☕️ SpaceXAI launches Grok 4.5

Hi there, this is your daily ☕️ Techpresso.

In today's Techpresso:

💥 SpaceXAI launches Grok 4.5

💻 Meta enters the AI coding race

🎙️ ChatGPT can now listen and talk at the same time

📊 Brown professor's chart exposes mass AI cheating in exams

🚨 US warns self-driving cars are a danger to the public

🤖 Mistral's first robotics model steers robots with one camera

🎁 + 13 other news you might like

🧰 + 5 trending tools

📚 + 4 trending papers

💥 SpaceXAI launches Grok 4.5 LINK

SpaceXAI has released Grok 4.5, its first model since going public, and set its price at half that of rival models by charging $2 per million input tokens and $6 per million output tokens.

The company says Grok 4.5 has "twice greater token efficiency" than other leading models, and can handle coding, app-building, office work, research, and writing, though its benchmarks fell just short of best-in-class.

Elon Musk called Grok 4.5 "roughly comparable to Opus 4.7, but much faster," noting that Anthropic's Opus 4.7 costs $5 per million input tokens and $25 per million output tokens.

💻 Meta enters the AI coding race LINK

Meta is challenging OpenAI and Anthropic in the AI coding market with Muse Spark 1.1, which AI chief Alexandr Wang called the company's strongest model yet for agentic and coding work.

The new model's API opens through a developer portal as a public preview, where users can join a waitlist, though Meta is limiting access to its own properties rather than third-party platforms like OpenRouter.

Every new API account starts with $20 in free credits, after which Meta charges $1.25 per million tokens of input and $4.25 per million tokens of output, pricing Wang described as aggressive.

🎙️ ChatGPT can now listen and talk at the same time LINK

OpenAI has released two conversational models, GPT-Live-1 and GPT-Live-1 mini, that can speak and listen at the same time, letting users interrupt naturally and enabling features like live translation.

ChatGPT's current Advanced Voice Mode is being replaced by GPT-Live-1 mini by default, while paid users get the larger GPT-Live-1 model, and both send queries to text models like GPT-5.5 for search and reasoning.

The new voice mode is built for longer conversations, can stay silent to absorb context until called upon, and includes safeguards giving age-appropriate responses to teens and resources for topics like self-harm.

📊 Brown professor's chart exposes mass AI cheating in exams LINK

A Brown University economics professor, Roberto Serrano, believes most of his class cheated with AI after his take-home midterm averaged 96 out of 100, then dropped to 48 on an in-person final.

Serrano offered take-home exams after a campus shooting left students anxious, but 40 of the 86 students scored a perfect 100 on a midterm whose historical average sits between 65 and 80.

When he announced the in-person final, 18 students dropped the course and nine skipped the exam, and 22 of those 27 had earned perfect midterm scores, leaving Serrano to count at least 50 cheaters.

🚨 US warns self-driving cars are a danger to the public LINK

America's top road safety regulator, the NHTSA, has told self-driving car companies to quickly fix what it calls a "clear pattern" of autonomous vehicles interfering with police, firefighters, and other first responders at emergency scenes.

In a letter to the industry, NHTSA chief Jonathan Morrison described autonomous vehicles blocking ambulances and firefighters and failing to respond to flashing lights, flares, smoke, fire, and traffic cones, calling this a "functional insufficiency."

The agency plans to meet with vehicle developers by month's end, and pointed to a late-May Dallas incident where a Waymo reportedly obstructed a fire truck heading to an apartment blaze.

🤖 Mistral's first robotics model steers robots with one camera LINK

Mistral AI has launched Robostral Navigate, its first robotics navigation model, which can guide robots through complex, changing environments using only a single RGB camera and simple language instructions.

The 8bn parameter model reaches a 76.6% success rate on the R2R-CE benchmark, which measures how well an agent follows instructions in environments not seen during training, beating models that use extra sensors like LiDAR.

Built in-house and trained on simulated data, the model uses a pointing-based method that predicts targets as image coordinates, and has been tested on wheeled, legged, and aerial robot platforms.

Other news you might like

• Despite ‘misgivings,’ judge approves Elon Musk’s $1.5M SEC settlementLINK

• I tried DuckDuckGo's new video player, and it blocks YouTube ads for freeLINK

• PC shipments just fell for the first time in two years, thanks to the memory shortageLINK

• Meta to put its own AI chip into production in September, aiming to double computing capacityLINK

• FTC reaches settlement that brings right-to-repair to John Deere farm equipmentLINK

• Meta CTO says employee-tracking data landed ‘where it wasn’t supposed to go’LINK

• Apple hiked Mac prices at exactly the wrong timeLINK

• Google pays $250k for Linux vulnerability allowing guest VM escapesLINK

• China to approve limited Nvidia H200 chip sales to top AI firms: ReportLINK

• Meta reportedly testing prototype AI specs that record everything the user sees and hearsLINK

• A new satellite wants to prove nuclear power can work in space without solar panelsLINK

🧰 Trending tools

Timbal AI: a unified platform for building, deploying, and monitoring AI agents and workflows without stitching together separate tools for each stage.LINK

Toyo: lives in iMessage and calls you by phone, helping triage your inbox, prep for meetings, and pull context from company tools.LINK

Lispr: lets you dictate or translate speech into any app instantly by holding a key, supporting 99 input languages with 346ms median latency and no audio storage.LINK

Monogram AI: an iOS app that answers your questions with dynamically generated visual interfaces instead of plain text, making everyday tasks like trip planning or recipe finding easier to navigate.LINK

GPT-Live: lets you build and scale AI-powered products using a centralized platform with access to multiple leading AI models.LINK

📚 Trending papers & reports

Sketches from 236 countries reveal that human concepts vary far more across cultures than language alone suggests, with visual drawings aligning 45% more closely with real cultural differences than text-based measures do.LINK

Crime report analysis now has a dedicated labeled dataset of more than 1,500 documents, helping software automatically spot key details like suspects, locations, and weapons in police and justice records.LINK

Video and audio compression in multimodal AI gets smarter by using sound as a guide to cut redundant data, keeping accuracy high while processing far fewer tokens.LINK

Medical image generators now produce fairer, more realistic scans across patient demographic groups, including combinations never seen in training, with up to 21% better image quality on those rare cases.LINK

Want to get the latest news differently? Find us on:

See you tomorrow for a new dose of ☕️ Techpresso!