SpaceXAI dropped Grok 4.5 and it's actually good enough to pay attention to now.
Coding and engineering: hangs with GPT 5.5 and Claude Opus 4.8. SWE Bench Pro hit 64.7%, first time beating GPT 5.5 on this one. In practice I'd put it around GLM-5.2 level.
$2/M input, $6/M output. Cheap relative to the rest of the frontier.
SpaceXAI dropped Grok 4.5 and it's actually good enough to pay
More posts
OpenAI is shutting down Atlas, its AI browser, less than 9 months
after launch. Atlas dropped in October 2025, meant to replace Chrome. ChatGPT sidebar, agent mode, browser memory. All built in. Service ends August 9. Turns out people don't switch browsers. Easier to put AI into Chrome, Edge, whatever people already use than to build a whole new browser. Atlas's core stuff is getting folded into the ChatGPT desktop app and browser extension.
Tencent's Hunyuan Hy3 is out. Best agent capability at its size.
295B total params, activates only 21B per inference. Simple queries answer fast. Complex ones tap deeper experts. 256K context, Apache 2.0 open source. Pricing sits near DeepSeek V4 Flash: ¥1/M tokens in, ¥4/M out, ¥0.25 cache hit. OpenRouter has two weeks free. Agent scores: BrowseComp 84.2, barely behind GPT-5.5 at 84.4. ClawEval 68.5, beating DeepSeek V4 Pro (62.4) and Qwen 3.7 Max (65.2). Tested it alongside DeepSeek V4 Pro, K2.7, M3. Coding isn't the strongest, but agent ability is real, and the price is right. Solid pick as a general purpose agent model for openclaw, hermes, etc. openrouter.ai/tencent/hy3:free
Meituan shipped LongCat-2.5
. 1.6T params, ~48B active, native multimodal, aimed at long-horizon agents, trained start to finish on Chinese hardware. Images get quantized into multi-level discrete tokens, and that's where Visual Coding and 3D pick up. On the RL side, failed trajectories get turned back into verifiable training tasks. W4A4 inference, 43% faster prefill. Meituan's own evals put it between Claude Opus 4.8 and Opus 5.
SenseTime open-sourced SenseNova Loop-9B-Preview.
It stacks layers instead of parameters. The decoder has 32 layers, but layers 4 through 27 run twice with shared weights, so a forward pass does 56 layers of work while the checkpoint stays at 9B. Their own eval numbers: 71.38 on BIRD-Dev and 13.72 on AutomationBench, both above Qwen3.5-27B. Terminal Bench is still behind.
