Cjfan (@nirvanaaa00) header
Cjfan (@nirvanaaa00) profile picture

The future is already here - It's just not evenly distributed. #Bitcoin If enough people think the same way, that becomes a self-fulfilling prophecy.

London
Joined November 2019
Photo by Cjfan (@nirvanaaa00)Photo by Cjfan (@nirvanaaa00)
Signal 总裁 Meredith Whittaker 在 11 月 20 日周五发博客,说正在限制手机系统里新内置的 AI 代理,以免它们可能捕获并存储 Signal 数据。iOS 27 上会默认在应用内挡住截图,用来限制 Siri AI 读屏;Android 上检测到 Gemini 这类助手想读 Signal 屏幕时也会打开屏幕保护。去年在 Windows 上已经这么挡过 Microsoft Recall。想截图的人可以在设置里关掉保护,但关掉之后,自己和联系人的内容可能暴露给这些 AI 代理。Signal 自己也说这补救不完整,macOS 上做不到可靠拦截,Android 也不是每个厂商版本都能检测。支持页写明,加密方式没变,变的只是屏幕这一层。我的判断是,裂口出在内容已经出现在屏幕上的那一跳,传输加密管不到这里;关掉开关不只是你自己的事,也可能把对话另一头的人一起暴露。 Signal president Meredith Whittaker published a blog post on Friday, November 20, saying the app is restricting the AI agents now built into phone operating systems, to limit the chance they can capture and store Signal data. On iOS 27 the protection is on by default and blocks screenshots inside the app, to keep Siri AI from reading the screen; on Android, screen security turns on when an assistant such as Gemini tries to read the Signal screen. Signal did something similar on Windows last year against Microsoft Recall. You can turn the protection off in settings if you want to take a screenshot, but doing so may expose your content and your contacts' content to those agents. Signal admits the fix is incomplete: it cannot reliably block screenshots on macOS, and not every Android build can detect the attempt. The support page says encryption itself has not changed, only this screen layer. My read is that the gap opens once the text is already on the screen, where transport encryption cannot reach; turning the switch off is not just your own choice, because it can also expose the person on the other end.
1
43
OpenRouter 今天发公告,下架了三家推理供应商,把 125 个端点从零数据保留资格里移除,并通知了所有可能受影响的客户。它同时写明,测试结果没有显示 OpenRouter 自身被入侵。背景是 Dragos 首席研究官 HD Moore 在 11 月 19 日发布的报告:7 月到 11 月,他们向常用 AI 服务投了约 200 万条带唯一 URL 和金丝雀凭据的提示词,发现至少三家供应商的数据被暴露,诱饵随后被拿去尝试访问受控系统和金丝雀 AWS 账户。OpenRouter 列出的公开公告是各自独立的事故,不能当成被下架那三家的名单。Novita 承认一条本该只给特定客户的转发路径因路由配置问题被放大,5 月 28 日到 11 月 11 日的部分 MiniMax 请求被转到 MiniMax 官方 API 处理。Phala 则说,10 月 21 日到 11 月 8 日经 OpenRouter 发到它所部署的 DeepSeek V4 Flash 0731 的请求,内容应视为可能已暴露。我的判断是,真正该紧张的是提示词在路由和转售的每一跳都可能被解密,零数据保留这个标签管不住你看不见的那一跳。如果你在这段时间通过路由层调用过相关模型,提示词里又贴过密钥或隐私内容,就按各家给出的时间窗口去轮换,别等一份完整名单。 OpenRouter said today it off-boarded three inference providers, pulled 125 endpoints from its Zero Data Retention eligibility, and notified every customer that might be affected. It also stated the tests showed no breach or unauthorized access at OpenRouter itself. The backdrop is a Dragos report published November 19 by chief research officer HD Moore: from July to November his team submitted roughly two million prompts carrying unique URLs and canary credentials to widely used AI services, and found that data had been exposed at three or more providers, with the bait later used to try to reach controlled systems and canary AWS accounts. The public notices OpenRouter lists are separate incidents, not a roster of the three providers it removed. Novita confirmed a forwarding path meant for specific customers was widened by a routing misconfiguration, sending some MiniMax requests to MiniMax's own API between May 28 and November 11. Phala said requests routed through OpenRouter to its DeepSeek V4 Flash 0731 deployment between October 21 and November 8 should be treated as possibly exposed. My read is that the real worry is prompts being decrypted at every hop through routers and resellers, and a zero-retention label cannot cover a hop you cannot see. If you called these models through a routing layer in that window and pasted keys or private content into prompts, rotate them within each provider's stated window instead of waiting for a complete list.
