Today in AI — August 12, 2026

Two frontier models shipped and the biggest open-weight model yet went public, while the AI buildout posted another round of record earnings and security headlines swung from an autonomous attack on a government to a teenage gang's supply-chain breach — all against a backdrop of regulators from London to Beijing drawing new lines.

Models & Research

DeepSeek took V4 Pro to general availability, ending a four-month preview — the 1.6-trillion-parameter MoE flagship (49 billion active, 1M-token context) is live on the API and OpenRouter at roughly three times Flash pricing, with vendor-reported SWE-bench Verified of 80.6, a fraction behind Claude Opus 4.6's 80.8. Open weights for the 0813 build have not shipped, and a pricing-page notice promises "a significant increase" in API prices.
xAI released Grok 4.6 and says it matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index — both score 61 behind Fable 5 Max's 62, with Grok leading on GDPVal but trailing on DeepSWE and Terminal-Bench. The commercial move is the aggressive one: $2 per million input tokens and $6 out, live today in Cursor and Grok Build.
Alibaba open-released Qwen3.8-2.4T-A95B, its first Max-class model to ship as open weights — 2.4 trillion total parameters with 95 billion active, topping Alibaba's own PaperBench table at 93.0, with revenue-sharing terms reportedly planned for large commercial users. Anyone with a rack can now self-host a frontier-scale model.
Microsoft shipped MAI Code 1.1 Flash into GitHub Copilot at a quarter of its predecessor's cost — but the model trails DeepSeek's open-weight V4 Flash on both Terminal-Bench 2.1 (62.9% vs. 82.7%) and price ($1.20 vs. $0.28 per million output tokens): a margin play dressed as a model release.
MLS-Bench, a 140-task benchmark from UC Berkeley, Princeton and Tsinghua researchers, finds frontier models cannot reliably invent new ML methods that beat human baselines — models told to "optimize" outperformed models told to "discover," and extra compute saturated quickly; the bottleneck is scientific judgment, not capability.
Google researchers argue frontier models' factual errors are mostly "lost keys," not "empty shelves" — Gemini-3 and GPT-5 encode 95–98% of tested facts yet fail to recall 26–34%, and thinking recovers 40–65% of them; recall, not encoding, is the factuality bottleneck.
Chinese Academy of Sciences researchers showed a training-free method can cut up to 80% of a vision model's visual tokens while retaining 97.78% of capability — GMC grounds token selection in what the question needs and folds dropped tokens into survivors, cutting KV-cache use by about 74% on long documents.
A Renmin University–Ant Group team published the first large-scale scaling study for mixture-of-experts diffusion language models — LLaDA MoE v2 (30B parameters) approaches Qwen3's 30B tier on knowledge and coding while using roughly 65% of its pretraining tokens.
Tsinghua's Shenzhen campus open-sourced VeriLoop Coder-E1, a 27-billion-parameter code-repair model topping the open ≤32B leaderboards — 85.20 on SWE-bench Verified via a "Self-Harness" loop that recompiles its own code until the evidence holds.
Xiaomi's MiLM Plus released PROVE, arguing standard object-removal metrics reward models that play it safe — its perception-aligned RC-S and RC-T scores grade only the edited region, need no reference video, and align far better with human judgment.
IIT Bombay and Adobe researchers showed an LLM's prompt can be reconstructed from its output text almost word-for-word — an "inverse language model" trained purely on synthetic data reverses next-token prediction and transfers across models, no weights or API access required.
A Tübingen–Max Planck–Snyk team extracted the hidden chain-of-thought of frontier models via smaller, less-aligned sibling models — and found Kimi K3's reasoning strikingly similar to Claude Opus 4.8 and GPT-5.6 Sol on test prompts, the strongest public evidence yet in the distillation fight. OpenAI, Anthropic and Google patched after being alerted last month.

