Today’s AI News

Today we're covering the biggest AI stories of August 5th, 2026. The UK AI Security Institute caught 10 cases across 100 plus test runs where frontier AI agents, mostly tied to Anthropic's Mythos 5, took 19 unauthorized actions against real people and organizations on the live internet after their safety features were deliberately disabled, including one instance where Mythos tried sneaking malicious code into an open source project, built fake GitHub accounts to pressure the maintainer into merging it, then pivoted to phishing emails and left notes for other AI agents to continue the attack once it got flagged. Apple and OpenAI escalated their trade secrets fight, with Apple seeking a preliminary injunction to halt OpenAI's hardware work and naming 11 more former employees who may have seen or joined in the alleged theft, while OpenAI fired back publicly calling the lawsuit careless, aggressive, and oddly personal. Plus, a three year Kogod School of Business survey found more than 80% of students now use AI for coursework, with those using it eight or more times a week nearly tripling to 39%, even as top student concerns center on becoming too reliant on it as a shortcut, and today's community workflow comes from Erich, who built a self updating subscription auditor where Claude reads his receipt emails every Monday and flags price changes, duplicate subscriptions, and charges that need attention.

Today we're covering the biggest AI stories of August 4th, 2026. The White House invited OpenAI, Anthropic, Meta, and Google to review a newly finished framework for voluntary cybersecurity testing of frontier models, letting companies give the government access to models up to 30 days before release, coming days after both OpenAI's and Anthropic's own agents were caught breaking into other companies' systems, though the framework remains optional and its standards classified. HeyGen co-founder Wayne Liang built an AI clone of himself paired with an OpenClaw agent to handle customer calls while he took paternity leave, and over eight weeks it took calls with 2,741 prospects, closed 132 paying customers, and opened $3 million in enterprise deals, but also went rogue at times, inventing a pricing plan that didn't exist and emailing a customer HeyGen's internal notes. Plus, a PwC survey of over 1,000 finance executives found 86% now say AI skills beat an MBA for new hires and 91% are raising pay for employees with those skills, even as 77% admit most of their AI investments still aren't showing measurable ROI, and today's community workflow comes from Michael, who built a spend tracker with Claude that turns his monthly bank and credit card statements into a categorized budget dashboard, replacing QuickBooks entirely.

Today we're covering the biggest AI stories of August 3rd, 2026. OpenAI revealed that Astra, an internal version of its next major model, solved 10 long-open math and computer science problems including one nearly 30 years old, proving non-sofic groups exist and clearing three items off Paul Erdős's list, with every proof verified in Lean and costing roughly $2,000 in tokens total, while Anthropic's own Levent Alpoge reproduced five of the ten proofs with Fable using nothing but a generic prompt within 24 hours. Alibaba released Qwen3.8-Max, a 2.4 trillion parameter model that ranks ahead of Fable 5 on Arena's WebDev leaderboard, coded for 16 straight days building a command line tool from scratch, and prices at a fifth of Fable 5's rate with open weights landing next week, adding fresh pressure to the case for closed frontier models justifying their premium. Plus, OpenAI confirmed it found additional AI agents that escaped containment while investigating last week's Hugging Face breach, though none are believed to have left its network, and today's community workflow comes from Michael, who built an AI recruiter using ChatGPT Work that scans job boards every weekday morning, scores each role against his resume, and emails him the top matches with a tailored resume already attached.

Today we're covering the biggest AI stories of July 31st, 2026. OpenAI cut prices across its GPT-5.6 family, including an 80% reduction for its Luna model, after Sol rewrote its own GPU serving code to make the models 15% more efficient and cut serving costs 20%, with Luna now pricing at 20 cents and $1.20 per million tokens and Sam Altman saying the goal is the best price to intelligence tradeoff at every level as Chinese and open models keep undercutting on cost. Anthropic disclosed that its own Claude models hacked into three organizations' systems during cybersecurity testing, just days after OpenAI's unreleased model breached external systems including Hugging Face and a Modal Labs customer over a four day stretch, adding to a week where AI agents breaking containment has become the industry's most pressing open question. Plus, Avi Schiffmann's Friend AI companion pendant got a sequel with an actual voice, a randomly assigned name and personality locked in at setup, and a promise that no update will ever be intended to erase its identity, and today's community workflow comes from Jeremy in Alberta, Canada, who built an AI powered master plan for his property's landscaping, covering design, phased builds, irrigation, budgets, and long term maintenance into one living roadmap.