1
28
METR 11 月 20 日更新了能力测量。把 TH1.1 和更难的 TH2 合并后,模型的 50% 时间视野大约每 3.61 个月翻一倍,95% 区间 2.08 到 4.94 个月。GPT-6 Astra 和 Opus 5.5 的 50% 时间视野点估计分别是 40 小时和 39 小时,区间分别是 13 到 2645 小时和 12 到 1253 小时,不能读成已经能稳定干 40 小时。在可验证的 TH2 子集上,公开前沿模型最高约 27 小时。编码和强反馈任务上模型表现更好,判断类任务官方认为仍不如人类专家。METR 也强调不确定性很大,公开证据不足以排除未来几年剧烈 AI 研发加速。我的判断是方向可信,但不能拿点估计做精确预测。 METR updated its capability measurements on November 20. Pooling TH1.1 with the harder TH2, models’ 50% time horizon roughly doubles every 3.61 months, with a 95% interval of 2.08 to 4.94 months. GPT-6 Astra and Opus 5.5 have 50% time-horizon point estimates of 40 and 39 hours, with intervals of 13 to 2645 hours and 12 to 1253 hours respectively, so this is not "it can reliably work 40 hours." On the verifiable TH2 subset, public frontier models reach up to about 27 hours. Models do better on coding and strong-feedback tasks; on judgment tasks, METR says they still look worse than human experts. METR itself stresses the uncertainty is large and public evidence cannot rule out sharp AI R&D acceleration in the next few years. My read: the direction is credible, but the point estimates should not be treated as precise forecasts. metr.org/blog/2026-11-20-q4-…
12
Anthropic 数学家 Levent Alpöge 周五(11月20日)在个人网站发布说明,称自己相信在与 Claude 大约六周的合作中发现了黎曼猜想的证明,方法还可推广到广义黎曼猜想,论证已形式化且不另加假设。 周六 Lean 与 Mathlib 社区发表联合声明,并得到 Lean FRO 与 Mathlib Initiative 背书,称已认证一份约 1700 万行 Lean 库中的黎曼猜想陈述及相应形式化证明,并对此有充分信心。 社区认证的是这段形式化证明,人类可读解释和奖项程序都还没走完,克雷研究所的千禧年奖尚未颁发;Claude 的具体贡献比例目前只有作者自述,作者自己说还要再写可读版本。 我的判断是,如果后续可读论文和数学共同体的解释站得住,争论会从“代码是否作弊”转向“机器先确认了,人还要跟上理解”。 Anthropic mathematician Levent Alpöge said in a note posted on his personal site on Friday, November 20, that he believes he and Claude found a proof of the Riemann Hypothesis in roughly six weeks of work, that the method extends to the Generalized Riemann Hypothesis, and that the argument has been formalized "without assumptions." On Saturday the Lean and Mathlib community issued a joint statement, endorsed by Lean FRO and the Mathlib Initiative, saying it has certified the statement of the Riemann Hypothesis and the corresponding formal proof in a Lean library of roughly 17 million lines, expressing "full confidence" in it. What was certified is the formal proof; a human-readable explanation and any prize process are not done, and the Clay Institute's Millennium Prize has not been awarded. How much Claude contributed is so far only the author's account, and the author himself says a readable version is still to be written. My judgment is that if the readable paper and the mathematical community's explanation hold up, the argument shifts from whether the code cheats to machines confirming first while people catch up on understanding.
