1. GPT-6.1 Sol pushes intelligence into a price war / GPT-6.1 Sol 把高能力模型推入价格战
OpenAI says GPT-6.1 Sol approaches GPT-6 Astra on agentic coding, computer use, and professional work while standard input and output pricing is only one fifth of Astra's. Cached input costs $0.10 per million tokens; DeepSWE improves by 6.4 percentage points over GPT-6 Sol, while low-reasoning factual error rate falls from 11.4% to 7.7%.
OpenAI 称 GPT-6.1 Sol 在代理编码、电脑操作和专业工作上接近 GPT-6 Astra，但标准输入与输出价格仅为 Astra 的五分之一。缓存输入低至每百万 token 0.10 美元；相较 GPT-6 Sol，DeepSWE 提高 6.4 个百分点，低推理强度下事实错误率从 11.4% 降至 7.7%。
链接：https://openai.com/index/introducing-gpt-6-1-sol/
2. Sonnet 5.5 improves completed-work economics / Sonnet 5.5 改善“已完成任务”的经济性
Claude Code builders report that Sonnet 5.5 fixes a coding bug about 30% faster with 30% lower usage, and users complete roughly 30% more tasks than with Sonnet 5. In one tool-use demo, the model finished 24 seconds faster while consuming 6,000 fewer tokens. Box's enterprise evaluation also found 2.4x faster completion and 12% fewer tokens on difficult knowledge-work tests.
Claude Code 团队的一线数据表明，Sonnet 5.5 修复代码问题约快 30%、用量低 30%，用户比使用 Sonnet 5 多完成约 30% 的任务；一个 tool-use 演示中，它快 24 秒且少用 6,000 token。Box 的企业评测也显示，在困难知识工作测试中交付约快 2.4 倍、token 少 12%。
链接：https://x.com/bcherny/status/2104638725317923228 · https://x.com/_catwu/status/2104639552170377399 · https://x.com/levie/status/2104648654074343480
3. An open-source AI security agent finds 24 Android vulnerabilities / 开源 AI 安全代理发现 24 个 Android 漏洞
GitHub Security Lab used targeted taskflows organized around mobile entry points and vulnerability classes to discover and report 24 Android vulnerabilities. A medium-sized repository takes about one to two hours per audit, although the workflow requires Copilot authorization and can consume substantial premium-model requests.
GitHub Security Lab 通过围绕移动端入口和漏洞类型设计定向 taskflow，发现并报告了 24 个 Android 漏洞。中型仓库单次审计约需一至两小时，但需要 Copilot 授权，也可能消耗大量高级模型请求。这说明安全 agent 的价值来自可复用审计流程，而不只是通用代码问答。
链接：https://github.blog/security/how-we-found-24-android-vulnerabilities-using-our-open-source-ai-security-agent/
4. GitHub cuts rendering cost by moving from CSS-in-JS / GitHub 渐进迁移 CSS-in-JS 后显著提速
GitHub is migrating toward CSS Modules through its design system, feature flags, and visual regression tests. After Primer components moved, server-side rendering time fell 55% and component initialization time fell 25%, offering a measurable template for paying down frontend performance debt without a risky big-bang rewrite.
GitHub 借助设计系统、功能开关和视觉回归测试，渐进迁移到 CSS Modules。Primer 组件迁移后，服务端渲染耗时下降 55%，组件初始化耗时下降 25%。这个案例说明，前端性能债可以通过可测量、可回滚的分段迁移偿还，而不必一次性重写。
链接：https://github.blog/engineering/architecture-optimization/improving-site-performance-by-shipping-more-css/
5. Infrastructure is becoming agent-readable / 基础设施正在变得可被 Agent 直接调用
Vercel opened domain search without authentication to make discovery easier for agents. It also reports an AI-assisted migration completed in under a week, with roughly 70% faster builds and 75% faster paints; the team extracted two reusable AI skills from the work. Meanwhile, ZenABM positions LinkedIn ad creation, optimization, and reporting as operations callable from any AI tool.
Vercel 开放无需登录的域名搜索，明确降低 agent 的发现门槛；一次 AI 辅助迁移在不到一周内完成，使构建速度提高约 70%、页面绘制提高约 75%，并沉淀出两个可复用 skill。与此同时，ZenABM 把 LinkedIn 广告创建、优化和汇报包装成可由任意 AI 工具调用的操作层。共同趋势是：基础设施和 SaaS 正从“给人使用的控制台”转向“给 agent 调用的能力”。
链接：https://x.com/rauchg/status/2104764419305796094 · https://x.com/rauchg/status/2104660502723072281 · https://www.producthunt.com/products/zena-by-zenabm-linkedin-ads-ai-chatbot
我的判断
今天最清晰的主线不是某个模型单独领先，而是单位任务经济性正在快速改善。更便宜的推理、更少的 token、更短的延迟，以及可复用的 taskflow / skill，会共同决定 agent 是否能进入生产环境。模型路由也将从 benchmark 排名问题变成针对具体工作流的产品决策。
对 opcpay.org 读者的意义
AI SaaS 创业者应重算三个指标：每个成功任务的真实成本、失败或人工接管率、能力升级能否转化为定价权。支付科技还需额外关注权限、审计、回滚和供应商依赖；成本下降只会加快 agent 执行动作，可信控制层才是长期壁垒。