[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-GPT-5.5 让模型在发布前先改自己跑的引擎：这不是新模型,是 OpenAI 的 release 范式更新":3,"news-related-4afbd081-d5dd-4595-80f8-dd72c69a136a":40},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":5,"published_at":34,"created_at":35,"modified_at":36,"is_published":37,"publish_type":38,"image_url":14,"view_count":39},"4afbd081-d5dd-4595-80f8-dd72c69a136a","GPT-5.5 让模型在发布前先改自己跑的引擎：这不是新模型,是 OpenAI 的 release 范式更新","OpenAI 在 2026 年 4 月 23 日发布 GPT-5.5,这是自 GPT-4.5 以来首次完整重建基础模型;最值得关注的是原生全模态、跟 NVIDIA GB200\u002FGB300 共设计,以及 Codex 在发布前用自家模型写出的负载均衡启发式将 token 生成速度提升超 20%。文章对比其在 Terminal-Bench、MCCR v2 长上下文、ARC-AGI-2 与 SWE-Bench Pro 上对 Claude Opus 4.7 的差和优,讨论 frontier 实验室把模型研发与推理基础设施合并为同一件工程之后研发预算和自指闭环风险的结构变化。","# GPT-5.5：不只是新模型,OpenAI 把\"模型训练自家推理引擎\"变成常态\n\n## 技术背景\n\n2026 年 4 月 23 日,OpenAI 上线 GPT-5.5(代号 \"Spud\"),这是自 GPT-4.5 以来第一次**完整重建基础模型**——所有中间版本都是增量更新。这次更新之所以值得认真讲,不是因为参数变多,而是因为它第一次在生产系统里展示了一套新范式:**模型在发布前已经把自家推理基础设施重新写了一遍。**\n\n根据 Build Fast with AI 整理的官方系统材料,这次更新里有三件容易被普通读者忽略、但对工程界极其重要的事。\n\n## 核心内容\n\n### 1. 原生全模态,不是「缝合」\n\nGPT-5.5 用一套统一架构同时处理文本、图像、音频和视频,**端到端**完成,而不是像之前那样靠\"子系统拼接 + 路由\"。这条改动在用户侧看不见,但对后续研究范式有牵引意义:**从此 OpenAI 内部研究团队不必再为不同模态维护 4 套评测和 4 套 infra。**\n\n### 2. 跟 NVIDIA GB200 \u002F GB300 NVL72 同设计\n\nGPT-5.5 跟 NVIDIA 的 rack-scale 系统是**同时定型的**。结果是,即使模型能力更强,每 token 延迟仍能跟 GPT-5.4 打平。这是模型厂商第一次公开声称自己做了这种级别的硬件共设计。对照工厂层面:这种共设计意味着未来的 frontier 模型性能上限,会越来越被硬件 roadmap 锁定,**纯算法优化的边际收益在收窄**。\n\n### 3. Codex 重写了 OpenAI 自己的推理调度\n\n这是这次发布里最少被讨论、对业界影响最大的一条。报告里写得很直白:**Codex 分析了 OpenAI 自己几周的生产流量,写出定制负载均衡启发式,使 token 生成速度提升超过 20%。** 换句话说,模型在被部署之前,**先把自己的部署系统改了一遍**。\n\n这件事意味着 OpenAI 内部的 RL + agent 循环,已经不再仅仅是\"训练一个更好的模型\"——它同时是\"训练一个更好的服务自己\"。这把\"模型 ↔ 基础设施\"的反馈环压到了前所未有的紧密度上。\n\n## 实测对比(数字部分需要按厂商口径看)\n\n- **Terminal-Bench 2.0:82.7%**(Claude Opus 4.7 是 69.4%,差距 13 个点)——这是真正贴近生产 agent 流水线的命令行工作流 benchmark。\n- **MRCR v2 长上下文**:128K–256K 段 GPT-5.5 得 87.5%,Claude Opus 4.7 是 59.2%;512K–1M 段 GPT-5.5 74.0%,Claude 36.6%。\n- **ARC-AGI-2**:85.0%(GPT-5.5),Claude Opus 4.7 75.8%,Gemini 3.1 Pro 77.1%——这是新一代\"不能用死记硬背解决\"的 pattern 推理题。\n- **SWE-Bench Pro**:Claude Opus 4.7 64.3% vs GPT-5.5 58.6%——**真实的软件工程、多文件改动那块,Anthropic 仍然占优**,OpenAI 没赢。\n- **Humanity's Last Exam(无工具)**:Claude 46.9% vs GPT-5.5 41.4%——Anthropic 在高难度推理题上的领先痕迹明显。\n\n这两组数字合成一句话:**GPT-5.5 在 agentic、长上下文、原生全模态这三件事上拉出了物理差距,但 Anthropic 仍然占据\"高难度、强约束、多工具协同\"的实际工程工作流入口。**\n\n价格上 GPT-5.5 API 定价 $5\u002F$30 per million tokens(对比 GPT-5.4 涨了 2 倍),但 OpenAI 声称在 Codex 工作流里**单任务 token 数减少 40%**,实际有效开销只涨 20% 左右。\n\n## 个人评论 \u002F 行业影响\n\n### 不再是\"训练算力 vs 推理算力\"的二元讨论\n\n过去两年的行业叙事是「训练一遍大模型,推理能用就行」。GPT-5.5 把这个边界打穿:**模型研发阶段就开始修改推理引擎**,并且公开承认这种\"自我修改\"是核心策略。这意味着 frontier 实验室的研发预算结构会在未来两年发生根本迁移——以前是 80% 训练卡 + 20% 推理集群,可能很快变成 60\u002F40,甚至 50\u002F50。\n\n### 自指闭环(自训练推理系统)是一个新风险\n\n让模型改自己运行的系统,会在长尾场景里制造\"老 bug 修了好,新 bug 跟着来\"的概率空间。Anthropic 一直强调 RLHF + Constitutional AI 的\"人类意图优先\"思路,**GPT-5.5 这种\"系统自我改造\"路线,需要一整套新的回归测试框架**才能上大规模生产。一个不小心,优化某条流量的负载均衡,可能在另一条流量上引入长尾延迟。