Chinese contentOpenAI 发布 GPT-6 Sol 和 Luna：能力对标 Claude，API 价格比 GPT-5.6 再降 50%
OpenAI 把 Astra 的能力做成大中小三档，Sol 跑分超 Claude Opus 5 但成本只有 9%，跑长任务智能体的开发者账单能省不少。
DailyVOL 2026.09.23282 items
OpenAI's GPT-6 Sol and Luna trigger AI price war with Opus 5.5
Singapore·Wednesday, 23 September 2026·DAILY · 08:00 SGT
2 Sep 2026. The day's biggest fact is an unprecedented same-day pricing clash: OpenAI launched GPT-6 Sol and GPT-6 Luna, with prices 50% below GPT-5.6 equivalents — Luna at $0.10/$0.50 per M tokens — and Sol scoring 33.2% on AutomationBench versus Claude Opus 5's 26.9% at roughly 9% of the per-task cost. About an hour earlier, Anthropic shipped Claude Opus 5.5 at $4/$20 per M tokens (down 20% from Opus 5, cache reads down 60%), claiming 40% lower cost and 30% faster output; xAI's Grok 4.7 at $2/$6 set the stage the day before. Second, agent infrastructure matured: SGLang, Qwen and NVIDIA applied NVFP4 to KV Cache, lifting decode throughput 26–78% on Blackwell, Mem0 released DolphinBench for agent memory, and OSWorld-Pro showed Claude Opus 5 dropping to 75.7% under process-based evaluation. Third, economics came into focus: Harvey turned -50% gross margins positive in one quarter, while an arXiv study of 325,000 trials found eight of 13 agents steer wealthier users toward pricier options — dubbed adversarial delegation.
01
Models
5 items
Chinese contentClaude Opus 5.5 与 GPT-6 Sol/Luna 同日发布，掀起价格战
Anthropic 和 OpenAI 前后脚发新模型还互相砍价，GPT-6 Luna 只要 $0.10/M，做应用的成本又降一半，附完整价格对比表。
Chinese contentAnthropic 发布 Claude Opus 5.5：成本比 Opus 5 低 40%，速度提升 30%
Anthropic 新旗舰 Opus 5.5 来了，价格降了 20%，跑得还快 30%，编码能力直接对标 Fable 5.1，写代码的可以试试。
IT之家Source
Chinese contentSGLang × Qwen × NVIDIA 三方合作：用 NVFP4 KV Cache 突破长上下文与 Agentic 推理瓶颈
SGLang、Qwen和NVIDIA联手推出NVFP4 KV Cache，4-bit精度下1.78倍容量提升，长上下文推理提速近80%。
TypeSafe launches System One model Jev
Jev offers a different approach to AI decision-making that could better handle everyday tasks, contrasting with text-based models like Claude Code and Cursor.
02
Products
3 itemsClaude Code v2.1.280 Adds Opus 5.5 Default and MCP Fixes
The Opus 5.5 default switch changes baseline cost and context behaviour for every Claude Code session without opt-in, so teams budgeting per-token spend should re-check their model settings. The new environment variable eases long MCP server definitions, and the safety-check retry fixes make auto mode more predictable. Developers running Persian or Arabic input or Windows terminals get targeted reliability fixes.
Amazon SageMaker AI adds concurrency sweeps for endpoint right-sizing
The CreateAIBenchmarkJob API is available to any SageMaker AI customer, so teams hosting open-source models like Llama can cut inference costs by matching instance types to real traffic instead of over-provisioning.
Mem0 releases DolphinBench for agent memory evaluation
Memory benchmarks have mostly measured recall on quiz-style questions, ignoring the cost and latency trade-offs that decide what teams actually ship. DolphinBench scores downstream actions after long histories, so results better reflect real agent behaviour. Builders choosing a memory layer for production agents can use it to compare systems on all three axes at once.
