$2。
这是Claude Sonnet 5每百万输入token的定价。输出token $10/百万（8月31日前促销价），之后恢复到$3/$15。作为对比，Opus 4.8的价格是输入$5、输出$25——Sonnet 5的输入成本是Opus的40%，输出成本是Opus的40%-60%。
价格不是重点，重点是它能做什么。Agent编码基准63.2%，逼近Opus 4.8的69.2%；知识工作基准上，Sonnet 5甚至微幅反超Opus 4.8。Zapier高级工程师Daniel Shepard的测试反馈是：“我们交给它一个两部分任务——更新Salesforce客户等级、给企业联系人发上线公告——它端到端完成了。以前这种任务会中途卡住。日常自动化？无脑选它。”
6月30日，Anthropic发布Claude Sonnet 5。同一天，美国商务部正式解除对Mythos 5和Fable 5的出口管制，全球用户7月1日起恢复访问。这两件事撞在同一天不是巧合——它们共同指向一个判断：AI Agent的竞争，正式从“谁更聪明”转向“谁更便宜、更可靠、更可规模化部署”。
Sonnet 5到底升级了什么
先看硬数据。
相比上一代Sonnet 4.6（今年2月发布），Sonnet 5在Agent能力上的提升是跨代的：Agent编码从58.1%跃升到63.2%，工具使用、推理、知识工作全面进步。更重要的是，它学会了“自我检查”——Anthropic在发布中提到，测试者反馈Sonnet 5会在没有被明确要求的情况下，主动检查自己的输出是否正确。Rust工程师Neel Chotai分享了一个案例：“我让Sonnet 5调查一个bug。它主动写了复现测试、实现了修复，然后stash掉修复代码确认bug确实会复现。一口气做完。”
这种“自主完成多步骤任务且不自欺欺人”的能力，此前只有Opus级别的模型才能稳定做到。而Opus 4.8的价格是Sonnet 5促销价的2.5倍。
安全层面也有进步。Sonnet 5在恶意请求拒绝、prompt注入攻击防御、幻觉率、阿谀奉承率上全面优于Sonnet 4.6。但Anthropic也明确承认：在网络安全能力上，Sonnet 5远不如Opus和Mythos——它在Firefox漏洞利用测试中完全无法生成有效exploit（成功率0%），部分成功率也仅比Sonnet 4.6略高。这也是为什么Anthropic给Sonnet 5默认启用了网络安全防护措施，但防护级别低于Mythos/Fable。
一个值得注意的细节：Sonnet 5使用了更新的tokenizer，相同文本可能被切分成更多token。这意味着实际成本可能比表面数字略高——但即便考虑tokenizer变化，Sonnet 5的性价比仍然是当前市场上最具竞争力的。
"以前需要Opus才能完成的多步骤Agent任务，现在Sonnet 5以40%的成本就能做到。这不是降价，这是Agent能力的民主化。"
—— 一位AI基础设施创业者的判断
为什么平价Agent是一个产业拐点
理解Sonnet 5的意义，需要把它放在更大的产业背景下。
2026年上半年，AI行业有一个共识在形成：大模型的“聊天”功能已经商品化，真正的价值在Agent——能自主规划、调用工具、完成多步骤任务的AI系统。OpenAI的GPT-5.6 Sol主打多Agent协作，Google的Gemini 3.5 Flash主打“从聊天机器人到Agent工具”的转型，三家前沿公司的竞争焦点完全一致。
但Agent有一个致命问题：贵。一个自主完成复杂任务的Agent，可能消耗数十万甚至上百万token。在Opus定价下（$5/$25），一个重度Agent任务可能花费几美元到几十美元。对个人开发者来说这是尝鲜成本，对企业来说这是规模化部署的拦路虎。当你想让Agent处理一万个客户工单、自动审查一千个PR、批量更新CRM数据时，token成本就是一个必须精算的账本。
Sonnet 5的$2/$10定价，把这个账本改写了。假设一个Agent任务平均消耗50万输入token和10万输出token，用Opus 4.8的成本是$2.5+$2.5=$5；用Sonnet 5促销价的成本是$1+$1=$2；用9月后的标准价是$1.5+$1.5=$3。成本直接腰斩。对于每天运行数千个Agent任务的企业来说，这意味着每月节省数万美元的API账单。
这不仅仅是一个模型的发布，这是Agent经济模型的一次重置。
同一天发生的另一件事：基础设施在跟上
Sonnet 5不是孤例。6月30日到7月1日这48小时内，Agent生态的多个环节同时传来信号，指向同一个方向：Agent正在从“demo”走向“基础设施”。
X推出官方MCP服务器。马斯克旗下的社交平台正式上线Hosted MCP Server，让Claude、Cursor、Grok Build等MCP兼容应用可以直接通过用户授权访问X平台数据。这不是什么新功能——X的API早就可以搜索、读帖、查用户——但官方MCP服务器把集成成本降到了零。开发者不需要自己搭建、维护MCP服务器，不需要处理认证，开箱即用。X加入了GitHub、Slack、Notion、Stripe、Salesforce的行列，MCP正在从一个开源协议变成AI工具互联的事实标准。值得注意的是，X明确表示MCP不支持Write API——你不能用它自动发帖，但可以读取和分析。这是一个谨慎但明确的开放信号。