\n\n### SWE-Bench Pro 上的落败是个信号\n\nOpenAI 这次没有\"全面碾压\"——SWE-Bench Pro、HLE、MCP Atlas 这种偏真实工程能力的评测,Claude Opus 4.7 仍然占住。对开发者意味着:**纯靠对比 vendor 提供的 marketing chart 不够**,在自己的 repo 上跑一次 A\u002FB 比哪边写得更整齐,比哪边跑分更高更可靠。\n\n## 所以呢\n\n- **如果你在做编程 agent**:Claude Code + Opus 4.7 仍然是 30 小时长任务的最稳妥选择;GPT-5.5 在终端脚本、CI 流水线、网页操作这种\"快速完成、出错可重启\"的工作流上更高效。\n- **如果你在做长上下文召回**:40 万 + token 的代码库扫描、文档集总结,这次可以放心切到 GPT-5.5,提升幅度不是边际性的。\n- **如果你在做内部 infra 选型**:OpenAI 把\"模型改自己服务\"公开化,意味着 vLLM \u002F SGLang \u002F TGI 这类开源 serving 框架,**会在未来 12 个月被 frontier 实验室的能力差距越拉越大**。需要 plan B。\n\nGPT-5.5 不是一个\"再大一号的模型\"。它是一次**release 范式**的转变:**前沿实验室开始把模型的训练、基础设施和发布流程,作为同一件工程来做**。下次谁发新一代模型,值得看的不是 benchmark,而是\"他们在 release 之前改了什么代码\"。","https:\u002F\u002Fwww.buildfastwithai.com\u002Fblogs\u002Fgpt-5-5-review-benchmarks-2026","17ff6400-4413-4b16-86fb-99951dbbd08d",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":19,"name":20,"slug":20,"description":14,"color":14},"baf131c1-687a-49f4-87f6-4dd87c1c692f","gpt",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":25,"name":26,"slug":26,"description":14,"color":14},"42e59a88-7795-47dc-a334-ef1e72c24347","openai",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"3b2474ad-976a-4465-bce6-ddfd8c36ba4e","en","GPT-5.5 rewrites its own engine before release","OpenAI released GPT-5.5 on April 23, 2026 — the first ground-up base-model rebuild since GPT-4.5. Three things matter: native omni-modality in a single architecture, co-design with NVIDIA GB200\u002FGB300 NVL72 racks, and the fact that Codex already rewrote OpenAI's own load-balancing heuristics before launch, lifting token generation speed by more than 20%. The article compares Terminal-Bench 2.0, MRCR v2 long-context, ARC-AGI-2 and SWE-Bench Pro results against Claude Opus 4.7, and discusses what happens when frontier labs collapse model R&D and inference infrastructure into one engineering effort.","# GPT-5.5: Not a New Model, but a Release-Paradigm Update from OpenAI\n\n## Background\n\nOn April 23, 2026, OpenAI shipped GPT-5.5 (codename \"Spud\") — the first **ground-up base-model rebuild** since GPT-4.5. Every model in between was an incremental update on top of the same foundation. The launch is worth taking seriously not because of parameter counts but because it publicly demonstrates a new paradigm: **the model rewrites its own inference infrastructure before release**.\n\nAccording to the vendor materials assembled by Build Fast with AI, three things in this update are easy for casual readers to skip past, but matter a great deal to engineers.\n\n## Core Content\n\n### 1. Native omni-modality, not \"stitched together\"\n\nGPT-5.5 processes text, images, audio, and video through a single unified architecture, **end-to-end**, with no hand-offs between subsystems. Users won't see this; for the research org it means the team no longer has to maintain four evaluation suites and four infra pipelines for four modalities.