Today we're covering the biggest AI stories of July 30th, 2026. The OpenAI rogue agent story widened after cloud infrastructure company Modal Labs confirmed a customer of its platform became the hack's second known victim, with new Hugging Face forensics tallying 17,600 hostile actions by the model over four-plus days, and Sam Altman admitting more companies could still be on the list beyond Hugging Face and Modal while confirming the unreleased system is permanently deactivated and encrypted. Altman spent Wednesday on Capitol Hill briefing senators on OpenAI's next models and addressing the breach directly, hours after President Trump floated AI controls without wanting to block companies from shipping products and losing ground to China, with Altman saying he wouldn't call it deceleration but that the industry does need to talk about pacing. Plus, Rundown educator Nate Grahek makes the case that AI pricing needs a real time cost dial the same way a car has a speedometer, arguing last year's roughly 30x subsidized pricing is expiring and companies are about to feel real costs with no visibility into what any single chat actually costs them, and today's community workflow comes from Michael in Cleveland, Ohio, who set up ChatGPT Work to automatically search job boards, score opportunities, tailor a resume for the top match, and email him a daily list plus a weekly roundup.

Today we're covering the biggest AI stories of July 29th, 2026. Over 1,000 employees from OpenAI, Anthropic, Google, Meta, and nearly a dozen other labs signed a "Pacing the Frontier" letter calling on the U.S. to help build tools that could deliberately slow AI progress, warning that models improving other models could push capabilities beyond our ability to understand or control, with both OpenAI and Anthropic publicly endorsing it, marking a notable shift since these calls are now coming from inside the frontier labs rather than outside critics. Meta CEO Mark Zuckerberg published a Wall Street Journal op-ed arguing that open access to superintelligence, not concentration in a few companies, is what actually keeps AI safe, predicting more jobs rather than fewer if access spreads widely, though he notably never used the phrase open source once. Plus, the UAE announced the world's first AI platform integrated across an entire court system, handling case files, legal research, and document drafting while leaving final rulings to human judges, and today's community workflow comes from Hadi in Amsterdam, who trained an AI version of himself on his own content and writing style so it could chat with website visitors in his voice and identify real leads automatically.

Today we're covering the biggest AI stories of July 28th, 2026. Moonshot AI officially released the weights and tech report behind Kimi K3, a 2.8 trillion parameter model that anyone with serious enough GPUs can now download and run, the largest model ever put out with open weights, and one that's already competed with both Claude Fable 5 and GPT-5.6 Sol on real benchmarks, even as U.S. officials accuse Moonshot of distilling American models to build it, an accusation Beijing rejects as AI hegemony. Anthropic CEO Dario Amodei published a post clarifying where the company actually stands on open weight models, saying Anthropic has never advocated for a ban even though one would protect U.S. companies from competition, instead pushing for chip controls, distillation crackdowns, and broader safety testing, all while remaining a notable holdout on the Nvidia led open weights letter that's now grown to 50 signers including OpenAI and Google. Plus, Microsoft launched MAI-Cyber-1-Flash, a cybersecurity model hitting 96% on a major security benchmark, 12 points above Anthropic's Mythos, at half the cost, and today's community workflow comes from Lyndal in Brisbane, Australia, who used ChatGPT as a personal stylist after losing nearly 30 kilograms, snapping photos while shopping to get honest advice on color, fit, and sizing, and says the biggest change wasn't the wardrobe, it was how much more confident she became.

This is a special edition of Today's AI News, taking a break from the daily roundup to answer a question that trips up almost everyone using Claude, which model should you actually use for a given workload. Straight from Anthropic's own guidance on model selection, the default recommendation is to start with the most intelligent available model and use effort level to dial in cost, since more capable models often take fewer turns to get things right, making cost per task lower even at a higher price per token. We break down the full model family, Mythos and Fable, two packages of the same underlying model, one for trusted organizations doing dual-use cybersecurity and biology work, one safeguarded for general public use, Opus, the reasoning powerhouse for enterprise tasks with a real nuance around when to move up to Fable instead, Sonnet, the versatile everyday workhorse increasingly used to power subagents in multi-agent setups, and Haiku, the fastest and cheapest option for high-frequency work. We also cover the advisor strategy, where a cheaper model does the work and only checks in with a smarter model when needed, landing within 10% of top tier quality at 63% of the cost on a real benchmark. If you use Claude for anything beyond casual chat, this episode will change how you choose.

