Independent testing on Qualcomm's newest laptop chip found it giving CPU cores more memory bandwidth than two higher-profile unified-memory AI chips, despite a narrower bus. Separately, Anthropic put real numbers on what it costs to actually run an open-weight model capable of writing its own exploits.
Qualcomm's smaller bus gives its CPU more bandwidth than NVIDIA's AI mini PC chip
Chips and Cheese measured the system-level architecture of Qualcomm's Snapdragon X2 Elite Extreme, the 18-core chip behind this generation's flagship Copilot+ laptops, on a unit sampled by ASUS. Its 192-bit LPDDR5X setup has 228 GB/s of theoretical DRAM bandwidth, a narrower bus than the 256-bit designs behind AMD's Strix Halo (256 GB/s) and NVIDIA's GB10 (273 GB/s), the unified-memory chip behind NVIDIA's AI mini PCs built for local model work. In the outlet's own conclusion, Qualcomm's chip still "beats out AMD's Strix Halo and Nvidia's GB10 in terms of CPU-side bandwidth."
Qualcomm groups its cores into three six-core clusters, two of "Prime" performance cores and one of "Performance" efficiency cores (its naming, not a typo), each running its own L2 cache at core clock instead of a slower shared mid-level cache. That gets L2 latency down to 20-21 cycles on every cluster and lets a CPU-only bandwidth test hit 176.98 GB/s, 77.6% of the chip's theoretical maximum. DRAM latency measured 115 ns, and even when every cluster competed for bandwidth at once, latency for any single core stayed under 200 ns, which AMD's own client designs don't always manage, per the same testing.
The difference comes down to how each chip splits bandwidth between CPU and GPU. When Chips and Cheese loaded both sides at once, Strix Halo and GB10 reached higher combined shares of their theoretical bandwidth (88% and 81%, against 74.2% for Qualcomm), but only because their interconnects let the GPU take the larger share. Qualcomm instead splits available bandwidth close to evenly, which is why its CPU cores keep more of it to themselves even on a smaller bus. The outlet calls the design "something of a warning to Intel and AMD" in the laptop chip's own weight class, while cautioning it is one reviewer's own synthetic benchmark suite on a single unit, not a production fleet.
What an exploit-writing open model can do and what it costs to run
According to Anthropic, Zhipu AI's openly downloadable GLM-5.3 can automate software exploit development at a rate closing in on Anthropic's own unreleased "Claude Mythos" model, Tom's Hardware reports, citing Anthropic's own testing alongside a separate report from the Center for AI Standards and Innovation. In Anthropic's sandboxed "Exploitbench" suite, which targets Google Chrome, GLM-5.3 produced a working end-to-end exploit in 50 of 410 attempts, against 56 of 410 for Mythos; on a second internal benchmark for chaining bugs into full control-flow hijacks, GLM-5.3 scored a 4% success rate to Mythos's 6%, while Kimi K3 and DeepSeek V4.1 Flash each scored 0%.
Anthropic says the stock, unmodified GLM-5.3 refuses harmful requests about as often as its own models do, but those refusals can be bypassed. A prompt that has the model role-play an unrestricted agent worked 64% of the time, and prefilling its "thinking" tokens to force a response worked 92% of the time, both against the stock model. Because the weights are public, the model can also be "abliterated," a technique that strips refusal behavior directly out of the weights: Anthropic says that drops the refusal rate to 6% for full GLM-5.3 and 14% for the lighter GLM-5.3-Flash.
Running the full, abliterated model is not casual. Anthropic's figures put FP8 weights at roughly 306 GB of VRAM, with a comparable allocation needed for KV cache, and 4 TB/s of memory bandwidth to sustain 100 tokens per second, a specification Anthropic says points to a cluster of eight NVIDIA H200 GPUs. Tom's Hardware puts the cost of that hardware, bought outright, in the hundreds of thousands of dollars. For a would-be attacker without that hardware, Anthropic says abliterating the smaller GLM-5.3-Flash took 2,200 GPU-hours, about $4,400 at roughly $2 per GPU-hour. Tom's Hardware, working from Runpod's published H200 rental rate of $3.79 an hour per card, calculates that generating around 100 million tokens on eight rented H200s would run about $8,422 and take roughly 11.5 days.
That leaves a real gap between the headline capability and who can use it. A state actor that can simply buy the hardware clears the bar easily; a lone attacker renting compute is looking at thousands of dollars and nearly two weeks of GPU time to generate around 100 million tokens, a cost Tom's Hardware argues is the bigger practical barrier here than the model's raw capability.