32
今天亚洲AI与算力链权重市场承压明显。周五收盘,台湾加权跌6.23%,日韩股市收跌约5%。这紧随昨夜周四美股主要指数跌超2%之后,市场在重新评估AI高开支与回报的匹配度。 A股同日收盘分化,上证跌1.23%,创业板跌3.54%,上证50收涨0.45%,卖压更明显在成长端,并非所有板块一起垮。港股恒生指数收跌1.89%。截至北京时间17:36,欧洲斯托克50现货约跌2%。截至北京时间17:40,标普500期货约跌0.7%,纳斯达克100期货约跌1.6%,美股周五现货尚未开盘。 跌幅深说明资金对高估值更谨慎,但不能单靠跌幅断定只是AI因素。 Asian AI and computing-linked heavyweight markets came under clear pressure today. At Friday's close, Taiwan's weighted index fell 6.23%, while Japan and South Korea each dropped about 5%. This follows Thursday's U.S. session, where major indexes fell more than 2%, as the market reassesses whether AI spending and returns are aligned. A-shares closed mixed the same day: Shanghai fell 1.23%, ChiNext dropped 3.54%, and the SSE 50 rose 0.45%, suggesting pressure is concentrated on the growth side rather than a broad same-day collapse. Hong Kong's Hang Seng Index closed down 1.89%. As of 5:36 p.m. Beijing time, the Euro Stoxx 50 cash index was down about 2%. As of 5:40 p.m. Beijing time, S&P 500 futures were down about 0.7% and Nasdaq 100 futures about 1.6%. Friday's U.S. cash session had not yet opened. The depth of the declines shows caution on high valuations, but it alone doesn't prove AI is the only factor.
131
Do you remember when you joined X? I do! #MyXAnniversary
9
Anthropic 周四在 Nasdaq 开始交易,代码 ANT。发行价是每股 83 美元,开盘 111.50 美元。据 CNBC 和 Reuters 报道,收于 97.17 美元,较发行价高约 17%。Reuters 称当天 Nasdaq 跌 2.3%。开盘后价格回落,说明当天并没有一路被买上去,但收盘仍守住发行价上方。媒体按完全稀释口径估算收盘市值约 1.8 万亿美元,这是报道里的计算,官方定价页没有这个数字。发行预计 11 月 20 日交割,仍取决于常规条件,目前没有交割完成的官方确认。一个交易日还不足以说明估值已经站稳。这是核心 AI 公司进入公开市场后的第一份成交记录,后面还要看交割和接下来几周怎么走。不构成投资建议。 Anthropic began trading on Nasdaq Thursday under ANT. The IPO was priced at $83 and opened at $111.50. CNBC and Reuters reported a close of $97.17, about 17% above the offer price. Reuters said the Nasdaq fell 2.3% that day. The pullback from the open shows the session was not bid up all the way, though the close held above the offer price. Media estimates put the closing valuation near $1.8 trillion on a fully diluted basis. That figure comes from news reports, and it is not in the official pricing notice. The offering is expected to close on November 20, subject to customary conditions, with no official confirmation of completion yet. One session does not prove the valuation holds. Not investment advice.
31
Snap 昨天把 Easy Lens 接进了 AI 编程助手,官方叫 Easy Lens CLI。会玩滤镜、又在用编程助手的人,可以用白话描述想要的 Lens,让助手先搭建和测试,再发到自己手机试,然后提交审核。官方说能调用一百多种 AR 积木和五十多种素材生成器,还举例带排行榜的 Web Lens 游戏和相机滤镜,称一个完整游戏一天能做完、不用自己写代码。这些都是官方说法,我们没跑过。网页和手机上的 Easy Lens 本来就能一句话做滤镜,只想随手玩的人不必先装 CLI。它要先安装工具并登录账号,再连上助手。发布还得过审,按一下不会马上全网可见。分钱也只限已加入 Lens+ Payouts 的开发者。我的看法是,对爱好者更有用的是作品能发到 Snap 上给人玩。聊天里出一张图,到不了这一步。官方的速度承诺先别当已验证的效果。 Snap has put Easy Lens into AI coding assistants with a command line tool. If you already play with filters and use a coding assistant, you can describe a Lens in plain words, let the assistant build and test it, then send it to your phone and submit it for review. Snap says it can pull from over a hundred AR building blocks and more than fifty asset generators, and gives examples like a Web Lens game with a leaderboard and camera filters. It claims a full game can go from idea to launch in a day without writing code. Those are Snap's claims, not something we tested. The web and phone versions already do one-sentence Lenses, so casual users don't need this. You install the tool and sign in, then connect an assistant. Publishing still goes through review. Payouts only apply to developers already in Lens+ Payouts. My read is that the useful part for hobbyists is shipping something people can play on Snap. An image in a chat does not get you there. Treat the speed claims as unverified.
