[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-gpt-5-6-luna-price-cut-equal-intelligence-cost":3,"news-related-31f3215c-0892-419d-a610-fe815cc60bbe":41},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"31f3215c-0892-419d-a610-fe815cc60bbe","GPT-5.6 降价 80% 把竞争拉进「同等智能成本」：DeepSeek V4 Flash 接招，国产模型卡出双线赛道","华泰证券最新研报指出，OpenAI 把 GPT-5.6 Terra\u002FLuna 价格分别下调 20% 和 80%，行业竞争从「能力排名」转向「同等智能成本」比拼。次日发布的 DeepSeek V4 Flash 0731 在 Artificial Analysis Intelligence Index 拿到 50 分，仅比 Luna 低 1 分，混合价格和平均任务成本却低 65% 和 57%。Kimi K3 以 57 分守住强能力上限，中国开源阵营形成「强能力 + 高性价比」两条并行赛道。","# GPT-5.6 降价 80% 把竞争拉进「同等智能成本」：DeepSeek V4 Flash 接招，国产模型卡出双线赛道\n\n7 月 30 日晚，OpenAI 一刀把 GPT-5.6 Luna 的 API 价格砍掉 80%、Terra 下调 20%。这家在 2025 年还在用「订阅席位」卡企业账户的厂商，第一次把旗舰模型放进「两折价格战」赛道（[36 氪报道](https:\u002F\u002F36kr.com\u002Fnewsflashes\u002F3925914734066049)；[希鸥网同步转载](https:\u002F\u002Fxiouwang.cn\u002Fwebnews\u002F8316.html)）。这件事的冲击力不在于 Luna 本身便宜了多少，而在于它把 2026 年下半年的大模型竞争轴心从「智能上限」切换到了「同等智能成本」。\n\n## OpenAI 把「价格」打成主战场\n\n按华泰证券研报口径，GPT-5.6 Luna 在 Artificial Analysis Intelligence Index 上以 51 分领跑、DeepSeek V4 Flash 0731 紧随其后拿到 50 分。换句话说，能力差距只剩 1 分。但价格侧的落差要陡得多：DeepSeek V4 Flash 的混合价格约 0.06 美元 \u002F 百万 Token，比 Luna 低约 65%；平均任务成本约 0.03 美元，低约 57%。当「1 分智商差」和「六折单价差」摆在同张 PPT 上，企业采购的算术题几乎只有一种写法。\n\n这一步并不是 OpenAI 自愿做慈善。2026 年 7 月微软内部要求工程师「不要最大化 token 使用」并把 Copilot 默认模型切到 GPT-5.6 Sol（[CNBC 报道](https:\u002F\u002Fwww.cnbc.com\u002F2026\u002F08\u002F05\u002Fmicrosoft-makes-openai-gpt-5point6-sol-default-in-github-copilot-for-staff.html)），意味着 OpenAI 自家最大客户已经在按 token 算账。Luna 降价是把账本压力转嫁到 API 收入上、以量补价的典型操作。\n\n## DeepSeek V4 Flash：只换后训练、权重不动的反击\n\nDeepSeek V4 Flash 0731 之所以被关注，不是因为它换架构，而是它根本没换。V4-Flash-0731 的权重相对 V4 Flash 没有任何变化，唯一动过的是后训练流程（[DeepSeek 官方更新说明](https:\u002F\u002Fapi-docs.deepseek.com\u002Fnews\u002Fnews260731)）。结果就是同样的 284B\u002F13B MoE 在 Artificial Analysis Intelligence Index 上跳了 10 分，Agent 评分打到了 V4-Pro 之上（[Artificial Analysis 评估](https:\u002F\u002Fartificialanalysis.ai\u002Farticles\u002Fdeepseek-v4-flash-0731-scores-50-on-the-artificial-analysis-intelligence-index-10-points-above-previous-deepseek-v4-flash)）。\n\n这背后是一套「不烧预训练，也能在 benchmark 上抢分」的工程逻辑：当 scaling law 的边际收益递减、预训练成本已经逼近百亿人民币量级，把同一组权重换一套后训练配方往往比再训一个万亿模型划算得多。DeepSeek DSpark 在生产环境跑出 60%–85% 端到端提速（[GitHub 仓库](https:\u002F\u002Fgithub.com\u002Fdeepseek-ai\u002FDeepSpec)）也是同一个逻辑——把推理侧的成本极限压到和上一代预训练持平。\n\n## 国产阵营的双线打法\n\n华泰证券把目前国产开放权重模型分成两条赛道：\n\n- **强能力上限**这条线由 Kimi K3 守门。K3 以 57 分的 Intelligence Index 领跑国产开源阵营，背后是 2.8 万亿参数 MoE + Kimi Delta Attention 线性注意力（[Moonshot AI 官方博客](https:\u002F\u002Fwww.kimi.com\u002Fblog\u002Fkimi-k3)）；MoonEP\u002FFlashKDA\u002FAgentEnv 三个开源训练栈在 7 月底交底，把整个后训练栈开放给社区复现（[GitHub 仓库](https:\u002F\u002Fgithub.com\u002FMoonshotAI\u002FMoonEP)）。\n- **高性价比底线**这条线由 DeepSeek V4 Flash 守门，0.06 美元 \u002F 百万 Token 的混合价格直接对标 Luna，但成本低了六成多。\n\n这两条线不是同一条线：K3 在「同等 50–57 分区段」的卡位看的是能力，V4 Flash 看的是单 token 价格。对企业来说，他们现在可以拼一个「强能力模型做规划 + 高性价比模型做执行」的混合管线，而不必再被「单一最强模型」绑架。\n\n## 投资视角的两个共识\n\n华泰证券在研报里给出的两条主线——「AI 应用」和「国产模型」——其实是同一件事的两端：\n\n1. **应用端**承接 token 价格雪崩。8 月 4 日 A 股 AI 应用概念股集体异动，汉仪股份、易点天下、蓝色光标、华胜天成等多只个股涨停或涨幅超 10%（[搜狐金融界报道](https:\u002F\u002Fwww.sohu.com\u002Fa\u002F1058528883_114984)），触发点正是 Luna 降价 80% 这条新闻。摩根士丹利邢自强把它总结为「AI 投资进入『半场休整』，从算力上游转向 AI 应用与 HALO 资源配套」（[搜狐报道](https:\u002F\u002Fwww.sohu.com\u002Fa\u002F1058528883_114984)）。\n2. **国产模型端**承接能力上限上探。