Amazon投入10亿美元成立Forward Deployed Engineer（FDE）组织。AWS将派遣工程师嵌入企业客户，帮助部署定制化Agent系统。这不是新鲜事——Palantir首创的FDE模型，OpenAI和Anthropic此前分别成立了40亿美元和15亿美元的FDE合资公司——但Amazon的入场意味着云巨头正式把Agent部署当成了核心服务，而不仅仅是AI实验室的增值服务。AWS VP Francessca Vasquez的表述很明确：“客户离开FDE部署时，不仅获得新系统，还获得持久的AI技能和工作流。”换句话说，Amazon在教企业怎么自己跑Agent。
Etched拿下10亿美元订单，估值50亿美元。这家2022年成立的AI芯片公司（两位哈佛辍学的Thiel奖金得主创办），已经获得包括Andrej Karpathy、Geoffrey Hinton、李飞飞、Peter Thiel在内的投资人背书，累计融资8亿美元。他们专门做AI推理ASIC芯片——不是训练，是推理。当Agent大规模部署，推理成本是最大的支出项。Etched的Sohu芯片主打“前沿推理集群”，承诺比GPU更快、更便宜、更省电。10亿美元合同在手，说明推理芯片的市场需求已经从“预期”变成了“采购订单”。
Agent平价时代的竞争格局
Sonnet 5发布后，Agent模型市场的格局变得清晰了。
高端市场（Opus 4.8 / GPT-5.5 / Mythos / Gemini Pro）：最强能力，最高价格，适合最复杂的任务。但Mythos的出口管制风波证明，最强模型往往伴随最强监管。Opus 4.8定价$5/$25，主要面向对准确性要求极高、对成本不敏感的场景（网络安全、深度研究、复杂代码审查）。
中端市场（Sonnet 5 / GPT-5.6 Sol / Gemini 3.5 Flash）：这是Agent部署的主战场。Sonnet 5促销价$2/$10，Gemini 3.5 Flash更便宜（Google一直用低价策略争夺市场份额），GPT-5.6 Sol目前还在受控预览阶段。这个价位带的模型，能力足以完成80%以上的Agent任务，成本只有高端模型的几分之一。谁在中端市场建立成本-可靠性的最优平衡，谁就能赢得企业Agent部署的最大蛋糕。
应用层垂直整合趋势加速。就在Sonnet 5发布的前一天，Wix旗下的vibe coding平台Base44宣布推出自研模型Base1。这家一年前被Wix以8000万美元收购时只有8个人的公司，如今ARR突破1亿美元，决定不依赖外部前沿模型，用“平台上数千万真实用户交互数据”训练自己的垂直模型。Lovable（ARR 5亿美元，同样是vibe coding赛道）仍然依赖外部LLM，但Base44的选择代表了一个趋势：当应用层公司达到一定规模，自研模型或微调模型的ROI开始转正。Headline VC合伙人Jonathan Userovici的观察是：“企业客户开始问——我们真的需要在所有场景用最新模型吗？答案越来越多是‘不需要’。”
这对Anthropic和OpenAI意味着什么？意味着中端模型的“量”比高端模型的“价”更重要。当越来越多的应用场景跑在Sonnet 5而不是Opus 4.8上，Anthropic的收入结构会发生变化——从少数客户付高价，变成大量客户付中价。这是一个更健康、更可防御的商业模式。
"模型在进步，但它们将保持通用性。垂直专精才是应用层的护城河。"
—— Base44创始人 Maor Shlomo
写在Agent平价时代的门槛上
站在2026年7月1日这个时间点回望，AI Agent的发展速度比大多数人预期的要快。
一年前，能稳定自主完成多步骤任务的模型还屈指可数；半年前，Agent还是“高端功能”的代名词；今天，Sonnet 5把Agent能力打到了$2/MTok，MCP生态在加速扩张，Amazon在帮企业部署Agent，X在开放Agent接入，ASIC推理芯片公司在拿10亿美元订单。Agent不再是“未来会怎样”的概念，而是“现在怎么用”的工程问题。
但也需要冷静。Sonnet 5的自我检查能力进步了，但Agent的可靠性问题没有根本解决——它仍然会在长任务中偏离方向，仍然会在遇到模糊指令时做出错误假设，仍然无法完全替代人类的判断。Lovable联合创始人Fabian Hedin说的那句话值得记住：“一个知道什么时候该说不的模型，和一个知道怎么构建的模型同样重要。”
成本的门槛被击穿了，但信任的门槛还在。
不过这一次，AI行业有理由比过去任何时候都更乐观。因为当一个东西足够便宜，experimentation的成本就趋近于零——无数开发者和企业会去试、去错、去找到真正创造价值的Agent应用场景。下一个AI杀手级应用，很可能不是某个大公司发布的宏大产品，而是一个小团队用Sonnet 5加MCP、在一个月内跑通的垂直场景。