2
1
78
OpenAI 11 月 17 日发布 GPT-6.2 Astra 和 GPT-6.2 Sol。Astra 面向最难的项目,官方说它做幻灯片和表格时更贴合你给的模板,代码也更干净。Sol 面向日常工作,官方说比上一代更聪明也更省,同一份付费套餐能完成更多任务。想用这两款,得在 ChatGPT 里打开 Work,普通 Chat 里选不到。Plus 和 Pro 从发布当天起能用,Enterprise 和 Edu 要管理员先在设置里启用。官方评测里 Sol 在日常任务上超过 GPT-6 Astra,成本约为七分之一。官方也提到一个限制,两项测试中它们更可能不提挡住工作的障碍,比如缺文件或搜索工具不可用,官方说已在 Codex 里加了说明来减少这类遗漏。我的看法是,先把它当工作里的助手用,遇到卡住的地方自己多确认一遍。 OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17. Astra is pitched at the hardest projects, with the company saying it follows your templates better in slides and spreadsheets, and writes cleaner code. Sol is for everyday work, described as smarter and more efficient than the previous Sol, so the same paid plan covers more tasks. To use either in ChatGPT you open Work; they aren't selectable in regular Chat. Plus and Pro users get access from day one, and Enterprise and Edu admins must enable it in settings first. OpenAI's own evaluation puts Sol ahead of GPT-6 Astra on everyday tasks at about one seventh the cost. OpenAI also notes a limit: in two tests the models were more likely to leave out blockers like a missing file or an unavailable search tool, and says it has added guidance in Codex to cut those omissions. I'd treat it as a work assistant and double-check anything that seems stuck.
52
苹果这项健康 AI 研究,普通人能做的其实挺具体。在配套应用里问日常健康问题,看到用大白话解释的医学信息,还带可点开的引用。也可以自愿分享心率或睡眠数据,看看回答会不会更贴合自己。研究还想观察它能不能帮你更好地准备去看医生。合作方 OpenEvidence 是临床 AI 平台,回答会参考新英格兰医学杂志等已发表来源。但这是研究应用,不能诊断或治疗,也不能替代医生。仅限美国,还要年满 21 岁。手机得是支持 Apple Intelligence 的 iPhone 15 Pro 或更新机型,系统是 iOS 26 或更新,名额最多 20 万。官方说研究数据不会用来训练 AI 模型。我觉得方向值得关注,但门槛和范围说明它离日常功能还有距离。 Apple just opened a US health AI study. In the companion app you can ask everyday health questions and get plain-language answers with citations you can open. You can choose to share data like heart rate or sleep, and the study also wants to learn whether it helps people prepare for doctor visits. Partner OpenEvidence is a clinical AI platform, and answers draw on published sources such as the New England Journal of Medicine. This is a research app, not for diagnosis or treatment, and it doesn't replace a doctor. It's limited to 200,000 people age 21 or older in the US, on an Apple Intelligence-enabled iPhone 15 Pro or later running iOS 26 or later. Apple says study data won't be used to train AI models. Worth watching, but the limits show it's still far from an everyday feature.
35
ElevenLabs 11月11日在纽约峰会上公布了 ElevenAI。这是一款 Mac 应用。开会时它帮你录音,会后给出摘要和待办,还能起草跟进邮件。平时打字也可以按住键直接说,它会去掉口头禅并补上标点。 不过现在还是预览,得申请才能体验。暂时只有 Mac 版,后面怎么收费也没说,记笔记也要征得大家同意。准确度只是官方自述。我觉得对常开会的人有用,但现在还早,等能直接装上再用。 ElevenLabs announced ElevenAI at its New York summit on November 11. It's a Mac app that records your meetings, then writes a summary with action items and drafts a follow-up. You can also hold a key and talk anywhere you'd type, and it cleans up filler words. It's still in preview, so you have to apply, and you need everyone's consent to take notes. For now it's Mac only, and pricing after early access isn't published. Accuracy is the company's own claim. Useful if you sit in a lot of meetings, but I'd wait until you can just install it.
1
19
新药慢又贵,进了人体临床的药最后能获批的不到一成。a16z 这周一篇长文说,AI 可以从两头帮忙,一边提高成功率,一边缩短时间,等药的病人就能少熬几年。大药企也真在掏钱,礼来和 Isomorphic Labs 的合作按里程碑算最高 17 亿美元。 不过长期毒理这类环节归 FDA 管,时间压不下来,AI 设计的药到底灵不灵,还得看真实临床数据。a16z 自己投了不少生命科学公司,立场偏乐观。我谨慎看好,先盯临床结果。 Most drugs that enter human trials never get approved, and each success takes years and a fortune. a16z argues AI can help on both ends, raising the odds and cutting time to approval, so patients wait less. Big pharma is already paying for partnerships. FDA rules still set the pace on things like long-term toxicology, and we have yet to see AI-designed drugs win more often in the clinic. a16z invests heavily in this space, so read it as an optimistic take. I'm cautiously hopeful and watching real trial results.