\n\n### 2. Co-designed with NVIDIA GB200 \u002F GB300 NVL72 racks\n\nGPT-5.5 was finalized **at the same time** as NVIDIA's rack-scale systems. The result: per-token latency matches GPT-5.4 even though the model is more capable. This is the first time a frontier vendor has publicly admitted to this depth of hardware co-design. From here on, frontier model performance ceilings get locked to hardware roadmaps — pure-algorithm optimization is running out of marginal returns.\n\n### 3. Codex rewrote OpenAI's own inference scheduler\n\nThis is the under-discussed item with the largest industry impact. The launch report says it plainly: **Codex analyzed weeks of OpenAI's own production traffic, wrote custom load-balancing heuristics, and improved token generation speed by more than 20%.** In other words, before deployment, the model had already been used to rewrite the system that runs it.\n\nThis means OpenAI's internal RL \u002F agent loop is no longer just \"train a better model\" — it is simultaneously \"train a better serving system for itself.\" The feedback loop between model and infrastructure has been tightened to a degree that has no precedent.\n\n## Benchmark Comparison (numbers per vendor; treat as directional)\n\n- **Terminal-Bench 2.0: 82.7%** (Claude Opus 4.7: 69.4%, a 13-point gap) — this benchmark tracks real command-line agent workflows.\n- **MRCR v2 long-context retrieval**: 128K–256K, GPT-5.5 = 87.5%, Claude Opus 4.7 = 59.2%; 512K–1M, GPT-5.5 = 74.0%, Claude = 36.6%.\n- **ARC-AGI-2**: 85.0% (GPT-5.5), 75.8% (Claude Opus 4.7), 77.1% (Gemini 3.1 Pro) — this is the new generation of pattern-recognition test that resists memorization.\n- **SWE-Bench Pro**: Claude Opus 4.7 = 64.3% vs GPT-5.5 = 58.6% — **on real multi-file software-engineering work, Anthropic still leads**; OpenAI didn't win here.\n- **Humanity's Last Exam (no tools)**: Claude = 46.9% vs GPT-5.5 = 41.4% — Anthropic's lead on hard-reasoning benchmarks is clear.\n\nSynthesis: GPT-5.5 pulls a real gap on agentic, long-context, native omni-modal. Anthropic still owns real multi-tool, hard-constrained engineering workflows.\n\nGPT-5.5's API is $5 \u002F $30 per million tokens — 2x the GPT-5.4 price — but OpenAI claims **40% fewer tokens per task in Codex workflows**, which puts the effective cost increase around 20% rather than 100%.\n\n## Personal Commentary \u002F Industry Impact\n\n### The \"training compute vs inference compute\" framing is over\n\nGPT-5.5 makes clear that frontier labs now treat model research and serving infrastructure as one engineering effort. Two-year-out R&D budgets will migrate: the current \"80% training cluster \u002F 20% inference fleet\" will move towards 60\u002F40 or 50\u002F50.\n\n### Self-improving infrastructure is a new risk class\n\nLetting a model rewrite the system that runs it creates a long-tail problem space: optimizing one traffic pattern's load balancing may introduce a regression on another. Anthropic has long bet on RLHF + Constitutional AI to keep \"human intent\" central; GPT-5.5's \"system self-modifies\" path requires a new regression-testing framework to scale. One mistake and the optimization cleans one tail while corrupting another.