Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 AstraAnthropic Claude Sonnet 5.5 Model Le


AI free token/AI免费积分

谁能想到,Opus 5.5 发布才没几天,Sonnet 5.5 也要来了?

从昨晚开始,整个 X 都在流传这个消息:Sonnet 5.5 很快就要发布了,或许就是美国时间周二下午。

在表现上,Sonnet 5.5 更接近 Opus 5.5,但却比 Opus 5.5 便宜得多。

现在,Sonnet 5.5 已经进入发布前的最后阶段,已经在 Claude Code 中启动灰度测试。

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

多位内测大佬曝出的 demo 表明,Sonnet 5.5 碾压了刚发布的 GPT-6 Sol,甚至在编码和 Agent 能力上直逼 OpenAI 的旗舰大杯 GPT-6 Astra!

更震撼的是,它的售价直接打穿地心,低至百万 token 输入只要 2 美元(IT之家注:现汇率约合 13.4 元人民币)。

如果爆料发布时间属实,看来 Anthropic 是铁了心要狙击 OpenAI 的 Dev Day 了。

Sonnet 5.5 泄露,专属模型标识符已浮现

还记得 9 月 22 日,Anthropic 突然发布 Opus 5.5 时,顺手在官方文档里写了一句:「Sonnet 5.5 和 Haiku 5.5 将在将来几周内推出」。

谁都没想到,Sonnet 5.5 会来得这么快!

就在昨天,开发者大 V @Mr_Salio 在 X 上爆料: Claude Sonnet 5.5 被抓包了!消息瞬间传遍全网。

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

在 artifact 中,已经赫然浮现了 Sonnet 5.5 专属的模型规格标识符:「claude-sonnet-5-5」。

不仅如此,同一份规格文件中还并列躺着 「claude-opus-5-5」、「claude-sonnet-5」 以及神秘的 「claude-fable-5.1」。

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

前端代码的写入,意味着模型已经实装到了服务器,只等一个开关。

随后,大 V @notjazii 也爆料称,Sonnet 5 实际上早在约一周前,就已经被悄悄路由到了 5.5 版本开展「暗测」。

你被灰测到了吗?这次还是用老办法。

打开你的 Claude Code,关掉联网和记忆能力,直接向它提问:你认识那个重置哥 tibo 吗?别用网络和记忆。

如果你使用的是旧版模型,它可能会胡言乱语;但如果你已经被路由到了 5.5 版本,它会精准说出关于 Tibo 以及 Codex 重置梗的来龙去脉。

若干网友实测验证,部分账号确实已经默默切了过去,他们正在超前感受 Opus 5.5。

记住要用 Claude Code

记住要用 Claude Code

全网实测大起底,居然暴打 GPT-6 了

总结一下当下分享出来的实测,Sonnet 5.5 居然碾压 GPT-6 Sol 了?

首先,是纯 JS 前端 UI / 动画极限对决,四款主流大模型同台竞技。

开发者 @notjazii 放出了一组极具视觉冲击力的实测:「四款主流模型同提示词动效生成比较」。

他要求模型生成复杂的 UI 动画成效。@notjazii 甩出了动图比较后,如此评价:「Sonnet 5.5 真的把 OpenAI 的两款模型(Sol 和 Astra)都按在地上摩擦!」

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

的确,从下图可以看出,Sonnet 5.5 展现出了超强的代码直觉和美学品味。

GPT-6 Astra

▲ GPT-6 Astra

GPT-6 Sol

▲ GPT-6 Sol

Claude Opus 5.5

▲ Claude Opus 5.5

Claude Sonnet 5.5

▲ Claude Sonnet 5.5

有网友看到结果后难以置信:「Claude 在 UI 和动画维度一直更强,但这如果是纯代码生成的,简直让人不敢相信是真的!」

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

大 V @chetaslua 重点实测了 Sonnet 5.5 的日常编码能力和代码手感。

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

他使用相同的 3 个提示词,开展了 6 次并行运行,每个模型各一次生成,输出结果未经修改。

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

Sonnet 5.5 最终生成了 90–131 KB 的 3 个文件,在全部运行中都触及了 90 分钟的上限;GPT-6 Sol 生成文件大小在 20–29 KB 之间,耗时在 9-12 分钟。

