Chinese editionAnthropic 发布 Claude Haiku 5.5：价格降 90%，上下文扩到 100 万 token
Anthropic 的新款小模型 Haiku 5.5 便宜了九成，上下文还扩到 100 万 token，跑高并发任务的成本直接打下来了，做客服、分类、子代理的团队可以看看。
DailyVOL 2026.10.08278 items
Anthropic cuts Haiku 5.5 price 90% as OpenAI math proofs stir debate
Singapore·Thursday, 8 October 2026·DAILY · 08:00 SGT
2 Sep 2026. Anthropic's Claude Haiku 5.5 delivered the day's biggest shock, cutting pricing roughly 90% to $0.10/M input and $0.50/M output tokens for requests under 100K, while expanding context to 1 million tokens and doubling max output to 128K. Benchmarks jumped sharply — OSWorld 2.1 rose from 15.7% to 72.4% and Terminal-Bench 4.0 from 0.0% to 39.2% — surpassing GPT-6 Luna on all shared benchmarks. Meanwhile OpenAI published 372 AI-generated mathematical proofs on GitHub, Lean-formally verified, claiming progress toward a quasi-Riemann hypothesis with results averaging just 3 hours of ChatGPT Pro compute each. Twenty-five Fields Medalists warned that mass-producing mathematical truths could erode research soil, and Gary Marcus criticized the report's missing methodology; an arXiv paper (2610.08144) cautioned that Lean verification does not guarantee natural-language proof correctness.
On the industry front, Huawei's Xu Zhijun said Ascend has overtaken Nvidia in China market share, unveiling the new Peerium compute architecture led by the Ascend 950 supernode, though overseas supply stays limited. In engineering practice, a Google DORA report found AI adoption speeds coding but shifts the bottleneck to trust and review, while a vibe-coding mishap left one developer with a US$10,000 Cloudflare bill after a Durable Objects alarm loop triggered 6 trillion read/write operations. Mistral Large 4 also climbed to 45th on Code Arena: WebDev, within 2 points of Claude Opus 4.8 (High) at roughly one-sixth the price.
01
Models
5 items
Chinese editionMistral Large 4 登陆 Code Arena：WebDev 排名升至第45位
Mistral 的新模型 Mistral Large 4 在 WebDev 榜单一口气从130名冲到45名，价格只有 Claude Opus 4.8 的六分之一，10月底还会开源权重，可以蹲一下。
Chinese editionLiquid AI 开源 d1-3B 与 d1-omni-600M 决策模型
Liquid AI 开源了两款不吐文字只给结构化判断的小模型，3B 版读图文、600M 版还能读音频，一次前向就出结果，适合做实时决策。
Chinese editionvLLM 三周优化 DeepSeek-V4.1-Flash：低并发提速 1.9 倍
vLLM 花三周把 DeepSeek-V4.1-Flash 跑快了 1.9 倍，AgentX 上吞吐 5.3 倍，还附交互图表拆解过程，做部署的可以看看。
02
Products
5 items
vLLM snapshot with CRIU checkpointing
This technical advancement addresses a critical performance bottleneck in large language model deployment by eliminating weight reloading during restarts. For developers working with quantized models in production environments, this feature could reduce infrastructure costs and improve service reliability. The implementation is particularly valuable for applications requiring frequent restarts or scaling in cloud environments.
Anthropic Computer SDK integrates browser-use
This SDK integration significantly reduces development overhead for web automation projects. Developers can now implement complex browser interactions with minimal code, accelerating development cycles. For ASEAN developers, this provides a more accessible way to build browser-based AI applications without needing deep browser automation expertise.
Visual Studio Code 1.141 Adds Zombie Agent Session Cleanup
The cleanup command addresses a real operational problem: long-running agent sessions quietly consume gigabytes of worktree storage before anyone notices. Developers running multiple Copilot agents daily can now set auto-cleanup policies instead of hunting down dead worktrees manually, while the sandbox and grid layout changes make running several agents in parallel more practical. Teams on GitHub Enterprise also gain true multi-instance login without switching windows.