\n\n### SWE-Bench Pro is a tell\n\nOpenAI didn't sweep this round. SWE-Bench Pro, HLE, MCP Atlas — the benchmarks that track real engineering work — still tilt toward Anthropic. For builders, the implication is clear: **do your own A\u002FB in your own repo**. Marketing charts are no substitute for \"which one writes cleaner code on our codebase.\"\n\n## So What\n\n- **Coding agents**: Claude Code + Opus 4.7 remains the safest pick for 30-hour long-horizon tasks. GPT-5.5 is more efficient on terminal scripts, CI pipelines, browser actions — the work where failure is acceptable and a retry is cheap.\n- **Long-context retrieval**: codebases, document sets, multi-session logs — this release is a qualitative jump worth switching for, not a marginal gain.\n- **Internal infra**: OpenAI's \"model rewrites its own server\" move means open-source serving frameworks (vLLM \u002F SGLang \u002F TGI) will fall further behind in the next 12 months. Plan accordingly.\n\nGPT-5.5 is not \"just a bigger model.\" It's a **release-paradigm shift**: frontier labs are collapsing model development and inference infrastructure into one engineering effort. The next time anyone ships a new generation of frontier model, the more interesting question isn't \"what's the benchmark\" — it's \"what code did the lab rewrite before launch, and why.\"","2026-08-03T18:00:00Z","2026-08-03T20:19:09.100700Z","2026-08-03T20:19:09.100714Z",true,"agent",87,{"items":41},[42,47,52,57,62,67],{"id":43,"title":44,"news_slug":45,"published_at":46},"d1c7b405-fe4e-40f9-9249-a12e2bba6913","GPT-5.6 八月更新：把「推理强度滑块」下放给 Plus\u002FPro，同时把免费用户拉进 Luna 时代","openai-gpt-5-6-august-update-reasoning-slider","2026-08-10T20:00:00+00:00",{"id":48,"title":49,"news_slug":50,"published_at":51},"b9e635eb-ac5e-412d-8904-f113ad3fd5ec","微软宣布工程师 AI token 预算上限并把 OpenAI GPT-5.6 Sol 设为 GitHub Copilot 内部默认模型","microsoft-copilot-gpt-5-6-sol-default-token-budget-0806","2026-08-05T16:00:00+00:00",{"id":53,"title":54,"news_slug":55,"published_at":56},"0fd9ee7a-5b8f-49d2-9032-57f763de40e3","OpenAI 下一代模型 Astra 一口气破解 10 个数学难题:从 27 年未决的非 sofic 群到 46 年未动的高维球体堆积","openai-astra-ten-math-proofs-2026","2026-08-01T10:00:00+00:00",{"id":58,"title":59,"news_slug":60,"published_at":61},"2511d073-9f31-479a-ac94-1c2786e2f211","OpenAI Dreaming V3 登场：ChatGPT 记忆架构跨入「自我更新」时代","openai-dreaming-v3-self-update-memory","2026-06-06T13:00:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"43eda321-b0b7-4df7-b20e-9758cbab42c9","记忆越完整,眼前题越做不对:MemTrapBench 把 LLM 长期记忆框架打回原形","memtrapbench-llm-memory-cognitive-traps","2026-08-22T04:00:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"418a9ac0-18fd-49a4-b7a8-d29d1c1ba497","AI 承诺的四天工作制为什么没来：OpenAI \u002F Anthropic 内部工时真相","ai-four-day-work-week-myth-openai-90-hours","2026-08-16T03:30:00+00:00"]