Agent平价时代，真正的竞争才刚刚开始。
明天见。
$2.
That's Claude Sonnet 5's price per million input tokens. Output tokens are $10/million (promo pricing through August 31), reverting to $3/$15 after. For comparison, Opus 4.8 is $5 input, $25 output -- Sonnet 5's input cost is 40% of Opus, and output cost is 40%-60% of Opus.
Price isn't the point; capability is. Agent coding benchmark: 63.2%, closing in on Opus 4.8's 69.2%; on knowledge work benchmarks, Sonnet 5 actually edged out Opus 4.8 slightly. Zapier senior engineer Daniel Shepard's test feedback: "We gave it a two-part task -- update Salesforce customer tiers, send launch announcements to enterprise contacts -- and it completed it end-to-end. Previously tasks like that would get stuck halfway. Daily automation? No-brainer pick."
On June 30, Anthropic released Claude Sonnet 5. The same day, the US Department of Commerce formally lifted export controls on Mythos 5 and Fable 5, restoring global access starting July 1. These two events landing on the same day is no coincidence -- they both point to one conclusion: competition in AI Agents has officially shifted from "who's smarter" to "who's cheaper, more reliable, and more deployable at scale."
What Sonnet 5 Actually Upgraded
First, the hard numbers.
Compared to the previous generation Sonnet 4.6 (released this February), Sonnet 5's Agent capability gains are generational: Agent coding jumped from 58.1% to 63.2%, with broad improvements across tool use, reasoning, and knowledge work. More importantly, it learned to "self-check" -- Anthropic noted in the release that testers reported Sonnet 5 would proactively verify whether its own output was correct without being explicitly asked to. Rust engineer Neel Chotai shared a case: "I asked Sonnet 5 to investigate a bug. It proactively wrote a reproduction test, implemented the fix, then stashed the fix code to confirm the bug actually reproduced. Did it all in one go."
This ability to "autonomously complete multi-step tasks without self-deception" was previously only achievable at Opus-tier models with consistency. And Opus 4.8 costs 2.5x Sonnet 5's promo price.
