Jev 决策模型与 Opus 5.5 实测
做法与步骤
- Jev 返回预定义值（类别、分数、概率）而非生成文本，每百万输入 token 仅 4 美分且无输出费用，Claire 认为这覆盖了约 90% 的软件工作流需求。
- Claire 用 Jev 分析 1,700 个 ChatPRD 拉取请求（17,000 对比较），花费 9 美分，两分钟内发现近 30% 工程工作投入平台、安全和基础设施。
- Claude Code 和 Codex 的本地会话记录可被 Jev 快速分类，Claire 发现工程占其 AI 使用比例从 1 月的近 100% 降至 9 月的不足 40%。
- Jev 与前沿模型搭配更强大：先用 Jev 分类、聚类、过滤和路由大数据集，再将重要组发送给 GPT-6 Astra 深度推理，处理 1,100 个信号和 20 万次操作仅约 4 美元。
- Jev 的低价使实时 AI 循环变得可行，Claire 构建的语音应用将短语转为颜色并匹配情感引语，Jev 决策速度极快，引语 API 反成最慢环节。
- Claire 用 Jev 分类 4,500 条 YouTube 评论，识别出 58 条包含剧集创意，并构建了不到一秒扫描全数据集的搜索，发现对 Grok vs Muse 对比的强烈需求。
对实践的启示
从业者应识别工作流中只需决策的环节，用 Jev 等决策模型低成本处理分类、路由和过滤，再将复杂推理交给前沿模型。同时，模型个性、价格和速度与智能同样重要，跨模型互审可提升质量。
作者承认的反方与局限
Jev 不会写文档或设计界面，只适合排序、路由、排名和过滤；Opus 5.5 在消费级应用美学、计算机使用和视频编辑上仍弱于 Codex，且安全限制可能拒绝某些请求。
原文摘引
Jev is a decision model, not a language model, and that distinction can make many tasks dramatically cheaper.
The real skill is recognizing where a pipeline only needs a decision.
A model’s personality can matter just as much as its intelligence.
▸ 展开原文全文（39 段，抓取自 Lenny's Newsletter）
Jev for beginners: how to use it and what to build
Listen now on YouTube • Spotify • Apple Podcasts
Brought to you by:
OpenArt—An all-in-one AI creation platform for images, videos, music, audio, and more
In this solo episode, Claire tests Jev, TypeSafe AI’s new decision model that returns structured choices, scores, and probabilities instead of generated text. She uses it to analyze 1,700 pull requests for 9 cents, map her Claude and Codex usage, triage email, search 4,500 YouTube comments, and process 200,000 classifications for about $4. She also explains why Jev works best alongside a frontier model and how its speed and pricing make entirely new kinds of real-time apps and large-scale analysis practical.
Biggest takeaways:
- Jev is a decision model, not a language model, and that distinction can make many tasks dramatically cheaper. Instead of generating text, it returns predefined values such as a category, score, or probability. Claire believes this covers roughly 90% of what many software workflows actually need, at 4 cents per million input tokens with no output-token fee.
- It cost Claire 9 cents to understand where two years of engineering work went. She used Jev to compare 1,700 ChatPRD pull requests across 17,000 pairs, then had Gemini Flash Lite label the resulting clusters. In about two minutes, she learned that nearly 30% of the company’s engineering work had gone toward platform, security, and infrastructure.
- Some of the most useful analysis is already sitting on a local computer. Claude Code and Codex store past sessions locally, allowing Jev to classify them in minutes. Claire discovered that engineering had fallen from nearly 100% of her AI usage in January to less than 40% by September, with agents and media publishing filling the gap.
- Jev becomes far more powerful when paired with a frontier model. Claire uses Jev to classify, cluster, filter, and route large datasets, then sends only the most important groups to GPT-6 Astra for deeper reasoning. For ChatPRD’s product insights graph, this approach processed 1,100 signals and completed 200,000 operations for about $4 on the Jev side.
- Jev’s pricing changes which ideas are worth building. Because it returns small predefined values instead of generating long responses, TypeSafe charges nothing for output tokens. Claire spent less than $10 on Jev during the week, making classification workloads that would normally be expensive at scale feel almost free.
- Jev makes real-time AI loops practical. Claire built a voice app that turns a spoken phrase into a color, matches it with a quote based on sentiment, and displays everything almost instantly. Jev made its decisions so quickly that the quote API became the slowest part of the workflow.
- YouTube comment analysis is an immediate use case for any podcast team. Claire classified 4,500 How I AI comments by sentiment, identified 58 containing episode ideas, and built a keyword search that scans the full dataset in under a second. The results showed strong demand for a Grok versus Muse comparison and an 80% positive response to the “Claude Code for product managers” episode.
