xAI Force-Migrates Every Grok Voice Call -- Tomorrow, No Opt-Out Required
xAI announced Grok Voice Think Fast 2.0 on July 29, its most capable speech-to-speech model yet, and starting tomorrow, August 5, every developer or user calling the default "grok-voice-latest" endpoint switches to it automatically -- no opt-in request, no separate rollout window, no per-account toggle. The change happens at the routing layer: whatever product is built on top of that endpoint wakes up running a different model.
Pika's Squish Effect Turns Any Photo Into a Stress Toy
Upload one photo -- a friend, a pet, a product shot -- and Pika's Squish effect renders a giant hand pressing down on it, the image flattening, wobbling, and springing back like memory foam. No prompt writing, no timeline editing: pick a photo, pick "Squish," wait under a minute. The same one-tap formula now powers a family of variants -- Melt, Inflate, Crush, "is it cake" -- each just a different physics joke applied to whatever's already on your camera roll. It's spread the same way Kling's dance-transfer trend did last week (Day 128): collapse the skill and time needed to join a format to almost nothing, and the feed does the rest.
1) xAI Switches Every Grok Voice Call to a New Model Tomorrow
xAI announced Grok Voice Think Fast 2.0 on July 29, its most capable speech-to-speech model yet, and starting tomorrow, August 5, every developer or user calling the default "grok-voice-latest" endpoint switches to it automatically -- no opt-in request, no separate rollout window, no per-account toggle. The change happens at the routing layer: whatever product is built on top of that endpoint wakes up running a different model.
The gains are substantial: time to first audio drops from 1.25 seconds to 0.70 seconds, transcription accuracy improves roughly 1.4x over the prior version and by as much as 10x versus rivals in noisy environments, and reasoning-token use falls to about 0.4x of the previous model's baseline -- xAI says production tool calls now typically execute before the agent finishes speaking its first sentence. Pricing holds at $0.08 per minute of audio. The tradeoff sits with anyone who didn't ask for the swap: a product pinned to the default endpoint inherits a new model's timing and voice characteristics with no window to test it against production traffic first.
2) OpenAI Gives 100,000 Scientists Free Access to Its Frontier Models
OpenAI launched ChatGPT for Academic Researchers this week, offering free access to frontier models -- including GPT-5.6 Sol Pro -- to as many as 100,000 scientists, mathematicians, and engineers. The rollout started with 10,000 researchers this summer at institutions like the Institute for Advanced Study and the École normale supérieure, with eligibility limited to research faculty and postdocs at degree-granting institutions with high research activity. To apply, researchers verify their institutional affiliation through SheerID and submit a recent paper from arXiv, bioRxiv, or ChemRxiv. Each approved researcher can invite up to four collaborators, and workspaces come with business-grade privacy protections -- OpenAI says the data isn't used to train its models by default. The program sits inside a broader $250 million commitment through 2027 to external scientific research, which also funds NextGenAI, a $50 million initiative for research institutions. It's a direct wager that the fastest way to find out what frontier models are actually good for is to hand them, free, to the people whose job is finding things out.
3) DeepSeek's Free Weights Close the Gap to Frontier Models Overnight
DeepSeek published V4-Flash-0731 on July 31 -- a training and post-training upgrade to the same 280-billion-parameter, 13-billion-active mixture-of-experts architecture, not a new model. What changed is capability: on Terminal-Bench 2.1 it jumps from 61.8 to 82.7, ahead of Zhipu's GLM-5.2 and within striking distance of Claude Opus 4.8's 85.0. CyberGym climbs from 38.7 to 76.7, and DeepSWE goes from 7.3 to 54.4.
Pricing holds at $0.14 per million input tokens and $0.28 per million output tokens -- with cache-hit input priced at $0.0028 per million -- and the weights are on Hugging Face under an MIT license, so anyone can download, fine-tune, or self-host the exact model behind those numbers rather than take a vendor's word for it. DeepSeek used its not-yet-released minimal-mode harness with reasoning intensity set to max to get these scores, which is worth remembering before assuming production traffic will see the same numbers. Even with that caveat, closing most of the distance to a frontier model at roughly a fortieth of the price is the story the rest of the industry has to answer.
Three different labs just answered the same question -- what's the fastest way to win share -- with three different levers. xAI competed on latency and reliability, cutting response time nearly in half without raising price. OpenAI competed on distribution, giving away access outright to a constituency it wants building habits and papers around its models before anyone else gets there. DeepSeek competed on price, undercutting frontier providers by more than an order of magnitude while closing most of the capability gap. None of the three needed to raise prices or restrict access to make the case -- which says as much about how much room is still left in the market as it does about any single release.
tomorrow's automatic switch on August 5 -- pin to a specific model version if your app depends on exact timing or voice characteristics.
Researchers program -- free frontier-model access with business-grade privacy is a real deal if your institution qualifies.
paying today -- the MIT-licensed weights mean you can self-host and verify the numbers yourself rather than trust a benchmark screenshot.
intensity -- validate on your own production traffic before switching providers on the headline scores alone.