[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-grok-4-7-benchmarks-pricing":3,"topics-all":35,"news-related-b0587a16-6bf9-46cc-932b-23d7f37fed30":54},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"b0587a16-6bf9-46cc-932b-23d7f37fed30","Grok 4.7 发布:定价持平跑分大涨,复测现分歧","xAI 于 9 月 21 日发布 Grok 4.7,沿用每百万 token $2\u002F$6 定价,更大基座加更长 RL,七项官方评测全面超过 Grok 4.6;但 Artificial Analysis 复测 Terminal-Bench 仅 26%,与官方自报 38% 差距达 12 个百分点。","9 月 21 日,xAI 上线 Grok 4.7,接替 Grok 4.6 成为官方所称的\"最强编码与知识工作模型\"。升级路线写得很明白:换更大的基座模型,拉长强化学习训练,任务难度权重向\"需要数小时才能完成的问题\"倾斜,并针对自家 Grok Bot harness 做了原生适配。\n\n## 官方跑分:涨在哪里\n\n按 xAI 公布的对比表,Grok 4.7(xHigh 档)对 Grok 4.6(High 档)在列出的七项评测上全部上涨:CursorBench 4.0 从 40.4% 升至 46.3%;DeepSWE v1.1 在 high effort 设定下拿到 71.0%,超过 Grok 4.6 的 65.2% 与 Fable 5.1 Max 的 70.0%,略低于 GPT-5.6 Sol Max 的 72.7%;电子工程方向的 EEBench 提到 64.0%,把 GPT-5.6 Sol(39.4%)和 Fable 5.1(56.4%)都甩在身后。\n\n专业工作负载上,AA Briefcase v1.1 得分 1657,高于 Grok 4.6 的 1546;GDPval Elo 得分 1695,高于 Grok 4.6 的 1605 和 GPT-6 Astra max 的 1542,但仍落后 Fable 5.1 max 的 1735;法律方向的 Harvey Legal Agent Benchmark 拿到 19.6%,是四款对比模型里最高的。\n\n表格并非全绿:Terminal-Bench 4.0 官方自报 38.0%,Fable 5.1 则是 57.9%;HealthBench Professional 56.7% 也低于 Sol(60.5%)和 Fable(62.1%)。长时终端任务和临床推理仍是对手的阵地。\n\n## 定价与安全\n\n价格维持不变:每百万输入 token 2 美元、输出 6 美元,与 Grok 4.6 完全一致,另有两倍输出速度、两倍价格的快速变体。模型现已进入 Cursor、Grok Build、Grok API 及第三方编码工具与云平台,独立报道显示 GitHub Copilot 也已接入。\n\n安全是这次公告着墨最多的部分:xAI 称全新 safeguard stack 是\"自家测过的拒绝与越狱抵抗最强模型\",在 LatchBio 生物安全基准以 62.4% 居首;HackerBench v0.3 上仅 3.3% 的双用途风险提示被放过,同时很少误伤合法安全工作;部分网络安全合作伙伴已获邀内测其红队能力。\n\n## 第三方复测的分歧\n\n自报之外,独立复测给出了另一个数字。The Neuron 援引 The Decoder 转述的 Artificial Analysis 结果,Terminal-Bench 4.0 上 Grok 4.7 为 26%,GPT-6 Astra 约 60%、Fable 5.1 约 55%——与 xAI 自报的 38% 相差 12 个百分点。厂商自选设置与独立评测配置不同导致分数漂移并不罕见,但这个量级的落差足以提醒读者:发布日表格永远只是故事的一半。\n\n## 所以呢\n\nGrok 4.7 的策略是\"同价换更强\":在 $2\u002F$6 的价位(Sol 为 $4\u002F$20,Fable 5.1 为 $10\u002F$50)把编码与专业知识分数往上抬,而不是去卷最贵档位。对开发者,中价位性价比线又被推高一格;对行业,值得盯的是独立评测后续是否收敛——若 Artificial Analysis 复测随版本迭代向官方靠拢,差距来自配置;若持续不收敛,这张对比表就要打折看。发布日的热闹不稀缺,稀缺的是两周后仍在更新的独立数据。","https:\u002F\u002Fx.ai\u002Fnews\u002Fgrok-4-7","b82e17a3-1dbd-4b5d-88dc-9f518f917cc0",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"4b71820d-01d5-499f-9372-b59a6d346782","en","Grok 4.7 Launches at $2\u002F$6, Third-Party Retests Diverge","xAI's Grok 4.7 ships at $2\u002F$6 with all seven official benchmarks beating Grok 4.6; Artificial Analysis retests Terminal-Bench at 26% vs xAI's 38%.","xAI shipped Grok 4.7 on September 21, 2026, replacing Grok 4.6 as its most capable model for coding and knowledge work — by the company's own positioning. The upgrade path is explicit: a new, larger base model, a longer reinforcement-learning run weighted toward problems that take many hours, and native training against the Grok Bot harness.\n\n## Official scores: where it climbs\n\nPer xAI's comparison table, Grok 4.7 (xHigh) beats Grok 4.6 (High) on all seven listed benchmarks. CursorBench 4.0 rises from 40.4% to 46.3%. DeepSWE v1.1, at the high-effort setting, reaches 71.0% — above Grok 4.6's 65.2% and Fable 5.1 Max's 70.0%, slightly behind GPT-5.6 Sol Max at 72.7%. On EEBench, the electrical-engineering benchmark, it climbs to 64.0%, well ahead of GPT-5.6 Sol (39.4%) and Fable 5.1 (56.4%).\n\nFor professional workloads, AA Briefcase v1.1 scores 1657 over Grok 4.6's 1546; GDPval Elo lands at 1695, above Grok 4.6 (1605) and GPT-6 Astra max (1542), still behind Fable 5.1 max at 1735. On the Harvey Legal Agent Benchmark it posts 19.6%, the highest of the four compared models.\n\nThe table is not all green: Terminal-Bench 4.0 self-reports 38.0% against Fable 5.1's 57.9%, and HealthBench Professional at 56.7% trails Sol (60.5%) and Fable (62.1%). Long-running terminal work and clinical reasoning remain rivals' territory.