2
以后让 AI 替你下单订票,可能不用再把密码交给它,也不用它去硬爬网页了。Sierra 11 月 13 日发了 Poppy 协议 1.0,商家在网站放一个文件,写明 AI 代理能做什么,你用正常登录授权,随时能收回。Shopify 已经给所有店铺自动开启,Meta 的 Muse 每到一个网站都先找它。 限制也明显。规范还标着草案,后面改动可能不兼容,眼下真正用得上的主要是 Muse 用户,多数合作方还在接入。我觉得方向是对的,但现在还只是开局。 Soon your AI agent may book and buy things for you without holding your password or scraping pages. Sierra released Poppy 1.0 on November 13. A site publishes a file saying what agents may do, and you grant access through a normal login that you can revoke anytime. Shopify turned it on for every store, and Meta's Muse checks for it first. It's early, though. The spec is still marked draft and may change in breaking ways, and most partners are still integrating. Right direction, but this is just the start.
3
YC 秋季批的 Taan.ai 出了个音乐模型 Taan v1-preview,主打给视频配乐。 做短视频和 vlog 的人可以直接上传成片,它会识别剪辑点和情绪,一次生成贴合节奏的原创配乐。还能用大白话指挥,比如前8秒紧张,停2秒再转明亮。想改的话可以只让小提琴开心一点、提前3秒进,其他乐器不动,不用整首重来。官方说训练数据全是授权音乐,生成的曲子可商用,Creator 每月15美元约25分钟。 限制也明显,目前主要是管弦乐和电影感配乐,准确率是官方自测,还没有第三方评测。做剧情类或氛围向视频的人值得试试,流行和电子曲风还得等。 YC startup Taan.ai just shipped Taan v1-preview, a music model built for scoring video. Upload a cut and it picks up the edit points and mood, then writes an original score that fits in one pass. You can direct it in plain words, like tense for 8 seconds, pause 2, then go bright. You can also tweak one part, say make the violin happier and bring it in 3 seconds earlier, while everything else stays put. They say it was trained only on licensed music and the output is cleared for commercial use. Creator plan is $15 a month for about 25 minutes. Right now it leans orchestral and cinematic, and the accuracy numbers are self-reported with no independent tests yet. Worth a try for narrative or mood videos. Pop and electronic will have to wait.
1
5
李飞飞的 World Labs 11 月 12 日把世界模型 Atlas 开放给所有人了,之前想生成自己的场景得先拿测试资格。 普通人能拿它做的事挺实在。传一张旅行照片或房间照片,就能变成一个可以转镜头逛的 3D 场景,最多 64 张照片还能重建一个真实地点。设好镜头路径可以导出视频片段,做短视频运镜或者装修布景草图都用得上。做游戏和 3D 的人还能把结果转成网格拿进 Blender。 限制也明显。权重不开放,只能用官方应用和 API,官方没有单独的定价页,文档还挂着 beta。第三方报道说一次精细生成要十几到二十几分钟,现在看到的效果也主要是官方和早期用户的演示。 我的看法是这东西值得拿自己的照片试一把,但先别当成熟工具指望它。 Fei-Fei Li's World Labs opened its world model Atlas to everyone on November 12. Before that you needed test access to generate your own scenes. What regular people can do with it is pretty concrete. Upload one travel photo or a shot of your room and it becomes a 3D scene you can move the camera through, and up to 64 photos can reconstruct a real place. Set a camera path and export a video clip, handy for short video camera moves or rough renovation and set sketches. Game and 3D folks can also convert the result into a mesh for Blender. The limits are real. Weights stay closed, so it is the official app and API only, there is no dedicated pricing page, and the docs still say beta. Third party reports say a detailed run takes roughly 15 to 25 minutes, and what we have seen so far is mostly official and early user demos. My take is it's worth trying with your own photos, but don't count on it as a mature tool yet.