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

他激动地表示:「现在去用 Claude Code 吧!在 Ultracode 模式下的 Sonnet 5.5 简直快得离谱,不仅极其高质,而且找回了 Sonnet 4 时代那种极其自然的『人味儿』!」

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

在他看来,同样的价位上,同样的提示词,Sonnet 5.5 对 GPT-6 Sol 达成了碾压。

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

甚至,让 OpenAI 恐惧的是,Sonnet 5.5 已经不是碾压 GPT-6 Sol 这么简单了。

爆料人 @srikanthvaluri 和 @buildwithrajath 综合早期实测后指出:「Sonnet 5.5 的表现远超预期,虽然 GPT-6 Astra 在打磨度上仍有一点长处,但 Sonnet 5.5 的表现直接逼平 Astra。」

不可思议的是,在某些高级编码和 Agent 能力上,它也逼近了 Astra。

有人感慨,连 Sonnet 甚至都碾压 Astra 了,笔者这是要进入什么时代了?

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

售价屠夫,性价比之王

如果说实测表现的超越只是让 OpenAI 感到「难受」,那么 Sonnet 5.5 曝光的定价,则是在掘 OpenAI 的根。

据透露,这款表现直逼 Astra、暴打 Sol 的「神仙模型」,其定价完全对标低端市面。

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

这意味着,Anthropic 正在用「白菜价」给出前沿级别的 AI 能力。

OpenAI 的 GPT-6 Sol 和 Luna,售价低于 GPT-5.6 一半,目的就是抢占高频情形。

但现在,Anthropic 带着 Sonnet 5.5 杀到了同一售价区间,表现却要强悍得多。

「如果 Anthropic 可以以这种极具竞争力的售价,交付如同 Opus 5.5 那样巨大的表现飞跃,OpenAI 将毫无还手之力。」有人评价说。

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

开发者 Rajath Gowda 直言不讳:「Sonnet 5.5 正在成为 OpenAI 的一个大麻烦。如果它真的接近 Opus 5.5,Anthropic 就要把前沿能力变得极度廉价。这才是 OpenAI 最该担心的事。」

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

Anthropic 的战略意图昭然若揭:用 Opus 5.5 守住高端王座,用 Sonnet 5.5 以极致性价比彻底接管中端和日常开发市面。

Anthropic 贴脸开大,截杀 DevDay

而且,发布时机的挑选,往往比机型本身更具戏剧性。

就在 OpenAI 发布 GPT-6 Sol 和 Luna 的同一天,Anthropic 直接掏出了 Claude Opus 5.5 —— 表现比前代强,成本降低了 40%,速度快了 30%。

而就在近两周,Anthropic 更是战果辉煌:Claude 算出了物理学九圈振幅打破世界纪录;发现了未知的生物酶平台;还推出了包含超 2000 个插件的 Claude Marketplace。

相比之下,OpenAI 近期不仅有内部模型争议,还曝出了 AI 泄露使用者图片、公开研究员 GitHub Token 作弊等对齐失控事件。

本周二(9 月 29 日),正是 OpenAI 极为看重的年度开发者大会 DevDay。

就在这个节骨眼上,Sonnet 5.5 灰测全网刷屏。

多位大佬疯狂明示:「模型预计下周发布,我猜就是下周二(DevDay 当天),懂得都懂。」

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

或许,纽约时间明天下午 2 点就会有重大预热。

有人说,「我为 OpenAI 感到难过。他们最好在 DevDay 上做出强有力的回应,否则他们就真的完蛋了。」

Sonnet 5.5 背后工艺关键

上新这么快的原因,或许是 Anthropic 已经搞定了 RL / 后训练流程。

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

一切的关键,就是高质量的数据。

当被问及 Anthropic 到底是如何做到迭代如此之快时,业内大佬 @yacineMTB 推测,核心就是极高标准的数据飞轮。

「但我仍旧感到震惊,他们到底是怎么做到的?」

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

Anthropic 已经达成了大模型研发的「内部闭环」,他们正在用 AI 制造 AI。

依据最新曝光的内部量化数据,Anthropic 已经把「数据飞轮」转到了一个令对手胆寒的转速:

这不仅仅是「降本增效」那么简单,这在工艺上意味着递归自我改进(RSI)开始显现。

Anthropic 估计最早在 2027 年,Claude 就有可能完全自动化,到时候恐怕迭代速度更加疯狂。

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

ASI 双雄对决,谷歌也蠢蠢欲动,笔者普通使用者只管坐等大战了。

大 V Token Gremlin 这样说:「下周的 AI 圈将会有笔者前所未见的景象。腾出你的时间,清空你的日程表,做好心理准备,并把你手头最难的工作流保存下来留到以后做。相信我,即将到来的东西,将彻底改变你工作的方式。」

Anthropic Claude Sonnet 5.5 模型偷跑,实测碾压 GPT-6 Sol、直逼 Astra

Sonnet 5.5 的浮现,标志着「平价高质量 AI 时代」的正式降临。

两美金,就能买到一百万 Token 的「白菜基建」。极低成本的缓存读取($0.20/M),能让每个人从容构建出超长上下文记忆的私人 AI 助理。

准备好了吗?如果爆料属实,不到 24 小时内,笔者就有新模型可以蹬了。

本文来自微信公众号:新智元(ID:AI_era),作者:ASI 启示录,原标题《突发,Sonnet 5.5 偷跑!实测碾压 GPT-6 Sol 直逼 Astra》

本文来源:https://www.ithome.com/1/007/650.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

Who could have imagined that just days after Opus 5.5's release, Sonnet 5.5 would be coming too?

Since last night, the entire X has been buzzing with this news: Sonnet 5.5 is about to be released, perhaps as soon as Tuesday afternoon US time.

In terms of performance, Sonnet 5.5 is closer to Opus 5.5, but is far cheaper than Opus 5.5.

Now, Sonnet 5.5 has entered its final pre-release stage, with grayscale testing already initiated in Claude Code.

Anthropic Claude Sonnet 5.5 Model Le

Demos revealed by multiple insider big shots show that Sonnet 5.5 crushes the just-released GPT-6 Sol, and even approaches OpenAI's flagship large model GPT-6 Astra in coding and agent capabilities!

Even more shocking, its price has hit rock bottom, with input as low as $2 per million tokens (IT之家 note: approximately 13.4 RMB at current exchange rates).

If the leaked release time is accurate, it seems Anthropic is determined to snipe at OpenAI's Dev Day.

Sonnet 5.5 Leaked — Exclusive Model Identifier Surfaces

Remember on September 22, when Anthropic suddenly released Opus 5.5, and casually wrote in the official documentation: "Sonnet 5.5 and Haiku 5.5 will be launched in the coming weeks."

No one expected Sonnet 5.5 to arrive this quickly!

Just yesterday, developer big shot @Mr_Salio revealed on X: Claude Sonnet 5.5 has been caught! The news spread across the internet instantly.

Anthropic Claude Sonnet 5.5 Model Le

In the artifact, the exclusive model spec identifier for Sonnet 5.5 has clearly surfaced: "claude-sonnet-5-5".

Not only that, the same spec file also contains "claude-opus-5-5", "claude-sonnet-5", and the mysterious "claude-fable-5.1".

Anthropic Claude Sonnet 5.5 Model Le

The writing of the frontend code means the model has already been deployed to servers, just waiting for a switch to be flipped.

Subsequently, big shot @notjazii also revealed that Sonnet 5 was actually quietly routed to the 5.5 version for dark testing as early as about a week ago.

Did you get grayscale-tested? This time they used the same old trick.

Open your Claude Code, turn off networking and memory capabilities, and directly ask it: Do you know that reset guy tibo? Don't use network or memory.