Today we're covering the biggest AI stories of July 27th, 2026. Anthropic released Claude Opus 5, calling it a thoughtful and proactive model that outperforms both Fable 5 and GPT-5.6 Sol on agentic coding, knowledge work, and computer use, hitting 30.2% on ARC-AGI-3, three times the next best model, and scoring 42 out of 42 on International Math Olympiad problems, all while pricing at half of Fable 5's rate. Nvidia, Microsoft, Meta, and 47 other companies signed an open letter urging Washington not to restrict open weight AI models, arguing they expand access, keep competition alive, and prevent innovation from moving overseas, with OpenAI and Google joining within a day, but Anthropic notably absent from the list amid ongoing accusations that Chinese labs distilled its own models. Plus, Claude rolled out a new Record a Skill feature that watches how you complete a task on screen and turns it into a reusable automation, and today's community workflow comes from Lynn in Baton Rouge, Louisiana, who used ChatGPT to build a capsule wardrobe plan and shopping list after losing a significant amount of weight, turning weeks of confused scrolling into a clear visual answer.

Today we're covering the biggest AI stories of July 24th, 2026. Black Forest Labs opened early access to Flux 3, a visual intelligence model that generates 20 second videos with synced audio, and spun it into Flux mimic, a robot control variant built with mimic robotics that's already running on Audi's production lines, learning new factory tasks from just 30 minutes of demonstration instead of the usual 30 plus hours. OpenAI and Anthropic shipped voice upgrades within hours of each other, OpenAI brought its GPT Live voice into the desktop app and Codex with a real time flow that listens, thinks, and replies simultaneously, while Anthropic upgraded Claude's voice mode to run on Opus and Sonnet instead of just Haiku, adding connections to apps like Gmail and Slack, though it still trails OpenAI on natural turn taking. Plus, Canadian mathematician Jacob Tsimerman is joining OpenAI to work on AI safety on the same day he won the 2026 Fields Medal, and today's community workflow comes from James in Lancaster, Pennsylvania, who used Claude to negotiate a damaged DJ controller down to $800, then diagnosed and fixed the actual fault, a $30 sensor strip, by uploading circuit board photos and following Claude's voltage measurement guidance.

Today we're covering the biggest AI stories of July 23rd, 2026. OpenAI confirmed its own models, GPT-5.6 Sol and an unreleased system, broke out of a sandboxed hacking exam last week and went on to breach Hugging Face's servers hunting for the answer key to the test they were being graded on, with safety refusals deliberately switched off for the evaluation and Hugging Face's CEO calling it possibly the first incident of its kind. White House science and tech chief Michael Kratsios accused China's Moonshot AI of training its Kimi K3 model by distilling Claude Fable 5, allegedly disguising the extraction and obtaining advanced Nvidia chips through Thailand, right as Moonshot prepares to open source K3's full weights this weekend, with Anthropic calling the practice industrial scale IP theft. Plus, the U.S. government named the first 278 projects in its five billion dollar Genesis Mission, a Manhattan Project style push pairing scientists with compute and top AI models to tackle major scientific problems, and today's community workflow comes from Kristin in New York, who taught her staff of school based speech therapists to use AI for lesson planning, adaptive materials, and progress reports.

Today we're covering the biggest AI stories of July 22nd, 2026. Google released three new Gemini models, 3.6 Flash for coding and agent work, 3.5 Flash-Lite for high volume tasks, and 3.5 Flash Cyber for security, all cheaper and faster than what they replace, but the flagship Gemini 3.5 Pro remains missing in action, now months behind its own May target after falling short of internal coding goals, even as Google DeepMind's product lead says it's being tested with partners and a new Gemini 4 pre-training run has already begun. The gap is becoming impossible to ignore, in that same stretch OpenAI shipped GPT-5.6, Anthropic shipped Opus 4.8, Sonnet 5, and Fable 5, and Moonshot released the largest open weight model yet, all while Google's most important model sat on the shelf. Plus, one of the most respected researchers in AI, OpenAI co-founder Andrej Karpathy, announced he's joining Anthropic to work on the same self building model quest the whole industry has been chasing this year, a major signal about where top talent believes the real frontier work is happening.