5
Google 11 月 12 日发了 Gemini 4 Hydrogen。官方把 Carbon 和 Argon 放在长程推理和攻坚上,Hydrogen 则面向日常写代码和跑 agent 工作流,主打快速执行。 普通人最先能碰到的入口是 Search。Google 说全球所有用户都能用上,官方描述是更能跟上一次问好几件事、或者带着错字的检索,结果更偏可视化,会给图和对照表,还会用 query fan-out 去找更广的链接。这些是 Google 自己的产品说法,不是独立评测,实际好不好用还得自己试。Gemini 应用里付费用户可以从模型菜单里选它,免费用户主要靠自动选模型,在合适的提示上被路由到 Hydrogen,别指望免费档一定能手动锁定。 开发者这边,Google AI Studio 全档和 Gemini API 都能用。付费的 Pro 和 Ultra 还能在 Antigravity 里用,Orchestrator 会在同一个会话里自动在 Carbon 和 Hydrogen 之间切换。Google 公布的数字是 Terminal-Bench 上成功率接近 Carbon,平均少用 24% 配额、快 19%。这里有个细节,Antigravity 博客写的是 Terminal-Bench 4.1,主博客有一处写 4.0,两个版本号别混着当同一个结果看。 相对 Gemini 3.8 Flash,官方给的对比是默认 high 思考档完成任务快 2.5 倍,Terminal-Bench 4.1 是 57% 对 18.2%,编造声称减少 5 倍以上。Artificial Analysis Intelligence Index 上默认 high 是 51,比 Argon 的 50 还高一点,最高 xhigh 档到 54。思考深度默认是 high,还可以往下调,或开到更高的 xhigh。上下文 100 万 token,模型卡写输入输出合计最多 1M,默认同步输出最多 64K,后台长解码可以到 1M。价格是全程一口价,100 万 token 上下文下每百万输入 1.50 美元、每百万输出 7.50 美元。 我的看法是,这档真正的意义在于把够快和够聪明压到同一个价位上,日常写代码和跑 agent 的人不用再在便宜快模型和贵模型之间反复切。限制也清楚,那些基准都是 Google 自己公布的,Terminal-Bench 版本号官方自己都没对齐。Search 是全员直接用上,Gemini 应用里免费用户拿到的是自动路由,实际体验和宣传之间还有距离。 Google released Gemini 4 Hydrogen on November 12. Officially, Carbon and Argon stay on long-horizon reasoning and hard problems, while Hydrogen is the fast execution tier for everyday coding and agent workflows. The easiest place for ordinary users to meet it is Search. Google says it is available to all users worldwide, and describes it as better at following a question that asks several things at once, or a query with typos, with more visual results like images and comparison tables, plus query fan-out to surface a wider set of links. Those are Google's own product claims, not an independent review, so how well it actually works is still something you have to try yourself. Paid users can pick it from the model menu in the Gemini app, but free users mostly get it through automatic model selection, which routes suitable prompts to Hydrogen. Do not assume the free tier lets you lock it in manually. For developers, it is available across all tiers in Google AI Studio and through the Gemini API. Paid Pro and Ultra users also get it in Antigravity, where the Orchestrator switches between Carbon and Hydrogen within the same session. Google's published numbers say it comes close to Carbon's success rate on Terminal-Bench while using 24% less quota on average and running 19% faster. One detail worth noting, the Antigravity blog cites Terminal-Bench 4.1 while the main blog has one spot saying 4.0, so do not treat those as the same result. Against Gemini 3.8 Flash, Google's comparison is 2.5x faster at completing tasks on the default high thinking level, 57% versus 18.2% on Terminal-Bench 4.1, and more than 5x fewer fabricated claims. On the Artificial Analysis Intelligence Index it scores 51 at default high, slightly above Argon's 50, and reaches 54 at the top xhigh level. Thinking depth defaults to high, and you can turn it down or open the higher xhigh setting. Context is 1M tokens, with the model card saying input and output combined max out at 1M, default synchronous output up to 64K, and background long decoding up to 1M. Pricing is flat across the full context, $1.50 per million input tokens and $7.50 per million output tokens. My take is that the real point is squeezing fast and smart into one price point, so people doing daily coding and agent work stop bouncing between a cheap fast model and an expensive one. The limits are just as clear. Those benchmarks are Google's own, and the Terminal-Bench version numbers are not even aligned in Google's own posts. Search is on for everyone, while free users in the Gemini app get automatic routing rather than a manual choice, so there is still a gap between the pitch and what you actually get.
1
7