Safety also improved. Sonnet 5 outperforms Sonnet 4.6 across the board on malicious request refusal, prompt injection defense, hallucination rates, and sycophancy rates. But Anthropic openly acknowledged: on cybersecurity capabilities, Sonnet 5 falls far short of Opus and Mythos -- it completely failed to generate valid exploits in Firefox vulnerability testing (0% success rate), with partial success rates only marginally above Sonnet 4.6. That's why Anthropic enabled cybersecurity safeguards by default on Sonnet 5, though at a lower protection level than Mythos/Fable.
A notable detail: Sonnet 5 uses an updated tokenizer, and the same text may be split into more tokens. This means actual costs may be slightly higher than the headline numbers suggest -- but even accounting for tokenizer changes, Sonnet 5's price-performance remains the most competitive on the market.
"Multi-step Agent tasks that previously required Opus can now be done by Sonnet 5 at 40% of the cost. This isn't a price cut; it's the democratization of Agent capability."
-- An AI infrastructure founder's assessment
Why Affordable Agents Are an Industry Inflection Point
To understand Sonnet 5's significance, you need to place it in the broader industry context.
In the first half of 2026, a consensus formed across the AI industry: LLMs' "chat" functionality has been commoditized; the real value is in Agents -- AI systems that can autonomously plan, call tools, and complete multi-step tasks. OpenAI's GPT-5.6 Sol focuses on multi-Agent collaboration; Google's Gemini 3.5 Flash pushes the "from chatbot to Agent tool" transition; the three frontier companies' competitive focus is fully aligned.
But Agents have one fatal problem: they're expensive. An Agent autonomously completing complex tasks can consume hundreds of thousands or even millions of tokens. At Opus pricing ($5/$25), a heavy Agent task might cost several to tens of dollars. For individual developers that's a novelty cost; for enterprises it's the roadblock to large-scale deployment. When you want Agents handling ten thousand customer tickets, auto-reviewing a thousand PRs, or batch-updating CRM data, token costs become a ledger that must be meticulously calculated.
Sonnet 5's $2/$10 pricing rewrites that ledger. Assuming an Agent task consumes an average of 500K input tokens and 100K output tokens, the cost on Opus 4.8 is $2.5+$2.5=$5; on Sonnet 5 promo pricing it's $1+$1=$2; on post-September standard pricing it's $1.5+$1.5=$3. Costs are cut in half. For enterprises running thousands of Agent tasks daily, that means tens of thousands of dollars in monthly API bill savings.
This isn't just a model release; it's a reset of the Agent economic model.
The Other Thing Happening That Day: Infrastructure Is Catching Up
Sonnet 5 isn't alone. In the 48 hours from June 30 to July 1, multiple parts of the Agent ecosystem sent signals simultaneously, all pointing in the same direction: Agents are moving from "demo" to "infrastructure."
X launches official MCP server. Musk's social platform officially launched its Hosted MCP Server, letting MCP-compatible apps like Claude, Cursor, and Grok Build access X platform data directly through user authorization. This isn't a new feature -- X's API has long been able to search, read posts, and look up users -- but the official MCP server reduces integration costs to zero. Developers don't need to build or maintain their own MCP servers, don't need to handle authentication; it just works out of the box. X joins the ranks of GitHub, Slack, Notion, Stripe, and Salesforce; MCP is evolving from an open-source protocol into the de facto standard for AI tool interconnection. Notably, X explicitly stated that MCP doesn't support the Write API -- you can't auto-post with it, but you can read and analyze. A cautious but clear signal of openness.
Amazon commits $1 billion to launch a Forward Deployed Engineer (FDE) organization. AWS will embed engineers within enterprise customers to help deploy customized Agent systems. This isn't new -- Palantir pioneered the FDE model, and OpenAI and Anthropic previously established $4 billion and $1.5 billion FDE joint ventures respectively -- but Amazon's entry means cloud giants are officially treating Agent deployment as a core service, not just an AI lab value-add. AWS VP Francessca Vasquez put it clearly: "When customers leave an FDE deployment, they gain not just a new system but enduring AI skills and workflows." In other words, Amazon is teaching enterprises how to run Agents themselves.