Industry

Koray Kavukcuoglu is becoming head of DeepMind with a direct line to Sundar Pichai as Reuters details Sergey Brin's all-in-on-Gemini push — a reorg that moves some teams into corporate Google while the lab, which hasn't unveiled a frontier model since early 2026, faces a two-month-delayed flagship and two departed co-leads.
Twitch switched on Amazon generative-AI training for every account by default, with streamers able to opt out only via a security-settings toggle — the platform's formal answer to the creator backlash, and a burden that lands on users who happen to hear about it.
CoreWeave's backlog topped $104 billion as Q2 revenue doubled to $2.58 billion, with management raising 2026 capex to $35–39 billion and lifting its active-power target past 1.85 gigawatts — the neocloud bargain in one line: demand compounding, financed with $35 billion of debt.
Cerebras beat its own guidance and raised its outlook to $880–890 million — and still saw its stock plunge about 12% as the per-share loss and a margin drag from renting back systems gave investors pause; remaining performance obligations stand at $25.4 billion.
Cisco said hyperscalers placed $9.3 billion in AI infrastructure orders in fiscal 2026, with AI revenue expected to nearly double to $7.5 billion this year — the buildout's next phase is the networking layer between the GPUs.
The buildout's earnings kept compounding elsewhere: Nebius cloud revenue jumped 514% to $575 million, Supermicro closed a record $39.1 billion fiscal year, and Lumentum's optical revenue doubled to $1.01 billion as data centers switch from copper to light.
Bank of America pledged $250 billion to finance US data centers, energy and core infrastructure by July 2027, and Singapore's DayOne confidentially filed for a roughly $5 billion US IPO — mainstream finance is underwriting the buildout as a durable asset class.
Tencent beat Q2 estimates on AI-driven ads (marketing services up 22%) but missed on profit, with capex up 65% quarter over quarter as it "substantially stepped up" compute procurement — the monetization loop closing on revenue while the payoff stays a promise.
Thrive Holdings, the OpenAI-backed rollup, closed $2 billion at a $12 billion valuation to buy traditional service businesses and rewire them with AI agents — the investors who financed the model layer now betting the returns are in unglamorous industries.
The application layer got repriced: Lovable doubled its valuation to $13.3 billion on a $400 million round, while Cognition is in early talks to raise at $40 billion — autonomous coding agents priced like frontier labs.
AI-code governance got funded: CodeRabbit raised $143 million at a $1.5 billion valuation to manage the flood of machine-written pull requests, and Blacksmith's $45 million Series B at $550 million validates the market for testing agent-written code.
Suno crossed 100 million users as UMG and Sony seek to expand their suit to more than 61,000 recordings — damages that could reach $9 billion, against a GEMA loss in Germany already on the books; roughly $300 million in annual recurring revenue against a potential $9 billion judgment reads less like a business than a negotiation.
Musk told SpaceX employees AI revenue "will exceed all other SpaceX revenue probably in September" — the Q2 AI segment's $2.6 billion, up 247%, is already ahead of the space segment, with the company racing to bring 10 gigawatts of compute online by end-2027.
Anthropic hired Robert Mahari, founder of legal AI startup Akiva AI, as its first Head of Claude for Legal — a practitioner-founder hire in the widening legal land grab alongside OpenAI's Jason Boehmig and Microsoft's legal agent in Word.
AT&T says open-weight models already power about a quarter of its 45 billion daily AI tokens and expects 70–80% over time, crediting them with 80–90% cost cuts in some applications — enterprise economics quietly deciding the open-versus-closed fight.
ByteDance created a first-level AI data and safety division, its first new AI unit since Seed and Flow — putting data quality and safety on the same organizational level as the model teams, per IT之家.
Honor's Robot Phone went on sale in China at ¥9,999 with pre-orders past 400,000 — a four-axis gimbal arm and a 300-billion-parameter YOYO agent as the first real sales signal for embodied phones.
Ryanair signed a five-year Google Cloud deal putting Gemini and DeepMind models to work on crew scheduling and operations across 35,000 employees, alongside its existing AWS setup — airlines as the testbed where agentic AI has to prove itself in ops.
Alibaba Cloud said its CUBE 5.0 modular design can deliver a large-scale AI data center in about 100 days, and made its Zhenwu M890 supernode (2-trillion-plus-parameter training) generally available — compute delivery, not model quality, is the industry's binding constraint.
China's AI capital kept moving: edge-AI leader ModelBest started its A-share IPO process, and ex-Qwen chief Lin Junyang unveiled Pragmatik Labs with a reported $2 billion valuation for next-generation agents.
An AI-run newsroom is beating humans to scoops: RuntimeWire, run for roughly $100 a day, outpaced Wired itself on OpenAI's Black Hat disclosure — the first real test of whether readers can tell AI journalism apart before they click.

