LWiAI Podcast #252 - GPT 5.6, Grok 4.5, Nemotron-Labs-Diffusion, AI 2040
Last Week in AI Last Week in AI ● Covered by 4 sources
OpenAI released GPT-5.6 and rebranded its coding tool ChatGPT Work, while rivals fired back with cheap models fast. Safety documentation looks thinner every week as the race speeds up.
Based on reporting by Last Week in AI, Last Week in AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI pushed GPT-5.6 out to the public this week, folding in the Sol and Luna variants and renaming its desktop coding product ChatGPT Work. The rollout came tangled in a messy dispute over whether the US government effectively signed off on the release, or actually delayed it, and that argument is really a symptom of a bigger problem: nobody seems to agree on what frontier-model oversight is supposed to look like right now, and jailbreak concerns are only adding to the unease.
Meanwhile the pricing and capability war got louder. SpaceX AI's Grok 4.5 landed as an Opus-class coding model at a very low price, but with barely any safety documentation attached. Meta answered with Muse Spark 1.1, matching the aggressive pricing and posting big gains on coding and cyber benchmarks, only this time backed by a lengthy safety evaluation. Put those two side by side and you get a pretty clear split in how seriously companies are treating disclosure even as they compete on the exact same turf.
Meta's image push didn't go as smoothly. Muse Video got previewed and Muse Image shipped, but Meta pulled back fast after backlash over how easily the tool could generate images of public Instagram accounts. At the same time, cost pressure is reshaping who actually uses these models: Chinese open-source options now account for more than 30% of weekly OpenRouter tokens, and that shift has people talking openly about risks like insider threats alongside the usual efficiency gains.
On the infrastructure and policy side, Meta is reportedly weighing whether to sell AI compute as its own cloud business, and US energy regulators are pressing grid operators harder about connecting massive new data-center loads. Anthropic published research on a 'global workspace' method for interpreting verbalizable internal representations inside models, a step toward actually understanding what's happening inside these systems rather than just benchmarking outputs. And there are reports that China may move to restrict overseas access to its top models, which would mark a notable turn from the current cost-driven openness.
All of this sits under the shadow of the AI 2040 proposal, which argues the US and China should coordinate to deliberately slow progress until alignment work catches up. Given a week where companies can't even agree on who approved a model's launch, that kind of coordination sounds both necessary and, frankly, a long way off.
My take — AI-written commentary, not fact-checked reporting
Grok 4.5 shipping as an Opus-class coding model with almost no safety documentation, right as Meta ships a competitor with a full safety evaluation, says everything about where this market actually stands: speed and price are winning, disclosure is optional. Nobody should be surprised that Chinese open-source models are eating into OpenRouter usage when cost is the only lever companies seem willing to compete on transparently. Proposals like AI 2040 asking the US and China to coordinate on slowing down are worth taking seriously, but they read almost naive next to a week where firms can't even agree on who approved a model's release.“}
Read more about this at: Last Week in AI