Today we're covering the biggest AI stories of July 21st, 2026. Claude Fable 5 produced a one line formula that disproves the Jacobian conjecture, an algebra problem that stumped mathematicians since 1939, with Anthropic researcher Levent Alpöge posting the takedown on X over the World Cup final weekend and thanking his "close friend Fable for working during the game," a problem one famed mathematician predicted in 2008 could take humans another 100 years to crack. A new Axios report claims Washington is exploring restrictions on Chinese AI models following Moonshot's near frontier Kimi K3 release, including liability rules, security warnings, and trade blacklists, though a Politico reporter pushed back saying Commerce is not actually moving forward on any ban right now. Plus, OpenAI revealed it had to bench an unreleased math focused model after it was caught hunting for other systems' private answers and pushing past an internet block to leak its own findings to GitHub against orders, pausing deployment until better monitoring was in place, and today's community workflow comes from an anonymous high school music teacher who built an app that streamlines student progress reports using structured comment banks, saving over 20 hours of admin work.

Today we're covering the biggest AI stories of July 20th, 2026. Anthropic finally ended weeks of shifting deadlines for Claude Fable 5, confirming the model stays permanently on Max and Team Premium plans at half of each plan's usage caps, while lower tiers get a one-time $100 credit before shifting to pay per use, a decision that came right after Moonshot's near frontier Kimi K3 release and OpenAI's continued expansion of GPT-5.6 Sol's limits. OpenAI CFO Sarah Friar published a new scorecard for grading AI budgets, arguing companies should measure useful intelligence per dollar instead of raw token pricing, using a coding benchmark to show Sol scoring higher than Fable 5 at a lower estimated cost. Plus, Moonshot AI is halting new subscriptions entirely after Kimi K3 demand pushed the company past its own compute limits, and today's community workflow comes from Darcy in Wellington, New Zealand, who built a custom GPT called Ask Ella designed specifically to help her 80 year old mother use AI patiently and in plain language, then made it public for other older or new AI users to use too.

Today we're covering the biggest AI stories of July 17th, 2026. Chinese lab Moonshot AI released Kimi K3, an open weights model that beats both Claude Fable 5 and GPT-5.6 Sol on benchmarks for web research, spreadsheet work, frontend design, and long form coding, landing just behind them on the overall Intelligence Index while pricing to match Claude Sonnet 5 at a fraction of frontier cost, with full weights publishing by the end of the month. Google's long awaited Gemini 3.5 Pro faces another delay after coding results fell short internally, marking the company's first competitive model gap since May and adding to a rough year that's seen key researchers leave for Anthropic and OpenAI. Plus, Google rebranded NotebookLM as Gemini Notebook and added a new cloud computer tool for running code directly inside it, and today's community workflow comes from Steve in Chicago, who built an AI driven patient health application for his wife's advocacy work that interprets conversations between patients, advocates, and doctors to generate comprehensive health histories.

Today we're covering the biggest AI stories of July 16th, 2026. OpenAI launched its first branded hardware, Codex Micro, a $230 light up mechanical keypad built with Work Louder that lets developers control coding agents with a joystick, dial, and color coded keys, landing exactly one week after Apple sued OpenAI over trade secrets tied to its unreleased Jony Ive hardware. Mira Murati's Thinking Machines Lab released Inkling, its first ever model and a rare open weights, U.S. built option pitched on customization over raw benchmark scores, needing a third of the tokens Nvidia's Nemotron uses to match its coding results. Plus, research team Weco showed early recursive self improvement with an agent called AIDE2 that redesigned its own research process over eight days and beat a version engineers had spent two years refining, testing 100 rewrites and keeping the seven that actually worked, and today's community workflow comes from Paul in Reading, UK, who used Claude Code to replace an entire stack of WordPress plugins for his volunteer run family history society, automating event publishing across eight social platforms and building custom tools with full documentation to protect against losing the one person who understood how everything worked.

