This approach offers a lightweight way to update models without full retraining, potentially reducing costs for personalized AI applications. Developers working with large language models could implement this technique for efficient knowledge updates, especially useful for applications requiring frequent factual updates or user-specific customization.
The standout finding is that verification-driven harness design, not raw model scale, drives much of the gain, since the harness alone added up to 10 points and a 4B fine-tune outperformed a 35B baseline. Teams building research agents can replicate the check-gaps-research loop without large models. It offers a concrete benchmark point against other deep-research agent stacks.
LightOn 的开源 OCR 模型 OCR-3 出了，0.8B 和 4B 两个版本都能自己部署，表格图表效果不错，价格比 Gemini 便宜不少，做文档处理可以试试。
MiniMax 把自家开源视频模型 H3 拿出来跟闭源 SOTA 比物理效果，敢公开叫板，感兴趣可以去实测下。
SpaceXAI 出了个便宜版视频模型，排名超过 Veo 3.1 但价格只有三分之一，10 秒 1080p 一分钟内就能生成，API 上就能调。
Anthropic 的小模型 Haiku 5.5 上榜了，价格和 GPT-6 Luna 一样但分数更高，比自家 Haiku 4.5 便宜九成还涨分，做 Web 开发的可以看看性价比。
Anthropic 的新 Haiku 5.5 在 Code Arena WebDev 榜拿了 1587 分，比上一代多 257 分，写前端页面的可以留意一下这个榜单。
Anthropic 把 95% 的调用交给便宜的 Haiku，只在 5% 关键时刻用 Opus，省钱思路讲得很透。
GPT-6.1 Sol 出了 ultrafast 版，steering 改成即时生效了，跑偏了能马上掰回来，不浪费算力。
OpenAI 给 GPT-6.1 Sol 出了个快 8 倍的 Ultrafast 版，OpenRouter 上已经能直接调，跑对话或编程场景会很爽。
OpenAI 给受信访问的 GPT-6 Sol（Daybreak Blue）几乎没有网络安全防护，跑分直接登顶还比 Grok 4.7 便宜好几倍，值得看看细节。
OpenAI 推出 GPT-6.1 Sol Ultrafast，生成速度最高快 8 倍，价格输入 12 美元、输出 60 美元每百万 token，对要低延迟的场景很有用。
The hallucination gate reshuffles the leaderboard: Muse Spark 1.3 would lead at 26.7% all-pass without it, but two thirds of its passes contain material hallucinations, while GPT-6 Astra keeps nearly all its passes and jumps from joint 10th to 3rd. Legal tech teams benchmarking models for drafting work should weight groundedness over raw pass rates, and cost data matters too — Grok 4.7 tops the table at roughly US$9.50 per task, under half the ~US$21.70 of Claude Fable 5.1.
OpenAI 给 GPT-6.1 Sol 加了 Ultrafast 模式，比标准版快 8 倍，Codex 和 API 都能用，价格也公开了。
GPT-6.1 Ultra fast 已经推送到 API 和 Codex 了，速度更快，写代码的朋友可以直接上手试试。
This result positions Perplexity Decider as a cost-effective leader in decision-making AI, potentially disrupting enterprise decision automation. Developers and product teams should evaluate it against alternatives like GPT-4 for time-sensitive applications where accuracy and latency are critical factors.
StepFun 的 Step 5 Preview 免费进 Cline 了，DeepSWE 成绩压过 Kimi K3 和 GLM-5.3，写代码的可以直接试试。
Odyssey-3 能生成会对你操作实时反应的互动世界，Flash 版免费可玩，还顺手在 Physics-IQ 拿了 66.1 的高分，去看看 demo。
Odyssey-3 能根据一句话生成可实时走动的交互世界，还接上了机器人和真车，Physics-IQ 拿了最高分，挺值得试试预览版。
The benchmark shifts agent evaluation from capability to behavioural trustworthiness, treating false task-completion claims and unauthorized actions as measurable failure modes. Teams deploying agents for multi-step workflows can use the leaderboard to weigh alignment alongside performance when picking a model. Arena says it will expand the index with more safety signals and models over time.
Adaptive environments change how agents are trained: instead of evaluating against fixed benchmarks, the training tasks evolve alongside the agent, which can reduce overfitting to static test sets. Researchers and engineers building agent training pipelines can attend the booth demo or look for the paper to assess whether EnvHarness fits their setup. Comparisons with existing static environment toolchains will likely emerge once the framework is released.
Scaling laws for robot foundation models are far less mapped than for language models, and this 370x parameter sweep across 12 embodiments offers rare empirical evidence on that front. Using latent prediction error as a proxy for planning ability could reduce costly real-world evaluation for robotics teams. Researchers building embodied or cross-embodiment models will find the scaling trends directly comparable to their own setups.
The framework demonstrates how extending temporal context dramatically improves robotic task success rates, potentially enabling more complex autonomous behaviors. Developers working on robotics and AI agents should evaluate how this approach could enhance their applications, particularly for tasks requiring longer-term planning and memory.
一周模型动态都在这：Gemini 4 追平 GPT-6 还便宜四成，Mistral 放了个 1T 参数的开源模型，做选型或追前沿的可以扫一眼。
StepFun 把旗舰模型 Step 5 Preview 放在 Nous Portal 免费开放一周，还配了 Hermes Agent 智能体，想试智能体调用的可以去薅一把。
StepFun 把 Step 5 Preview 放上 OpenRouter 了，写代码跑 Agent 的朋友可以在 Cline、Kilocode 里免费用一周，顺便试试任务成本能省多少。
Kokoro-82M 蒸馏出个 8M 参数的小模型，CPU 单线程就能实时跑语音，没 GPU 也能玩，还有在线试听。
GPT-6 全量开放了，ChatGPT 现在能现场给你生成按钮、表单、图表这些小工具，做计划和算账时特别顺手。
有人让 Claude Opus 5.5 自己列出 20 个可能被它替代的岗位，清单挺有意思，不过只是模型自说自话，当个参考看看。
Anthropic 放出 Haiku 5.5，比上一代便宜 75%，基准和速度都压过 GPT-6 Luna，小模型优先的话可以看看这条。
有人拿 gpt-sol-6.1 和 opus-5.5 在多 Agent 调度里当主控对比，差距大到离谱，做编排的可以参考下。
Most serving stacks still bolt a diffusion server onto an LLM server with glue code, and vLLM-Omni proposes a single control plane covering both plus stateful robot and world-model loops. Engineers building multimodal agents or real-time voice and robotics systems can evaluate it as one runtime instead of maintaining separate stacks. Code is open source and contributions are accepted, so teams can inspect the architecture directly in the repo.
The takeaway is that Anthropic is now competing on price, not just capability, in the small-model tier where Chinese labs have dominated. Builders choosing a cheap workhorse model for high-volume tasks can add Haiku 5.5 to a shortlist alongside DeepSeek-V4.1 flash and GLM 5.3 flash and re-run their own workloads to compare latency and cost. Pricing and availability for ASEAN API users should be checked against local provider rates.
Liquid AI开源了两款小尺寸决策模型，最小只有600M还支持多模态，在边缘设备上几十毫秒就能出结果，做端侧Agent的可以试试。
GPT-6 正式全量推送了，回复里能直接生成按钮、图表和小工具，跟 GPT-5.6 的纯文本对比差别很明显，可以上手试试
Only the latest 2000 items are listed; for older ones see the daily archive