- The real skill is recognizing where a pipeline only needs a decision. Jev will not write documentation or design an interface, but it can sort, route, rank, and filter enormous datasets quickly and cheaply. Claire now asks one question before every build: Where does this workflow simply need to make a decision? That is where Jev belongs.
Blog and detailed workflow walkthroughs from this episode:
Jev: AI Data Analysis and Product Insights: https://www.chatprd.ai/how-i-ai/jev-ai-data-analysis-product-insights
↳ Jev GitHub PR Analysis: https://www.chatprd.ai/how-i-ai/workflows/jev-github-pr-analysis
↳ Jev YouTube Comment Analysis: https://www.chatprd.ai/how-i-ai/workflows/jev-youtube-comment-analysis
↳ Jev Multi-Model Product Insights: https://www.chatprd.ai/how-i-ai/workflows/jev-multi-model-product-insights
I left Claude for months. Opus 5.5 is why I’m back.
Listen now on YouTube • Spotify • Apple Podcasts
Claire tests Claude Opus 5.5 after months of leaving Claude out of her daily workflow. She puts it through long-running agentic tasks, frontend prototyping, writing, SVG illustration, computer use, and video editing to see where it earns a place back in her stack. She also shares why she is pairing it with Codex for cross-model code review, where Claude’s safety limits still get in the way, and which tasks remain firmly in Codex territory.
Biggest takeaways:
- A model’s personality can matter just as much as its intelligence. Claire stopped using Claude for months because its rambling, preachy, and overly verbose replies made it unpleasant to work with. Opus 5.5 is the first model in the family that no longer makes her blood boil, which is a meaningful improvement even if no benchmark captures it.
- Opus 5.5’s lower price and faster performance make long-running agent work more practical. It is 40% cheaper than Opus 5, and Claire found it noticeably faster. It successfully completed four complex tasks spanning inbox triage, backend development, research, and computer use, including runs of up to 82 steps from a single prompt.
- Silence during long-running tasks creates its own user experience problem. Opus 5.5 sometimes remains quiet for eight or nine minutes, leaving users unsure whether it is still working. It is a reminder that perceived latency matters alongside actual latency, especially when agents run for extended periods.
- Opus 5.5 is the strongest frontend designer Claire has tested so far. Its ChatPRD homepage redesign was bold and polished enough that she plans to ship it. The model handles hierarchy, white space, and visual rhythm exceptionally well, though it still struggles with consumer-app aesthetics and defaults to “Claude orange” without direction.
- SVG illustration is an unexpected strength of Opus 5.5. It was the only model Claire tested that produced clean, charming, and animatable character SVGs with consistent styling across multiple expressions. The characters remained visually coherent, and their anatomy mostly made sense.
- Opus 5.5 has a clear safety posture, and sometimes that means saying no. It refused when Claire asked it to skip testing and push directly to production, and it may route cybersecurity work to Opus 4.8. Whether that feels reassuring or frustrating depends on the workflow, but its boundaries are consistent.
- The best use of Opus 5.5 may be as an adversarial reviewer for another model. Claire now has Codex and Opus review each other’s work rather than using one to replace the other. This cross-model loop catches issues either model might miss alone, making the additional cost worthwhile when quality matters.
- Computer use and video editing still belong to Codex in Claire’s workflow. Opus 5.5’s ElevenLabs MCP video test produced weak color grading, too few jump cuts, and sloppy overlays. Codex also remains stronger at computer use in her current setup, giving her no reason to shift either category to Claude.
- Claude is back, but it has not replaced Codex as Claire’s daily driver. Opus 5.5 has earned a role in pull-request reviews, architecture questions, and frontend development. Codex’s desktop experience, computer use, and workflow integration still keep it in the primary position.
Blog and detailed workflow walkthroughs from this episode:
Claude Opus 5.5 Review: https://www.chatprd.ai/how-i-ai/claude-opus-5-5-review
↳ Claude Opus 5.5 SVG Illustrations: https://www.chatprd.ai/how-i-ai/workflows/claude-opus-5-5-svg-illustrations
↳ Claude Opus 5.5 Frontend Prototypes: https://www.chatprd.ai/how-i-ai/workflows/claude-opus-5-5-frontend-prototypes
Opus 5.5 vs. GPT-6 Sol: which model won my blind taste test?
Listen now on YouTube • Spotify • Apple Podcasts
Claire takes the How I AI bench live to compare GPT-6 Astra, GPT-6 Sol, Claude Opus 5.5, and more across the work she actua