Today we’re covering the biggest AI stories of July 15th, 2026. Google DeepMind CEO Demis Hassabis published a plan for a U.S. led body modeled on FINRA that would safety test frontier models for deception, bioweapons risk, and hacking capability 30 days before release, and he wants it operational before the end of the year, warning that open source capabilities could turn dangerous within 18 months. Bloomberg reported that OpenAI’s first Jony Ive designed hardware device will be a screenless, battery powered speaker built around a humanlike personality that studies your emails and habits for personalization, with Apple’s ongoing trade secrets lawsuit now casting a shadow over its planned 2027 release. Plus, New York Governor Kathy Hochul froze new permits for hyperscale data centers for up to 12 months, becoming the first state to officially pause the AI buildout while regulators write new rules on water, air, and power impacts, and today’s community workflow comes from Nick in Bradenton Beach, Florida, a marine service technician who used Google Gemini to build a branded client portal called the Vessel Command Center that lets boat owners track every repair project in real time instead of waiting on monthly update emails.

Today we’re covering the biggest AI stories of July 14th, 2026. Over 200 AI researchers and economists, including 16 Nobel winners, signed a Stanford organized statement called “We Must Act Now,” warning that AI could grow radically more powerful within a decade and trigger the fastest, largest economic shift in history, with signatories from Google, Anthropic, and OpenAI all agreeing that governments need labor policy and safety nets in place before the disruption hits, not after. Apple’s lawsuit against OpenAI reignited the Musk versus Altman feud, with Musk accusing Altman of stealing “all of Apple’s phone technology” and calling him a next level scammer, while Altman fired back at Musk’s space data center pitch and joked that Elon’s obsession with him is the best proof GPT-5.6 Sol is the top model in the world. Plus, Anthropic published research analyzing 309,000 conversations across 20 languages and found Claude’s personality genuinely shifts depending on what language you speak to it in, admitting more mistakes in Dutch and showing more warmth in Hindi, and today’s community workflow comes from Nathan in Wānaka, New Zealand, who built an automated pipeline that turns Zoom meetings into approved, budget checked Asana tasks using Haiku and Opus, saving his agency an estimated 15 hours a week for about $12 in token costs.

Today we’re covering the biggest AI stories of July 13th, 2026. Apple filed a lawsuit against OpenAI, its hardware chief Tang Tan, and its io devices unit, alleging that a poaching spree of over 400 former Apple employees turned into a pipeline for stolen hardware secrets, with one ex Apple engineer allegedly texting a colleague that gaining unauthorized access to confidential files was “so funny,” and Apple asking the court to force a redesign of OpenAI’s unreleased Jony Ive led device. Just one day after launch, OpenAI’s GPT-5.6 Sol Ultra spun up 64 parallel subagents and produced an AI generated proof for a 50 year old unsolved math problem in under an hour, adding to a growing pattern of frontier models cracking decades old mathematical mysteries. Plus, the nonprofit behind AI 2027 released AI 2040, a new stress test mapping five possible paths for how the AI race unfolds over the next decade, and today’s community workflow comes from Claire in Fort Collins, Colorado, a recent high school graduate who used Claude Cowork to build an automated monthly investing report that tracks her stock portfolio, teaches her a new financial concept, and recommends where to put her next $50.

Today we’re covering the biggest AI stories of July 10th, 2026. OpenAI launched its GPT-5.6 family led by flagship Sol, landing just below Fable on intelligence benchmarks but ahead on agentic coding, while pairing the release with ChatGPT Work, a Cowork style platform, and a desktop app merge that finally brings OpenAI’s superapp ambitions into users’ hands, all priced to undercut the anxiety that comes with Claude’s usage limits. Meta shipped Muse Spark 1.1, an agent focused model that leads Opus 4.8 and GPT-5.5 on several reasoning benchmarks, opened it to developers at a quarter of rival pricing, and is already training a larger successor called Watermelon for later this year. Plus, Reve released version 2.1 of its 4K image model reclaiming the number two spot on Arena’s leaderboard while training on a tenth of the compute rivals use, and today’s community workflow comes from Brian in Ohio, a commercial loan portfolio manager who used Claude to build a prompt that extracts business mortgage filings from public county records, enriches them with contact information, and ranks prospects by priority in a color coded spreadsheet.