Portable Computer for Windows is now available on @AMD Ryzen AI Max Series processors. It makes it easy to run local AI agents that work with your connected apps and local files. Kick off tasks or schedule recurring work that runs entirely on your device.

Any enterprise working on @AnthropicAI Claude Opus for agent work gets more of the work done for less money by moving to yesterday's brand-new Opus 5.5, and in our testing the gap is...pretty large. Opus 5.5 gets more work done for more people, at a noticeably lower price than Anthropic offered before, as measured by @Signal_65 PINNACLE. ➡️ 2.3x fewer weighted errors than Claude Opus 5, and 100% of the multi-step jobs finished end to end against 95% ➡️ 34% less per correct task (!!), $0.40 against $0.60, on a list price 20% below Opus 5 with cache reads at 5% of input rather than 10% ➡️ Fabrication in retrieval answers down from 7.6% to 2.8% ➡️ It also passes Claude Fable 5.1 with 1.3x fewer weighted errors at 63% less per correct task Full comparison on the Signal65 PINNACLE results views at pinnacle.signal65.com

Enterprises that want to utilize the largest open-weight models in an agentic fleet have a narrower choice of nodes and infrastructure that can support that level of quality. Kimi-K3 at 2.78T parameters is the largest model in the @Signal_65 PINNACLE GPU results, served from its MXFP4 weights on both current-generation platforms, and it is the one model where the comparison runs @AMD Instinct MI355X ahead of comp on both supported agent-count and speed. ➡️ 32 concurrent agents at the service level on one MI355X node against 20 on the B300 in our testing, 1.6x the headcount, with first token still in fractions of a second at the ceiling ➡️ From 16 agents up the MI355X is also the faster platform per agent, 20.4 against 15.1 tokens per second at 16 ➡️ The B300 holds a small edge at the lightest shared load, 31.2 against 26.4 tokens per second at 8 agents Full results: pinnacle.signal65.com/articl…

I was going through a storage unit where I have been keeping a lot of PC hardware stuff, not organizing it yet. Came across some items like this. Who knew I had been holding appreciating assets this whole time!

🚨🚨NVIDIA Vera CPU Agentic Performance🚨🚨 If you are planning capacity for agents (and aren't we all??), the CPU decides how many sessions a rack can keep busy, and the GPU spends much of its day waiting on it. Over the past few weeks Signal65 worked with @nvidia to measure that work on the new Vera CPU, one sandbox per core, no model in the loop, timed from sandbox creation to teardown. In our testing Vera finished the agentic task lifecycle 1.64x faster per core than a leading x86 server processor across 69 Terminal-Bench 2 tasks, and 1.87x on the 20 that look most like an agent's daily work. Our own runs on the same hardware came in faster than the NVIDIA figures, so the headline number is conservative for this comparison. Full write-up and per-task data: signal65.com/research/ai/val…

🚨🚨NVIDIA Vera CPU Agentic Performance🚨🚨 If you are planning capacity for agents (and aren't we all??), the CPU decides how many sessions a rack can keep busy, and the GPU spends much of its day waiting on it. Over the past few weeks Signal65 worked with @nvidia to measure that work on the new Vera CPU, one sandbox per core, no model in the loop, timed from sandbox creation to teardown. In our testing Vera finished the agentic task lifecycle 1.64x faster per core than a leading x86 server processor across 69 Terminal-Bench 2 tasks, and 1.87x on the 20 that look most like an agent's daily work. Our own runs on the same hardware came in faster than the NVIDIA figures, so the headline number is conservative for this comparison. Full write-up and per-task data: signal65.com/research/ai/val…

You cannot pace a frontier you have not measured, and in our testing the open weights are moving closer to it 💡 An enterprise choosing a model for agent work has more good options this month than in August, and a louder argument around them. Three of the labs at the top of the Signal65 PINNACLE board, @OpenAI, @SpaceXAI and @AnthropicAI, backed a call last week to pace frontier AI development, and the same week we scored eight new configurations, five of them open weights from @Alibaba_Qwen, @deepseek_ai and @Zai_org. Any agreement to pace capability starts from an independent measurement of where it stands, and in our testing the open weights landed closer to the hosted frontier than the commentary allows. Both hosted additions land on the cost frontier as well. ➡️ Gemini 3.8 Flash has drawn online skepticism as an agent. In our testing it finished 99.6% of the multi-step enterprise jobs it started at high effort and 98.9% at medium, second and third on the board behind only GPT-6 Astra at max effort. On the IT Professional persona, the most agentic of the five, the high setting makes about half the weighted errors of Claude Fable 5.1. It takes a median 59 tool calls per task against 8 to 11 for Astra, at 54 to 65 cents per correct task. ➡️ Grok 4.6, our first look at xAI, lands fourth on the board at 47 cents per correct task with 0.0% fabrication in the retrieval suite. That is the same score as Astra at medium effort for half the price, and it puts Grok on the cost frontier alongside both Gemini settings. But the open-weight additions matter most to a buyer running its own hardware, and they will matter more if the industry pushes away from the idea of slowing down. ➡️ Qwen3.8-2.4T-A95B is the strongest open-weight result we have scored, with 15% fewer weighted errors than Claude Opus 5 in our testing, and it did that at medium reasoning effort. The better comparison for a buyer is Qwen3.8-Flash-Next, a 180B model within 5% of its 2.4T sibling that runs on a desktop system like the NVIDIA DGX Spark. ➡️ DeepSeek-V4.1-Flash at maximum reasoning effort makes 1.5x fewer weighted errors than the previous DeepSeek V4 Flash and 1.8x fewer than its own high setting, so the generation gain is real. It pays with 20x the output tokens per correct answer and a median of more than 40 tool calls per task, the lowest token efficiency on the board, and whether that trade pays depends on what those tokens cost on your hardware. ➡️ GLM-5.3-Flash, 320B total and 18B active, lands between the full GLM-5.3 at maximum and low effort in our testing, so the small model gives up little to the large one, although both remain below the previous GLM-5.2. A paced frontier, if one is ever agreed, will be negotiated over numbers like these, and in our testing the open weights are close enough behind the hosted models that there is less room to pace than the debate assumes. The score and price of a correct task for all 62 scored builds are live on the PINNACLE results views at pinnacle.signal65.com