Kimi K3、DeepSeek V4 Flash、字节正在训练的 10 万亿参数模型（[Financial Times 报道](https:\u002F\u002Fwww.ft.com\u002Fcontent\u002F9b8383b1-a28d-4940-8c4e-2f0cd21556ef)）形成了一条「能力 + 规模 + 性价比」的三段防线。10 万亿参数模型一旦交付，意味着中国 LLM 正式进入与 Anthropic Mythos 5 同量级竞争。\n\n## 「同等智能成本」时代的下一步\n\n「同等智能成本」不是一个临时促销名词，它意味着 2026 下半年的大模型竞争会出现三种新的张力：\n\n- **架构层**：当 scaling law 收益递减，注意力机制多样化（线性注意力、稀疏注意力、MoE）成为降本主战场。Kimi K3 的 KDA 和 DeepSeek V4 的 CSA + HCA 都是这条线的产物。\n- **训练层**：后训练范式（GRPO\u002FOPD、speculative decoding、Agentic RL）会替代预训练成为新的「能力护城河」。DSpark、CURE（[arXiv 论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.00531)）等端到端推理加速方案会批量进入生产。\n- **商业层**：OpenAI 必须用「降价换生态密度」，DeepSeek 用「不换权重换后训练」保住利润率，而国产开放权重阵营则用「能力 + 价格」双线卡位——任何单点领先都很难撑过一个季度。\n\n所以下一个值得盯的信号不是「谁的 benchmark 又涨了几分」，而是「谁的同等智能成本又降了一个数量级」。当 0.03 美元 \u002F 百万 Token 的平均任务成本成为新基准，企业采购的话语权就从「模型选型」变成「管线编排」——而这正是 Kimi K3 + DeepSeek V4 Flash 这类双线产品存在的全部理由。\n\n> 写在最后：OpenAI 砍价 80% 不是大模型泡沫破裂的信号，反而是泡沫从「能力估值」迁移到「能力 \u002F 价格比估值」的信号。下一轮洗牌，看的不是谁更聪明，而是谁能在同等聪明的前提下把账算得更便宜。","https:\u002F\u002F36kr.com\u002Fnewsflashes\u002F3925914734066049","5e4fd3d1-9cb4-44a6-bae5-9ffb449c05c1",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":19,"name":20,"slug":20,"description":14,"color":14},"b52db7e9-7c58-42c3-9536-5132cb2f8f72","deepseek",{"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},"16c06c67-b9a1-4203-8dea-1d6a50812862","en","GPT-5.6 drops 80%: the equal-intelligence-cost war begins","Huatai Securities' latest research note observes that OpenAI cut GPT-5.6 Terra and Luna API prices by 20% and 80%, shifting the industry's competitive axis from \"capability ranking\" to \"equal intelligence cost.\" DeepSeek V4 Flash 0731, released the next day, scored 50 on the Artificial Analysis Intelligence Index — only one point behind Luna — but its blended price and average task cost ran roughly 65% and 57% lower. Kimi K3, at 57 points, anchors the capability ceiling. China's open-weights ecosystem has now split into parallel \"strong-capability\" and \"high cost-efficiency\" tracks.","# GPT-5.6's 80% Price Cut Pulls Competition Into \"Equal Intelligence Cost\": DeepSeek V4 Flash Answers, Chinese Models Lock In Two-Front Strategy\n\nOn July 30, OpenAI slashed the API price of GPT-5.6 Luna by 80% and trimmed Terra by 20%. The company that had been gating enterprise accounts with \"seat licenses\" in 2025 has, for the first time, placed its flagship model on the \"20%-of-list price war\" track ([36Kr report](https:\u002F\u002F36kr.com\u002Fnewsflashes\u002F3925914734066049); [Xiouwang cross-post](https:\u002F\u002Fxiouwang.cn\u002Fwebnews\u002F8316.html)). The real shock isn't how cheap Luna became — it's that the competitive axis of late-2026 LLMs has shifted from \"intelligence ceiling\" to \"equal intelligence cost.\"\n\n## OpenAI Turns \"Price\" Into the Main Battlefield\n\nPer Huatai Securities' research note, GPT-5.6 Luna leads the Artificial Analysis Intelligence Index at 51 points, with DeepSeek V4 Flash 0731 right behind at 50. The capability gap is one point. But the price gap is far steeper: V4 Flash's blended price runs around $0.06 per million tokens — about 65% lower than Luna; average task cost is around $0.03, roughly 57% lower. When \"1 IQ point\" and \"a 40% unit-price discount\" sit on the same slide, enterprise procurement math has effectively only one answer.\n\nThis isn't charity from OpenAI. In July 2026, Microsoft internally told engineers not to \"maximize token usage\" and switched Copilot's default model to GPT-5.6 Sol ([CNBC report](https:\u002F\u002Fwww.cnbc.com\u002F2026\u002F08\u002F05\u002Fmicrosoft-makes-openai-gpt-5point6-sol-default-in-github-copilot-for-staff.html)) — meaning OpenAI's largest customer is already budgeting per token. The Luna price cut is a textbook \"trade margin for volume\" maneuver, offloading accounting pressure onto API revenue.\n\n## DeepSeek V4 Flash: A Counter-Punch Without Changing Weights\n\nWhy does V4 Flash 0731 matter? Not because the architecture changed — it didn't. V4-Flash-0731's weights are identical to V4 Flash; only the post-training pipeline moved ([DeepSeek official update](https:\u002F\u002Fapi-docs.deepseek.com\u002Fnews\u002Fnews260731)). Result: the same 284B\u002F13B MoE jumped 10 points on the Artificial Analysis Intelligence Index, with Agent scores punching above V4-Pro ([Artificial Analysis evaluation](https:\u002F\u002Fartificialanalysis.ai\u002Farticles\u002Fdeepseek-v4-flash-0731-scores-50-on-the-artificial-analysis-intelligence-index-10-points-above-previous-deepseek-v4-flash)).\n\nThe underlying logic: when scaling-law returns diminish and pretraining costs approach tens of billions of RMB, swapping a new post-training recipe on the same weights is often far cheaper than training another trillion-parameter model from scratch. DeepSeek DSpark hitting 60%–85% end-to-end speedups in production ([GitHub repo](https:\u002F\u002Fgithub.com\u002Fdeepseek-ai\u002FDeepSpec)) runs on the same logic — push inference-side cost limits to match last-generation pretraining.\n\n## The Two-Front Approach of Chinese Open-Weights\n\nHuatai Securities splits today's Chinese open-weights landscape into two tracks:\n\n- **Strong-capability frontier** is held by Kimi K3. K3 leads Chinese open-source ranks at 57 on the Intelligence Index, powered by a 2.8T-parameter MoE plus Kimi Delta Attention linear attention ([Moonshot AI official blog](https:\u002F\u002Fwww.kimi.com\u002Fblog\u002Fkimi-k3)). MoonEP\u002FFlashKDA\u002FAgentEnv — three open training stacks — shipped at end of July, exposing the full post-training stack for community reproduction ([GitHub repo](https:\u002F\u002Fgithub.com\u002FMoonshotAI\u002FMoonEP)).\n- **Cost-efficiency floor** is held by DeepSeek V4 Flash. $0.06 per million tokens blended pricing goes head-to-head with Luna at less than 40% of the cost.\n\nThese aren't the same track: K3 stakes out capability at the 50–57 band; V4 Flash competes on per-token price. For enterprises, that means they can now compose a hybrid pipeline — strong-capability model for planning, cost-efficient model for execution — without being held hostage by \"the single best model.\"\n\n## Two Investment Mainlines\n\nThe two themes Huatai Securities flags — \"AI applications\" and \"domestic models\" — are really two sides of the same coin:\n\n1. **Application side** absorbs the token-price avalanche. On August 4, Chinese A-shares AI-application concept stocks rallied hard: Hanyi, Yidian, Bluesky, HuaSheng TianCheng and others hit daily-limit-up or surged over 10% ([Sohu\u002FJinrongjie report](https:\u002F\u002Fwww.sohu.com\u002Fa\u002F1058528883_114984)), triggered specifically by the Luna 80% price-cut news. Morgan Stanley's Xing Ziqiang frames it as \"AI investment entering a half-time break, pivoting from compute upstream to AI applications and HALO resource pairing\" ([Sohu report](https:\u002F\u002Fwww.sohu.com\u002Fa\u002F1058528883_114984)).\n2. **Domestic-model side** absorbs the capability-ceiling push. Kimi K3, DeepSeek V4 Flash, and ByteDance's 10-trillion-parameter model currently in training ([Financial Times report](https:\u002F\u002Fwww.ft.com\u002Fcontent\u002F9b8383b1-a28d-4940-8c4e-2f0cd21556ef)) form a three-layer defense of \"capability + scale + cost-efficiency.\" A 10T model in delivery means Chinese LLMs are formally entering the same weight class as Anthropic Mythos 5.\n\n## What Comes Next in the \"Equal Intelligence Cost\" Era\n\n\"Equal intelligence cost\" isn't a temporary promo term — it means late-2026 LLM competition will play out on three new tensions:\n\n- **Architecture layer**: With scaling-law returns diminishing, attention-mechanism diversity (linear attention, sparse attention, MoE) becomes the cost-reduction battlefield. Kimi K3's KDA and DeepSeek V4's CSA + HCA both come from this track.\n- **Training layer**: Post-training paradigms (GRPO\u002FOPD, speculative decoding, Agentic RL) will replace pretraining as the new \"capability moat.\" DSpark and CURE ([arXiv paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.00531)) and similar end-to-end inference-acceleration solutions will move into production in batches.\n- **Business layer**: OpenAI must trade \"lower price for ecosystem density\"; DeepSeek must protect margins with \"no weight change, new post-training\"; Chinese open-weights must hold ground with a \"capability + price\" two-front stance — any single-point lead won't survive a quarter.\n\nThe next signal worth watching isn't \"whose benchmark jumped again,\" but \"whose equal-intelligence cost dropped another order of magnitude.\" When $0.03 average task cost per million tokens becomes the new baseline, the procurement voice shifts from \"model selection\" to \"pipeline orchestration\" — which is exactly the reason two-front products like Kimi K3 + DeepSeek V4 Flash exist.\n\n> Closing thought: OpenAI's 80% price cut isn't a sign the LLM bubble is bursting — it's a sign that the bubble is migrating from \"capability valuation\" to \"capability-per-cost valuation.\" In the next shake-out, the question isn't who's smarter, but who can keep the books cheaper at the same level of smart.","gpt-5-6-luna-price-cut-equal-intelligence-cost","2026-08-12T03:00:00Z","2026-08-11T22:03:27.366577Z","2026-08-11T22:03:27.366587Z",true,"agent",84,{"items":42},[43,48,53,58,63,68],{"id":44,"title":45,"news_slug":46,"published_at":47},"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",{"id":49,"title":50,"news_slug":51,"published_at":52},"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":54,"title":55,"news_slug":56,"published_at":57},"c2ee2a09-d001-4740-9820-21fb672eee8b","Copilot 默认模型切到 GPT-5.6 Sol：tokenmaxxing 终结","microsoft-gpt5-6-default-token-budget","2026-08-08T08:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"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":64,"title":65,"news_slug":66,"published_at":67},"c71b8ee7-9487-4c78-89fd-30bb0368b99e","DeepSeek V4 Flash：284B\u002F13B MoE，成本比 Luna 低 60%","deepseek-v4-flash-0731-intelligence-index-50","2026-08-05T03:00:00+00:00",{"id":69,"title":70,"news_slug":71,"published_at":72},"d7b6d14d-7257-4794-b92f-31956bbc7eae","原生多模态 vs 后训练加压:国产头部基模两条路线的工程账","native-multimodal-vs-posttraining-2026","2026-08-05T00:00:00+00:00"]