Welcome to June 28, 2026

The Singularity has stopped having a ceiling and started having a horizon. OpenAI’s Noam Brown notes the plateau “is actually really far out these days,” with a well-scaffolded GPT-5.5 able to think for weeks before benchmarks flatten. Nous Research turned that depth into product, exposing mixture-of-agent presets as virtual models that claim to beat Opus 4.8 by 8% and GPT-5.5 by 11% on their own upcoming benchmark. The same frontier is becoming a commodity, fragmenting downward in price, as DeepSeek’s MIT-licensed V4-Pro-DSpark folds 1.6 trillion parameters and a 1M context into speculative decoding that sips a tenth the KV cache, and a field guide to open weights crowns DeepSeek, GLM 5.2, MiniMax M3, and Nemotron 3 Ultra as frontier-class coders holding the same three-to-six-month gap they have kept for eighteen months, now at a sliver of the cost. Up top, Shopify’s CTO finds GPT-5.6 beats Opus at everything yet still cedes coding to Fable 5, pairing them so Fable writes code while 5.6 runs experiments, and Elon Musk says Grok 4.5 is already nipping at Opus inside SpaceX and Tesla.

Capability this sharp cuts both ways. Anthropic looks set to restore Fable 5 within the week, days after Mythos 5 returned for trusted users, the pause itself a backhanded tribute to real cyber teeth. Those teeth are no longer uniquely American, with Chinese systems reportedly matching Mythos in cybersecurity, and as the export ban drags on, China’s 360 ships Tulongfeng and Tokyo’s Sakana ships Fugu to fill the Mythos-shaped hole. One analyst warns the real fallout is China waking up, reasoning that a CCP on the receiving end of NSA-embedded offensive operations will stop believing it is four months behind and start sprinting, even as the NSA’s own cyber benchmark, due by early August, looks like the gate that finally clears both Fable 5 and GPT-5.6 for the public.

The models that can bankrupt you can also budget for you. Coinbase’s Brian Armstrong nearly halved AI spend while usage soared, using cheaper defaults, smarter routing, and warm caches rather than usage caps, in one case dragging a cache hit rate from 5% to 60%. Perplexity’s CEO sees every enterprise spinning its own model-harness-sandbox-eval flywheel tuned for token value per watt, while engineers lean into “loop engineering” so completely that Claude Code’s creator no longer writes his own prompts, Claude does. Anthropic’s Claude Tag, an AI teammate inside Slack channels, has left staff at Slack owner Salesforce confused even as the company promotes it. The bill still looms, as Gartner expects AI coding costs to outrun the average developer’s salary by 2028, which helps explain founders running half a dozen agents at once who say they have never worked harder while shipping 100x more and burning out anyway.

All of it runs on a physical plant scrambling to keep pace. IBM is industrializing quantum, seeding its Anderon foundry with $2 billion, half from Washington, and pledging $9 billion more over five years toward its 2029 Starling machine. Compute is so scarce that Google throttled Meta’s Gemini access after Meta asked for more than it could spare, and Masayoshi Son is betting against Musk’s orbital data centers, asking “What’s the point?” when power is barely 7% of operating cost and orbit is a decade too slow. Power is going exotic and abundant anyway, as General Fusion tripled its plasma to 8.4 million degrees by mechanical squeeze alone, renewables reached 30% of US generation with solar overtaking wind, and SpaceX began laying an eight-mile “Starpipe” to pump methane straight to Starship.

Atoms are catching up to bits. A $300 wristband called ForceBand teaches robots force-aware hands from human muscle signals, hitting 87% on pick, squeeze, and place, while AGIBOT rolled out its 15,000th robot already holding a 39% share of humanoid shipments. The machines are taking the field too, as Taiwan budgeted $6.6 billion for an “unmanned shield” of 208,200 one-way attack drones, and in a national first a Sacramento sheriff’s drone used a magnet to lift a knife from a suspect’s hand before deputies moved in.

Discovery itself is automating, inward and outward. Erdős problem #870 fell to GPT-5.5-Pro, then held up across 180,000 sorry-free lines of Lean 4, a scale no one had formalized before. A lone PhD synthesized PAC-832, the first selective GalR1 antagonist for Alzheimer’s, in a garage lab run by Claude Code and a liquid-handling robot. Clinicians in Haifa used focused ultrasound to drop an opioid patient’s cravings to zero in 20 minutes, and SETI’s Andrew Siemion proposes sifting one cubic meter of lunar regolith for micron-scale alien debris, the nearest place to look for proof we were never alone.

The proof is in the powder.