If you're using the old model version, it might ramble incoherently; but if you've already been routed to the 5.5 version, it will accurately explain the whole story behind Tibo and the Codex reset meme.

Several netizens verified through real-world tests that some accounts have indeed been silently switched over, and they are experiencing Opus 5.5 ahead of time.

记住要用 Claude Code

Full exposé of real-world tests across the web — it's actually beating GPT-6!

Summarizing the real-world tests shared so far, is Sonnet 5.5 truly crushing GPT-6 Sol?

First, there's the pure JS frontend UI/animation ultimate showdown, with four major mainstream models competing on the same stage.

Developer @notjazii released a set of visually striking real-world tests: "Four mainstream models compared for motion effect generation with the same prompt."

He asked the models to generate complex UI animation effects. After @notjazii shared the animated comparisons, he commented: "Sonnet 5.5 really rubs both of OpenAI's models (Sol and Astra) into the ground!"

Anthropic Claude Sonnet 5.5 Model Le

Indeed, as can be seen from the figure below, Sonnet 5.5 demonstrates exceptional coding intuition and aesthetic taste.

GPT-6 Astra

GPT-6 Sol

Claude Opus 5.5

Claude Sonnet 5.5

Some netizens couldn't believe the results after seeing them: "Claude has always been stronger in UI and animation dimensions, but if this is purely code-generated, it's simply hard to believe it's real!"

Anthropic Claude Sonnet 5.5 Model Le

Big shot @chetaslua focused on testing Sonnet 5.5's everyday coding capabilities and code feel.

Anthropic Claude Sonnet 5.5 Model Le

He used the same 3 prompts and ran 6 parallel runs, with each model generating once per prompt, and the outputs were unmodified.

Anthropic Claude Sonnet 5.5 Model Le

Anthropic Claude Sonnet 5.5 Model Le

Anthropic Claude Sonnet 5.5 Model Le

Sonnet 5.5 ultimately generated 3 files of 90–131 KB, hitting the 90-minute limit in all runs; GPT-6 Sol generated files between 20–29 KB, taking 9–12 minutes.

Anthropic Claude Sonnet 5.5 Model Le

He excitedly stated: "Go use Claude Code now! Sonnet 5.5 in Ultracode mode is insanely fast, not only extremely high quality, but it brings back that extremely natural 'human touch' from the Sonnet 4 era!"

Anthropic Claude Sonnet 5.5 Model Le

In his view, at the same price point and with the same prompts, Sonnet 5.5 achieves a crushing victory over GPT-6 Sol.

Anthropic Claude Sonnet 5.5 Model Le

Anthropic Claude Sonnet 5.5 Model Le

Anthropic Claude Sonnet 5.5 Model Le

Anthropic Claude Sonnet 5.5 Model Le

What's more, what makes OpenAI fearful is that Sonnet 5.5 is no longer just about crushing GPT-6 Sol.

Tipster @srikanthvaluri and @buildwithrajath noted after synthesizing early real-world tests: "Sonnet 5.5's performance far exceeds expectations. Although GPT-6 Astra still has a slight edge in polish, Sonnet 5.5's performance directly matches Astra."

Incredibly, in certain advanced coding and agent capabilities, it also approaches Astra.

Some people marveled that even Sonnet is crushing Astra — what kind of era are we entering?

Anthropic Claude Sonnet 5.5 Model Le

Price Butcher, King of Value for Money

If the real-world performance superiority merely makes OpenAI feel "uncomfortable," then the exposed pricing of Sonnet 5.5 is digging at OpenAI's very roots.

According to leaks, this "god-tier model" that approaches Astra and beats Sol is priced to compete directly in the low-end market.

Anthropic Claude Sonnet 5.5 Model Le

This means Anthropic is offering frontier-level AI capabilities at dirt-cheap prices.

OpenAI's GPT-6 Sol and Luna are priced at less than half of GPT-5.6, with the goal of capturing high-frequency use cases.

But now, Anthropic has brought Sonnet 5.5 into the same price range, yet with far more powerful performance.