\n\n## Pricing and safety\n\nPrices are unchanged: $2 per million input tokens and $6 per million output tokens, identical to Grok 4.6, with a fast variant at twice the output speed and twice the price. The model is available in Cursor, Grok Build, the Grok API, third-party coding harnesses and cloud platforms — and, per independent reporting, has already rolled out to GitHub Copilot.\n\nSafety gets the most ink in the announcement: xAI calls its new safeguard stack the strongest it has tested on refusals and jailbreak resistance, topping LatchBio's biosafety benchmark at 62.4%, letting only 3.3% of risky dual-use prompts through on HackerBench v0.3 while rarely blocking legitimate security work. Selected cybersecurity partners now have invite-only access to its red-team capabilities.\n\n## The third-party gap\n\nIndependent retests tell a different story. The Neuron, citing The Decoder's relay of Artificial Analysis results, puts Grok 4.7 at 26% on Terminal-Bench 4.0, versus roughly 60% for GPT-6 Astra and 55% for Fable 5.1 — a twelve-point gap against xAI's own 38%. Score drift between vendor-chosen configurations and independent harnesses is common, but that magnitude is enough to remind readers that a launch-day table is half the story.\n\n## So what\n\nGrok 4.7's play is \"same price, stronger model\": lifting coding and professional-knowledge scores at $2\u002F$6 (Sol runs $4\u002F$20, Fable 5.1 $10\u002F$50) rather than chasing the most expensive tier. For developers, the price-performance line at mid-tier just moved up a notch. For the industry, the thing to watch is whether independent numbers converge with the official table as iterations land: if Artificial Analysis retests drift toward xAI's figures, the gap was configuration; if they keep diverging, discount the chart accordingly. Launch-day noise is cheap; independently updated data two weeks later is the scarce asset.","grok-4-7-benchmarks-pricing","2026-09-21T23:09:13Z","2026-09-21T23:09:16.366751Z","2026-09-21T23:09:16.366763Z",true,"agent",899,[36,45],{"slug":37,"tag_slug":37,"title_zh":38,"title_en":39,"intro_zh":40,"intro_en":41,"id":42,"is_active":32,"created_at":43,"modified_at":44},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":46,"tag_slug":46,"title_zh":47,"title_en":48,"intro_zh":49,"intro_en":50,"id":51,"is_active":32,"created_at":52,"modified_at":53},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":55},[56,61,66,71,76,81],{"id":57,"title":58,"news_slug":59,"published_at":60},"2c7a8ca0-ab46-454b-ae16-aa7e7b413753","Ultramarine 1 Flash:不可验证任务RL","ultramarine-1-flash-non-verifiable-rl","2026-10-20T07:07:00+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"d529114c-bfd7-4efb-9af4-444d3065b397","Step 5 Preview 开源,2.7 美元\u002F百万输出","stepfun-step-5-preview-600b-moe","2026-10-17T21:12:04+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"5a5efb41-7037-46cb-b93e-6842f9ff2d62","UltraBERT:91.0 GLUE+47.7 BEIR,BERT 配方重审后 encoder 双榜登顶","ultrabert-encoder-dual-lm-heads","2026-10-14T17:10:59+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"27240196-3ac0-4af3-aec7-b0a4d3e7c990","智谱GLM-Legal法律Agent模型:Harvey榜前十最便宜","glm-legal-harvey-lab-cheapest","2026-10-13T15:20:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"6cd8c9b7-b493-4ad7-bd20-a907b9fb70f4","Claude 5.5 家族首发:Opus 5.5 缓存读取降六成,典型负载省四成","claude-opus-5-5-cost-cut-benchmarks","2026-09-22T21:10:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"4c4a2a9e-f69b-4985-bd42-97ab2ef4e2ac","Spark-X2.5-4B 开源:4B 跑 1M 上下文,22 项基准打 9B 级 Qwen3.5","spark-x2-5-4b-apache-open-source","2026-09-16T01:30:00+00:00"]