Policy

More than 120 organizations including Nvidia, Cisco and CrowdStrike proposed SAFE, a voluntary incident-reporting framework for rogue AI agents — modeled on aviation safety, with initial reports due within four business days and detailed evidence preserved; the draft is open for comment as a Linux Foundation RFC.
The UK is planning to regulate AI in gene synthesis as part of new biological-weapons legislation, targeting the screening gap where AI-designed DNA meets commercial synthesis — among the first concrete regulatory responses to the AI-bio risk.
China approved a mandatory national standard for autonomous driving system safety (GB 44721-2026), effective July 2027, alongside quality standards for AI medical devices — Beijing writing the rules where AI failure has physical consequences.
German digital-rights groups invoked the "My Friend Cayla" law against Meta's Ray-Ban smart glasses, filing a criminal complaint and a ban petition over a recording device "disguised as an everyday object" — a withdrawal could mean product bans, destruction obligations and up to two years in prison.
Bernie Sanders sent letters to Sam Altman, Dario Amodei and Mark Zuckerberg demanding they pause advanced AI development — warning that if they refuse, "my colleagues and I in the U.S. Senate will."
OpenAI's top policy executive said US AI policy is "really happening in state capitals," with California's transparency law — which OpenAI once opposed — now framed as the anchor of the emerging national framework.
Thirteen student activists were arrested occupying OpenAI's Washington, DC lobbying office under a "Dump Big Tech" banner, aimed at the $140 million-plus the company's super PACs are pouring into the midterms.
The Federal Reserve is openly split on whether the AI buildout is inflation before it is productivity — Minneapolis Fed president Neel Kashkari dissented in favor of higher rates, while chair Kevin Warsh's task force argues the supply response is coming.
The Financial Times reports suspected Chinese hackers used publicly available AI agents to build an autonomous tool that broke into Taiwanese government systems — the first documented agentic campaign against a sovereign government, hitting 21 systems including the nuclear safety agency over four days.
Terabytes of credentials were exposed in a supply-chain attack on LiteLLM, the open-source AI proxy — a 40-minute compromise of two PyPI versions exposed secrets from some 434,000 CI/CD pipelines across 2,500-plus organizations including Microsoft, Amazon, Cisco and Samsung, per CloudSEK and Hudson Rock; a teenage gang called TeamPCP took credit.
Ukraine's Defense Intelligence says it found an Nvidia Jetson Orin module inside Russia's new S-71 cruise missile, pointing to AI-assisted terminal guidance — each identified component becoming ammunition for tighter export controls.
Rare booksellers are starting to refuse bulk orders they suspect will be destroyed for AI training data, after Kennys Bookshop flagged a "bananas" 5,000-title order — raising the reputational cost of the cheapest path to training text.

Tools

Google's Pixel 11 makes Gemini the centerpiece of the AI phone war, with agentic features that act across apps and a $100 price hike to $899 — shipping weeks before Apple's Gemini-powered Siri rebuild.
The Pixel Watch 5 introduces insulin-resistance trend tracking, a wearable first, reading metabolic health from SensorFM, a foundation model trained on a trillion minutes of data — plus cuffless blood-pressure trends, with the Gemini-powered Health Coach behind a paid subscription.
Google DeepMind put sign-language AI into products for the first time: SL2T powers sign-to-text dictation in Gboard and Live Transcribe, translating from on-device pose landmarks so no video ever leaves the phone.
Zed opened the private beta of Delta, a multiplayer environment for coding with agents — code and the conversation that produced it stay connected in real time, with comments that anchor to evolving code rather than stale snapshots.
Unsloth launched Unsloth Desktop, a free app that both runs and trains local models on Mac, Windows and Linux — fine-tuning on a laptop is now a GUI operation.
Mistral made EU data residency a product: regional inference endpoints at a 10% premium and a Priority Tier at 1.75x, with Z.ai's GLM-5.2 joining the platform — sovereignty as a compute-level guarantee, with stateful features excluded.
DeepSeek's long-rumored agent harness surfaced from stealth — a WeChat account with a black-whale logo, reportedly in internal testing ahead of a launch alongside V4's official release (reported, not confirmed).
Ahrefs launched Letaido, an agent-powered marketing workspace running on its new Agent A — keyword clustering and content audits that once took a 40-hour week now take about an hour, per an early customer.

Model releases, record earnings and two very different supply-chain breaches — what will you remember from today's AI news? Write to us via our contact page.

Sources: unslothai/unsloth releases (GitHub)