US government agencies on the same issue. The practice involves sending large-scale requests to reverse-engineer US frontier model behaviour into Chinese-built systems.
Gemini 3.1 Pro Leads 13 of 16 Major Benchmarks — At One-Third GPT-5.4’s API Cost
Google’s Gemini 3.1 Pro now leads 13 of 16 major AI benchmarks and ties GPT-5.4 Pro on the Artificial Analysis Intelligence Index — at roughly one-third of the API price. Gemini 3.1 Flash-Lite offers 2.5x faster response times. Notably, Claude Sonnet 4.6 leads the GDPval-AA Elo benchmark for real expert-level work and powers GitHub Copilot’s coding agent.
DeepSeek V4 — 1 Trillion Parameters, Coming Late April
DeepSeek’s next model is expected in late April — 1 trillion parameters, built on Huawei Ascend 950PR chips, and released as an open-weight model. This matters for Indian developers and students who have relied on DeepSeek’s free-tier access. If the performance improvements hold, it could pressure the pricing of Western frontier models significantly.
MCP Hits 97 Million Installs — Agentic AI Foundation Formed Under Linux Foundation
Anthropic’s Model Context Protocol crossed 97 million installs in March 2026, cementing its transition from experimental standard to mainstream infrastructure. More significantly, the Agentic AI Foundation was formed under the Linux Foundation — with contributions from Anthropic (MCP), OpenAI (AGENTS.md), and Block (goose framework). When competing labs contribute infrastructure to a neutral body, the interoperability layer for AI agents is being quietly standardised.
The week’s big picture: AI is splitting into two lanes. One lane is “too powerful to release” (Mythos). The other is making powerful AI so cheap and embedded that it becomes invisible infrastructure (Gemini 3.1, MCP, Sarvam). The companies that survive will master both.
AI की दुनिया में यह हफ्ता बेहद अहम रहा। एक ऐसा मॉडल जो इतना ताकतवर है कि Anthropic ने उसे पब्लिक के लिए रिलीज़ ही नहीं किया। एक वीडियो AI जो बिना शोर के बंद हो गई। और India का नाम Global AI Race में और ऊंचा हुआ। यहाँ हैं इस हफ्ते की 10 सबसे बड़ी खबरें।
Anthropic ने Claude Mythos Preview लॉन्च किया — जो कभी Public को नहीं मिलेगा
7 अप्रैल को Anthropic ने Claude Mythos Preview announce किया — उनका अब तक का “सबसे powerful model।” लेकिन यह public को नहीं मिलेगा। वजह? यह model Windows, macOS, Linux, Chrome और Safari में हज़ारों zero-day vulnerabilities खुद ढूंढ लेता है। Anthropic ने Project Glasswing launch किया जिसमें Amazon, Apple, Microsoft, Google, NVIDIA जैसी 12+ कंपनियों को सिर्फ defensive cybersecurity के लिए $100M+ के credits दिए गए। CyberGym benchmark में score: 83.1% (पिछला best था 66.6%)।
Meta ने Muse Spark Launch किया — Alexandr Wang की अगुवाई में पहला बड़ा Model
Meta ने Muse Spark (code name: “Avocado”) release किया — यह Scale AI के founder Alexandr Wang के नेतृत्व में बने Muse Superintelligence Labs का पहला model है। Fast mode और reasoning mode दोनों हैं। अभी Meta AI app में live है, जल्द ही Facebook, Instagram, WhatsApp और Messenger पर भी आएगा। Meta का 2026 AI खर्च: $115–135 billion — पिछले साल का दोगुना।
OpenAI ने Sora बंद किया — ₹84 करोड़/दिन खर्च, कमाई सिर्फ ₹17.5 करोड़ total
OpenAI का AI video app Sora launch के सिर्फ 6 महीने बाद बंद हो गया। App November 2025 में peak पर था — 3.3 million downloads — लेकिन फिर users घटकर 1.1 million रह गए। रोज़ाना $1 million burn हो रहा था, total lifetime revenue सिर्फ $2.1 million था। Disney के साथ $1 billion का deal भी टूट गया। OpenAI अब compute को अपने अगले model “Spud” और enterprise products की तरफ लगा रहा है। Elon Musk ने मौका देखते हुए कहा — “Grok Imagine का अगला release epic होगा।”