03
Industry
5 itemsHarvey turns gross margins from -50% to positive in one quarter
The margin turnaround came from engineering levers — routing, harness optimisation and post-training — rather than pricing hikes, which is a playbook other AI startups can copy. Teams building on rented frontier models should take note of the argument that custom models, harnesses and evals are becoming core company assets. The spend-control tooling Harvey shipped, like per-matter cost attribution, is also directly relevant to any B2B AI product managing enterprise budgets.
Jeff Dean's First Public Talk Post-Google
Dean's insights on research direction and AI safety directly impact developers and researchers working on autonomous systems in the region.
Chinese contentGoogle 科学家 John Platt 谈 AI for Science：从 sklearn 到自动化科学
sklearn 里的 SMO 算法作者、拿了奥斯卡的 Google 科学家聊 AI 怎么做科学、解决气候问题，访谈内容很扎实。
Vista8 Podcast Takeaways: China Pretraining Leads, Post-Training Trails
The piece compares Chinese and overseas labs on data and compute gaps and cites Tencent Hunyuan's reorganisation, useful context for anyone tracking how China's models compete with US fronts like GPT and Claude.
Agent Workloads Push CPU-to-GPU Ratios Toward 1:1
The takeaway is that agent infrastructure spending is quietly rebalancing toward CPU-heavy server configurations, a signal hardware buyers and cloud capacity planners can act on when sizing fleets for inference. Vendors like Intel and AMD stand to gain share as tool-calling pipelines scale. Data-centre operators in Singapore and Malaysia building out AI capacity should factor CPU budgets into their GPU-led procurement plans.
04
Papers
5 itemsChinese contentGoogle Research提出RRSI：给智能体harness递归自我改进加正则化
Google开源了RRSI：给自动进化智能体harness的方法加正则化，分布外基准还能涨4.7分，还省30%的token。
Physics-Informed Training Framework Uses Neural Operators to Avoid Bad Minima
An imperfect neural operator only needs to place a PINN in the right solution basin, a practical recipe for anyone solving PDEs in engineering or research where labeled data is scarce and out-of-distribution accuracy matters.
Chinese contentHarness-Zero：把专用 Agent Harness 的能力蒸馏进模型权重
把专用 harness 的本事直接练进模型权重，部署时甩掉笨重框架，成绩反而从 23.3% 涨到 44.3%。
05
How-to
5 itemsHamel Husain and Shreya Shankar publish advanced AI evals guide
Eval skills are shifting from niche research work to a stated hiring requirement for product managers, which signals how AI teams now allocate engineering resources. Practitioners building or maintaining AI features can use the guide's automation advice and free plugin to set up eval loops without dedicated infrastructure. The cited cost and accuracy gains offer concrete baselines for teams benchmarking their own pipelines against peers.
Chinese contentJev是什么：TypeSafe AI 推出只做判断不输出文本的模型
前 OpenAI 研究员搞了个不会打字的模型 Jev，专门做分类和判断，几十毫秒出结果还带置信度，比拿 ChatGPT 硬干省快多了。
Chinese contentOpenAI 发布提示词缓存调优指南
OpenAI 出了个提示词缓存的调优教程，教你设 cache breakpoints、调 reasoning effort 还不丢缓存，做应用的开发者可以直接抄作业。
Fable and Opus Loop Cuts Token Costs in Coding Workflows
The loop turns model review into an automated gate, so the expensive model spends tokens on planning and judgement rather than routine execution. Developers on metered Claude plans can replicate it with a single advisor command and one review prompt, cutting manual back-and-forth. The /compact timing tip matters too: compressing right after the plan is written lands while prompt caching is still warm, so it costs less than compacting after the cache expires.
Chinese contentMCP 是坏主意吗？Simon Willison：受限场景下仍不可替代
Willison 拿 Claude Code 举例：全权限智能体确实不用 MCP，但要管权限、密钥和审计日志就离不开。
282Stories
76First-party
72New models
87Sources
AITOP · generated by the editorial system