Tag
Skyfall AI released SKY-001, a world model that designs physical products by planning parts and build steps in 3D, matching GPT-6 Astra's CAD accuracy at roughly 150x lower cost per design.
Operon-S1 introduces a System 1-style fast decision controller that separates judging from doing in long-horizon biomedical LLM agents, posing 13 recurring judgements as closed questions answered by a local model in 0.6s median latency while cutting decision-related cost from 8.0%–35.2% with no measurable accuracy loss on five of six benchmarks.
Gemini 4 Carbon is announced as the first model to surpass Claude and GPT, outperforming Opus 5.5 on nearly every benchmark at half the cost, marking an 18% improvement in just two weeks since the Argon release.
A tech user shares that after 100+ hours of testing, they believe DeepSeek V4.1 Flash performs on par with frontier models while being 13-20x cheaper and fully open weights down to the inference stack.
OpenAI 发布 GPT-6.1 Terra，一款兼顾视觉推理、计算机操作与自然写作的高效模型，API 价格较 GPT-5.6 Terra 低 50%，在 Agents' Last Exam、MMMU-Pro、OSWorld 2.1 和 Terminal-Bench 4.0 等基准上大幅提升，同时显著降低幻觉与事实错误率。
该论文将生成模型的推理 effort 视为'信任品'，通过反事实实验量化不同 effort 级别对成本与准确率的影响，发现从低 effort 起步、失败后再升级的策略在多数模型和基准上以更低成本达到或超过固定最高 effort 的效果。
Perplexity releases pplx-decider-v1.2 and pplx-decider-flash-v1.2 under Apache 2.0, claiming 10x lower cost than GPT 6 Luna (low reasoning) with identical decision quality on the Artificial Analysis Decision Index. The new version is also 2.5x cheaper than Jev.
Mouse-1 是一个面向 agent harness 的轻量决策模型，能处理 agent 执行过程中的大量小决策，宣称比 GPT-6 Sol 便宜 163 倍、快 65 倍，可与现有技术栈并行接入。
The Elorian team released the visual reasoning model Glimpse, claiming Astra-level capability at 17x lower cost, capable of analyzing physical laws in videos frame by frame, targeting scenarios such as manufacturing, robotics, gaming, and content moderation.
Skyfall announced its new frontier engineering model Sky-001, claiming it outperforms Astra with 100x greater speed and 150x lower cost.
MiniMax 正式发布 M3.1 Flash，宣称在编码任务上接近 Opus 和 GPT 水平，而成本低约 25 倍；此前以代号 Space Bunny 登顶全球使用榜单。该模型现已在 Cline 中免费提供。
MiniMax 宣布此前以 "Space Bunny" 匿名现身的模型其实是 M3.1 Flash 的早期版本，其在 DeepSWE 编码上得 74%、BabyVision 视觉推理上得 91%，速度是 M3 的 3 倍而成本仅三分之一，适合长时间运行的智能体任务。
Cohere 的一篇新论文发现，多智能体编码团队中，额外的智能体仅负责研究和规划时收益最大，且廉价模型即可胜任规划工作，从而在性能接近的情况下大幅降低成本。
HarnessPro 提出一种无标签的递归自改进方法，通过轨迹诊断与加权 Pareto 选择生成 lite/medium/heavy 三档 harness，在 Terminal-Bench 2、SWE-Bench Pro 等基准上相比 RHO 平均提升 2.0 个百分点通过率并降低 26.1% 单任务成本。
GPT-6 Astra ranks #1 in scientific research benchmarks, beating Claude Sonnet 5.5 while using 22× fewer tokens, and GPT-6.1 Sol also outperforms Sonnet at nearly 1/10th the cost, suggesting scientific research is scaling with compute.
DeepSeek V4.1 Flash 1017 matched Opus 5.5 on long-horizon browser use benchmarks while being 19x cheaper per task, establishing a new cost-performance Pareto frontier.
Meta paper evaluates frontier agents on how well they learn from experience across chess, Go, Hex and NetHack with frozen weights, defining plasticity as held-out gain per dollar. Claude Fable 5 showed the largest absolute gains while GPT-5.6 Sol was far more cost-efficient, showing the best performer isn't the most efficient learner.
A benchmark put 17 AIs through 7,700 hands of heads-up no-limit Texas hold'em, finding that cheap reasoning LLMs matched frontier models, while a simple 30-line Python equity script beat dedicated decision models and held Claude Opus 5.5 to a draw.
The paper evaluates Jev, a non-generative decision model, against a frontier LLM (gpt-6-astra) for RAG claim verification on 495 LLM-AggreFact claims, finding statistically indistinguishable headline accuracy at a 187-fold cost difference but highly unequal error profiles, and showing that a confidence-gated cascade can cut costs without beating a matched random-escalation control.
Paper Instruments 发布 Ultramarine 1 Flash，一款面向金融、法律、咨询等知识工作的开源模型，称其达到 Fable 级别性能，但每任务成本仅 $0.13，比 GLM-5.3 Flash 便宜 60 倍且少生成 6.5% 的 token。