IT之家Source
03
Industry
3 itemsChinese editionOpenAI 用 ChatGPT Pro 生成数学手稿，声称逼近黎曼猜想
OpenAI 让 ChatGPT Pro 写数学证明，平均每个结果只花 3 小时算力，还声称搞定拟黎曼假设，数学圈已经炸了。
Gary Marcus Criticizes OpenAI Vague Unreleased-Model Math Report
The missing methodological details mean the result cannot yet be judged for generality outside mathematics, where verification via Lean makes success cheap to fake relative to open-ended domains. Researchers and AI evaluators should treat vendor-reported results on unreleased models with explicit requests for procedure, failure rates and architecture before drawing AGI conclusions. The critique matters for anyone benchmarking OpenAI against DeepSeek or Anthropic, where disclosure practices differ.
Chinese edition华为徐直军：昇腾在中国市场份额已超英伟达，950 超节点暂不全面出海
华为徐直军亲口说昇腾国内份额已经超过英伟达，950 超节点国内都不够卖，出海计划也说了，信息量不小。
IT之家Source
04
Papers
5 itemsChinese editionLean 验证通过不代表自然语言证明正确：从 Navier-Stokes 看自动形式化局限
一篇用 SCI 层级理论拆解自动形式化局限的论文，直接指向 OpenAI 的 Navier-Stokes 证明——Lean 验证过了，原文可能还是错的。
Epoch AI Launches InnovationEval to Test Automated AI Research
Most AI benchmarks score narrow capabilities; this one grades whether agents can reproduce research-level innovation, a much harder and rarer test. Teams building research agents can use it to calibrate how far current systems are from autonomous AI research. AI labs and investors tracking agent progress now have a concrete yardstick instead of anecdotal claims.
Paper shows agent harness choice shifts results and costs
Harness overhead, not model choice, may be the biggest lever on your agent bill—the repeated prompt and schema payloads compound with every step. Teams building coding agents can audit their harness size first before switching models or paying for extra compute. The finding also cautions against treating harness benchmarks as stable model comparisons, since results flip at similar rates across runs.
Chinese edition研究：LLM 智能体存在来源偏好，会因网站域名选更差的商品
有人测了 12 个智能体模型，发现它们买东西认网站不认货，Booking.com 比 Expedia 多 10 票，标签一改结果就翻。
Chinese editionGoogle DORA 报告：AI 让代码写得更快，信任成了新瓶颈
Google 用 1100 多份工程师反馈说明：AI 写码快了，但交付反而变慢，卡在验证这步。写代码用 AI 的团队都该看看。
05
How-to
5 itemsCloudflare Durable Objects alarm loop drives US$10,000 bill
The billing risk comes from design choices AI code generators routinely make, such as refreshing alarms before expiry, which silently chains invocations forever. Teams deploying on Cloudflare should add alarm-stop rules to agents.md and have a second model review any alarm-renewal logic. Unlike AWS Step Functions, Durable Objects lack a hard spending cap, so cost control depends entirely on the developer's own guardrails.
Chinese editionLauren 30天消耗 1.5T Token，用量最大的模型是 Opus 5.5
宝玉拆解了 Lauren 的 Cursor 用量数据：30天烧掉 1.5T Token，主力居然是 Opus 5.5。还讲了什么时候才敢学她不 Review PR。
SpaceXAI Cookbook adds 5 new Grok API sample apps
The cookbook doubles as an architecture reference: the pattern of using Grok for decisions and Imagine models for execution, plus cross-shot consistency via image editing instead of regeneration, applies to any agentic media pipeline. Developers building voice agents or content workflows can clone the repos on GitHub and adapt them directly. The itemised per-run costs also make budgeting for video generation pipelines more concrete.
vLLM-Omni streaming cuts first audio chunk latency to ~47 ms
The shared-memory ring buffer pattern shows that speech synthesis latency is limited by pipeline design rather than model speed alone. Developers building realtime voice agents or TTS services can reuse this chunked-consumption approach with vLLM-Omni to achieve sub-50 ms first-audio response.
html-plan plugin turns Claude Code plans into reviewable HTML pages
Plans as interactive HTML rather than walls of Markdown cut review time and make the feedback loop structured instead of ad-hoc, which could reshape how teams gate agent coding work. Developers running Claude Code on real projects can install the plugin and start routing plan approvals through numbered decision points. It comes from the Anthropic community plugin marketplace, so setup is a two-command install rather than custom tooling.
278Stories
61First-party
74New models
123Sources
AITOP · generated by the editorial system