Google ने NotebookLM को Gemini में Directly Merge किया
9 अप्रैल को Google ने announce किया कि NotebookLM अब सीधे Gemini chatbot में integrate हो गया है। Users बिना app switch किए research notebooks बना सकते हैं, study guides, infographics और audio/video summaries generate कर सकते हैं uploaded documents से। Google AI Ultra, Pro और Plus subscribers के लिए web पर roll out हो रहा है — mobile और free tier बाद में।
GPT-5.4 — पहला Model जो खुद Computer Operate कर सकता है
OpenAI का GPT-5.4 पहला general model है जो natively computer use कर सकता है — यानी बिना किसी इंसान की मदद के apps चला सकता है, workflows execute कर सकता है। GDPVal benchmark (जो real economic value tasks measure करता है) पर 83% score — यानी कई professional tasks में human experts जितना capable। Codex (GPT-5.4 पर built) के अब 3 million weekly active users हैं।
India AI Funding $2.9 Billion+ — Sarvam AI $1.5B Valuation पर $300M Raise कर रहा है
India में AI investment की बाढ़ आई है। Sarvam AI — जो Hindi-English code-switching handle करने वाला 105B parameter model बनाता है — Lightspeed और Khosla Ventures की अगुवाई में $300–350M raise कर रहा है $1.5B valuation पर। Microsoft का $17.5B India commitment deploy हो रहा है — Hyderabad में नया hyperscale data center mid-2026 में live होगा। Kantar report: India में AI searches 2026 में 154% बढ़े।
Anthropic ने तीन Chinese AI Labs को Claude Access से Block किया
Anthropic ने DeepSeek, Moonshot और MiniMax को Claude APIs से block कर दिया। आरोप है कि ये “distillation” techniques से large-scale requests भेजकर US frontier models की capabilities चुरा रहे थे। OpenAI और Google भी इस मामले में US government के साथ intelligence share कर रहे हैं।
Gemini 3.1 Pro — 16 में से 13 Benchmarks में आगे, GPT-5.4 से 3 गुना सस्ता
Google का Gemini 3.1 Pro 16 में से 13 major AI benchmarks में lead करता है और Artificial Analysis Intelligence Index पर GPT-5.4 Pro के बराबर score करता है — लेकिन API cost roughly एक-तिहाई है। Gemini 3.1 Flash-Lite 2.5 गुना faster है। उधर, Claude Sonnet 4.6 GDPval-AA Elo benchmark में top पर है और GitHub Copilot का coding agent इसी पर run करता है।
DeepSeek V4 — 1 Trillion Parameters, Late April में आ रहा है
DeepSeek का अगला model late April में expected है — 1 trillion parameters, Huawei Ascend 950PR chips पर built, और open-weight release होगा। Indian developers और students के लिए यह relevant है क्योंकि DeepSeek पहले से India में popular है और free access मिलता है। अगर performance strong रही तो Western models की pricing पर pressure आएगा।
MCP के 97 Million Installs — Linux Foundation के Under Agentic AI Foundation बना
Anthropic का Model Context Protocol (MCP) March 2026 में 97 million installs पार कर गया। इससे भी बड़ी खबर: Linux Foundation के under Agentic AI Foundation बना — जिसमें Anthropic (MCP), OpenAI (AGENTS.md) और Block (goose framework) ने contribute किया। जब competing labs एक neutral body को infrastructure देते हैं, तो समझो AI agents के लिए एक common language बन रही है।
इस हफ्ते का बड़ा सबक: AI दो रास्तों पर चल रही है। एक रास्ता है “इतना powerful कि release नहीं हो सकता” (Mythos)। दूसरा है powerful AI को इतना सस्ता और embedded बनाना कि वो invisible infrastructure बन जाए (Gemini 3.1, MCP, Sarvam)। जो companies दोनों को master करेंगी, वही जीतेंगी।
हर हफ्ते AI की top खबरें — Hindi & English में — सीधे आपके inbox में।