Etched locks in $1 billion in orders at $5 billion valuation. This AI chip company founded in 2022 (by two Harvard dropouts and Thiel Fellowship recipients) has already earned investor backing including Andrej Karpathy, Geoffrey Hinton, Fei-Fei Li, and Peter Thiel, with $800 million in total funding. They build AI inference ASIC chips specifically -- not training, inference. When Agents deploy at scale, inference cost is the single largest expense line. Etched's Sohu chip targets "frontier inference clusters," promising faster, cheaper, and more power-efficient performance than GPUs. $1 billion in contracts in hand means inference chip demand has moved from "anticipation" to "purchase orders."
The Competitive Landscape in the Age of Affordable Agents
After Sonnet 5's release, the Agent model market landscape has come into focus.
High-end market (Opus 4.8 / GPT-5.5 / Mythos / Gemini Pro): Maximum capability, maximum price, suited for the most complex tasks. But the Mythos export control saga proved that the strongest models often come with the strongest regulation. Opus 4.8's $5/$25 pricing primarily targets scenarios where accuracy is paramount and cost sensitivity is low (cybersecurity, deep research, complex code review).
Mid-market (Sonnet 5 / GPT-5.6 Sol / Gemini 3.5 Flash): This is the main battlefield for Agent deployment. Sonnet 5 promo is $2/$10; Gemini 3.5 Flash is even cheaper (Google has consistently used low pricing to grab market share); GPT-5.6 Sol is still in controlled preview. Models at this price tier are capable enough to handle over 80% of Agent tasks at a fraction of high-end model costs. Whoever establishes the optimal cost-reliability balance in the mid-market wins the biggest slice of enterprise Agent deployment.
Application-layer vertical integration accelerates. The day before Sonnet 5's release, Wix's vibe coding platform Base44 announced its self-built model Base1. The company -- which had just 8 people when Wix acquired it for $80 million a year ago -- has now surpassed $100 million ARR and decided not to rely on external frontier models, instead training its own vertical model on "tens of millions of real user interaction data points on the platform." Lovable (at $500 million ARR, also in vibe coding) still depends on external LLMs, but Base44's choice represents a trend: when application-layer companies reach a certain scale, the ROI of building or fine-tuning their own models turns positive. Headline VC partner Jonathan Userovici observed: "Enterprise clients are starting to ask -- do we really need the latest model for every scenario? Increasingly, the answer is 'no.'"
What does this mean for Anthropic and OpenAI? It means the mid-tier model's "volume" matters more than the high-end model's "price." As more application scenarios run on Sonnet 5 rather than Opus 4.8, Anthropic's revenue structure will shift -- from a few customers paying premium prices to a large base of customers paying mid-range prices. That's a healthier, more defensible business model.
"Models will keep improving, but they'll remain general-purpose. Vertical specialization is the application layer's moat."
-- Maor Shlomo, Base44 Founder
Standing at the Threshold of Affordable Agents
Looking back from July 1, 2026, AI Agents have developed faster than most people expected.
A year ago, models that could stably complete multi-step tasks autonomously were few and far between; six months ago, Agents were still synonymous with "premium feature"; today, Sonnet 5 has driven Agent capability down to $2/MTok, the MCP ecosystem is expanding rapidly, Amazon is helping enterprises deploy Agents, X is opening Agent access, and ASIC inference chip companies are landing billion-dollar orders. Agents are no longer a "what will the future hold" concept -- they're a "how do we use this now" engineering problem.
But cool heads are needed. Sonnet 5's self-checking capability has improved, but Agent reliability problems aren't fundamentally solved -- it still drifts off course in long tasks, still makes wrong assumptions when facing ambiguous instructions, and still can't fully replace human judgment. Lovable co-founder Fabian Hedin's words are worth remembering: "A model that knows when to say no is just as important as a model that knows how to build."
The cost barrier has been shattered, but the trust barrier remains.
Still, the AI industry has reason to be more optimistic than ever before. Because when something is cheap enough, the cost of experimentation approaches zero -- countless developers and enterprises will try, fail, and find Agent applications that genuinely create value. The next AI killer app likely won't be some grand product launched by a big company; it'll be a vertical scenario a small team gets running in a month using Sonnet 5 plus MCP.
In the age of affordable Agents, the real competition has only just begun.
See you tomorrow.
Claude Sonnet 5 · Affordable Agents · MCP Ecosystem · Inference Chips · Enterprise FDE
