SpaceXAI Ships Grok 4.5 as OpenAI Launches Real-Time Voice Model and GPT-6 Reportedly Weeks Away

SpaceXAI launched Grok 4.5 this week, which they're calling their smartest model yet. It's built specifically for coding, agentic tasks (meaning tasks where the AI runs on its own without needing hand-holding), and knowledge work. It was trained on tens of thousands of Nvidia's newest GB300 chips alongside the team from Cursor, and it runs at about 80 tokens per second while being roughly twice as efficient with tokens as its competitors. Pricing is $2 per million input tokens and $6 per million output tokens.

The benchmark performance actually backs up the hype. Grok 4.5 debuted at number 4 on a major AI capability ranking called GDPval-AA, at a tenth the cost of the three models sitting above it. It also took the number 1 spot on AutomationBench, becoming the first model ever to complete more than half of real-world software workflow objectives without breaking business rules. And it more than doubled Grok 4's performance on complete end-to-end professional tasks, now finishing about 1 in 3 of them. Aditya Gupta credits two things: massive synthetic training environments, and a tight feedback loop where "a smarter model behaves better, and better behavior unlocks more intelligence."

The team also has a hardware advantage that hasn't fully hit yet. Their custom C/C++ rewrite of the underlying software gives them a 4x full-cycle speedup, which makes launching a new model every month realistic through basic engineering alone. Musk also noted that their bespoke inference stack (the part that actually runs the model when a user talks to it) isn't even plugged in yet, and that it should double their speed again when it comes online. Observers are now ranking SpaceXAI third among AI labs, ahead of a stumbling Google, and warning that with a 10 trillion parameter model reportedly in training, money can still buy the frontier but only for another 12 to 18 months. One asterisk on the trophy: the training run accidentally pulled in the Cursor codebase, benchmark tasks included. In the era of AI-generated training data, contamination now sometimes comes from inside the house.

The rest of the frontier is talking back, literally. OpenAI launched GPT-Live this week, which is a new voice model that can listen while it's talking (full-duplex, meaning it doesn't wait for you to finish before it can respond) and hand off harder questions to a bigger model in the background. The team's most excited about the fact that the voice sounds smarter, not just chattier.

Behind the scenes, releases are compressing. GPT-5.6 is reportedly landing today or in the next few days, closing out the 5.x line. GPT-6, built on a much larger training run, is expected within weeks. Claude Fable 5.1 is close behind. And DeepSeek V4 should also be out imminently. Veteran AI researchers are confirming the sprint, saying "Mythos changed everything. Everyone is going big," and that both leading labs are seeing "no ceiling" on how far current techniques can go. Users are already picking and choosing between tiers. One user compared Fable to an F1 car and GPT-5.6 to a Tesla Model X Plaid, and admitted they drive the Tesla 95% of the time.

The cheaper end of the market is also closing in fast. Cognition just released SWE-1.7, an open-source coding model that gets near-frontier performance built on top of an open Kimi base, running at 1000 tokens per second. And Prime Intellect raised a $130M Series A for its open superintelligence stack after crossing $100M in revenue in less than a year.

The hardware layer is getting more strategic on both sides. China is now planning to let its top AI companies buy limited numbers of Nvidia's H200 chips to ease a shortage caused by soaring domestic demand. Meta, meanwhile, is dealing with rising memory prices by building a custom "bridge chip" that lets them use older memory inside their newer servers. That said, Meta is going big where it can't go cheap. The company just broke ground on a $9 billion data center in Alberta, Canada, which is their largest outside the US, and it'll be powered by its own dedicated natural gas plant.

Robots are graduating from demos into real duties. Teleoperated humanoid robots just performed live surgery on animals for the first time, including in a scenario where two robots operated side by side. The general-purpose brain for robots is also going open source. Ant Group released LingBot-VLA 2.0, an AI model pre-trained on 60,000 hours of physical world data across 20 different robot body types. And 1X, the robotics startup, just teased what they're calling "the most advanced robotic hand in human history." Fittingly, the right to fix machines is being returned to their owners. A landmark settlement will force John Deere to hand farmers dealer-grade repair tools for a full decade.

The machines are also watching us right back. Meta is testing new "super sensing" glasses that can quietly record every moment, possibly even without the little LED that's supposed to signal recording is happening. It's total recall coming to consumers before total consent has been sorted out.

Watching is also becoming a growth industry in orbit. There's a new proposal to send small satellites equipped with neutron detectors to check whether a suspicious foreign satellite might be hiding a nuclear device. And engineers just demonstrated a superconducting "Supertorquer" that accelerated a satellite without any fuel, by pushing against the Earth's magnetic field. That kind of endless-run capability is exactly what an inspection satellite would need.

Back on Earth, the AI dividend is compounding but unevenly. Founders now estimate they'd need 55% more employees to do the same amount of work without AI. But the payoff really depends on whether managers are willing to actually delegate work to it. The IMF is seeing the same split globally, where countries that export AI hardware are blowing past forecasts while less-exposed economies are sagging.

Intelligence may also be unevenly distributed across the galaxy. Congress is now formally pressing the CIA and FBI for records on the 1996 Varginha UFO incident in Brazil. And NASA's chief said this week that he "can't hate" UAPs, and that he sees "a very real possibility" that we'll conclude in our lifetime that "there's life everywhere out there."