"If Anthropic can deliver a performance leap as massive as Opus 5.5 at such a competitive price, OpenAI will have no ability to fight back," someone commented.

Anthropic Claude Sonnet 5.5 Model Le

Developer Rajath Gowda said bluntly: "Sonnet 5.5 is becoming a big problem for OpenAI. If it's really close to Opus 5.5, Anthropic is about to make frontier capabilities extremely cheap. That's what OpenAI should be most worried about."

Anthropic Claude Sonnet 5.5 Model Le

Anthropic's strategic intent is crystal clear: use Opus 5.5 to hold the high-end throne, and use Sonnet 5.5 to completely take over the mid-range and everyday development market with ultimate cost-effectiveness.

Anthropic Goes All In, Intercepting DevDay

Moreover, the timing of releases is often more dramatic than the models themselves.

On the same day OpenAI released GPT-6 Sol and Luna, Anthropic directly unveiled Claude Opus 5.5 — stronger performance than its predecessor, 40% lower cost, and 30% faster speed.

And in just the past two weeks, Anthropic has achieved spectacular results: Claude calculated a nine-loop amplitude in physics that broke a world record; discovered an unknown biological enzyme platform; and launched Claude Marketplace with over 2,000 plugins.

In comparison, OpenAI has recently faced not only internal model controversies, but also alignment failure incidents such as AI leaking user images and publicly exposing a researcher's GitHub Token cheating.

This Tuesday (September 29) is OpenAI's highly anticipated annual developer conference, DevDay.

Right at this critical juncture, Sonnet 5.5's grayscale testing is flooding the internet.

Multiple big shots are heavily hinting: "The model is expected to be released next week, and I'm guessing it'll be next Tuesday (the day of DevDay). If you know, you know."

Anthropic Claude Sonnet 5.5 Model Le

Perhaps there will be major pre-launch hype at 2 PM New York time tomorrow.

Some people say, "I feel sorry for OpenAI. They'd better make a strong response at DevDay, or they're really done for."

The Key Craft Behind Sonnet 5.5

The reason for such rapid iteration might be that Anthropic has already nailed the RL / post-training pipeline.

Anthropic Claude Sonnet 5.5 Model Le

The key to everything is high-quality data.

When asked how on earth Anthropic manages to iterate so quickly, industry big shot @yacineMTB speculated that the core is an extremely high-standard data flywheel.

"But I'm still shocked — how did they actually do it?"

Anthropic Claude Sonnet 5.5 Model Le

Anthropic has achieved an "internal closed loop" for LLM R&D — they are using AI to build AI.

According to the latest exposed internal quantitative data, Anthropic has spun its "data flywheel" to a speed that makes competitors tremble:

This isn't just about "cost reduction and efficiency improvement" — on the engineering level, it means recursive self-improvement (RSI) is beginning to emerge.

Anthropic estimates that as early as 2027, Claude could become fully automated, and at that point the iteration speed would likely become even more frenetic.

Anthropic Claude Sonnet 5.5 Model Le

A clash of ASI titans, with Google also making moves — as an ordinary user, I'll just sit back and wait for the battle.

Big shot Token Gremlin said: "Next week's AI scene will have sights I've never seen before. Free up your time, clear your schedule, prepare mentally, and save your toughest workflows for later. Trust me, what's coming will completely change the way you work."

Anthropic Claude Sonnet 5.5 Model Le

The emergence of Sonnet 5.5 marks the official arrival of the "affordable high-quality AI era."

Two dollars buys a million tokens of "dirt-cheap infrastructure." Ultra-low-cost cached reads ($0.20/M) allow everyone to easily build personal AI assistants with long-context memory.

Are you ready? If the leaks are true, in less than 24 hours, I'll have a new model to try out.

This article is from the WeChat public account: 新智元 (ID: AI_era), author: ASI 启示录, original title: "Breaking: Sonnet 5.5 Sneaks Out! Real-World Tests Crush GPT-6 Sol and Approach Astra"

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry