AI News
Important AI releases, clear comparisons, and the evidence behind the headlines.
13 articles for “Agents”
OpenAI launches GPT-6.2 Astra and Sol for complex projects and everyday work
GPT-6.2 Astra and Sol are available in ChatGPT Work, Codex and the API. Astra cuts standard input and output rates by 20%, while Sol retains its previous pricing.
Tencent launches Hy4 flash with 224B parameters and a million-token context
Hy4 flash launched on November 16 with 224B total and 12B active parameters, multimodal support and a 1M-token context. Standard API rates are RMB 0.5 input and RMB 2 output per million tokens.
Qwen4-Max launches with a million-token context for coding and long-running agents
QwenCloud lists qwen4-max with multimodal understanding and a million-token context, available through Token Plan and pay-as-you-go APIs.
openPangu-2.1-Pro launches on Huawei Cloud MaaS for agents and HarmonyOS development
Huawei Cloud MaaS adds openPangu-2.1-Pro, with updates targeting agents, AI coding and inference optimization, including HarmonyOS ArkTS and Ascend operator development.
Baidu releases ERNIE 5.5 Flash Preview with 280B parameters and 1M context
Released November 11, ERNIE 5.5 Flash Preview has 280B total and 14B active parameters, plus a 1M-token context. Its Qianfan API is live. Baidu plans to release weights by November 25; post-training is still ongoing.
Gemini 4 Carbon expands access for complex coding and long-horizon agent workflows
Google expanded Gemini 4 Carbon access on November 10 for paid API developers and rolling access for AI Ultra subscribers. Introductory input/output rates are USD 2/10 per million tokens, rising to USD 4/20 after the promotion ends November 30.
LongCat-2.5 explained: native multimodality, 1M context and agent gains over 2.0
Meituan’s LongCat-2.5 combines native vision and long-horizon agents with 1.6T total and about 48B active parameters. We examine scores versus 2.0, evaluation caveats, context limits, limited-time API pricing and weight availability.
Mistral ML4 open-weight release explained: Le Chonk, coding, security and reasoning costs
The updated open-weight Mistral Large 4 (ML4, Le Chonk) improves coding, defensive security and token efficiency. We examine its 1.05T total and 52B active parameters, reported benchmarks, preview comparisons, API rates and deployment constraints.
Nanbeige4.3-33B explained: a 512K-context open-weight model for coding and office agents
A closer look at Nanbeige4.3-33B: official agent benchmarks, comparisons with Qwen3.8-27B and peers, thinking and tool-use settings, and the memory costs behind its 512K context.
KRAFTON Raon-1 mini explained: Korean, agents and single-GPU deployment in a 30B-class open-weight model
A source-checked look at Raon-1 mini preview: 28.8B total and 4.3B active parameters, official comparisons, speculative decoding, deployment requirements and the limits of its free API offer.
Kimi K3.1 Pro explained: long-horizon coding, 3D creation, factuality and the upgrade from K3
A source-checked look at Kimi K3.1 Pro and Highspeed: official scores, comparisons with K3, RMB API pricing and integration limits, with internal evaluations clearly separated from public benchmarks.
MiniMax M3.1 Flash explained: agent benchmarks, pricing, and whether to upgrade
A source-checked look at MiniMax M3.1 Flash: official agent, coding and computer-use scores, token pricing, speed claims, and the limits of comparisons with M3.