DeepSeek下一张牌,我猜是19.9元Coding PlanDeepSeek's Next Move: A $19.99 Coding Plan?


AI free token/AI免费积分

本文来源:https://mp.weixin.qq.com/s/lCyEMmSaR_iANXQ6XwJzwA
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

人们好,我是赛博李同学。

最近几天忙着项目和搬家的事情,终于可以坐下来聊聊DeepSeek-V4-Flash-0731。

下面一起聊下我被种草的原因。

先把方向纠正过来

这两天我也翻到了 DeepSeek-V4-Flash-0731 的正式版更新,第一反应是,不是哥们,售价一分钱没动,Agent能力怎么涨成这样了???DeepSeek下一张牌,我猜是19.9元Coding Plan

先把一个很容易搞错的地方讲清楚。

DeepSeek-V4-Flash-0731和之前的Preview版,上下文长度都是1M。现在每百万tokens输入0.14美元,输出0.28美元,缓存命中0.0028美元,这套售价也没有因为0731正式版再次下调。DeepSeek下一张牌,我猜是19.9元Coding Plan

更准确的说法是,加量不加价。

这套售价原本就很能打,尤其是缓存命中价。Coding Agent每走一步,都要带着平台提示、工具定义、项目文件和对话历史继续跑。任务越长,重复浮现的上下文越多。

缓存命中只有普通输入价的五颇为之一,同一套项目背景反复复用,长任务确实更敢一直挂着跑。

但这只是底盘,不是这次更新真正新增的东西。

这次最让我兴奋的,是DeepSeek在原来的上下文和售价上,把Agent能力又往前推了一截。

它开始往工作流里钻

更骚的是,它原生兼容Responses API,还专门适配了Codex。官方文档写得很直白,Codex CLI、ChatGPT桌面端和VS Code里的Codex扩展共用同一份规格,配一次就能用。DeepSeek下一张牌,我猜是19.9元Coding Plan

这才是我觉得最有意思的地方。

DeepSeek现在显然不满足于只做聊天框,它盯上的是整仓库分析、跨文件修改、终端操作、测试修复这一整套 Coding Agent 工作流。

公开社区里已经有人拿0731版本接进Codex CLI,一次生成了可运行的小游戏。成效有人夸,也有人指出动画和达成还很粗糙。坦率的讲,这种个人案例不能证明它已经碾压谁,但至少说明,这玩意不是只会在榜单里好看。

OpenCode也很快把新版DeepSeek V4 Flash接进了付费版和免费版,公开反馈里最集中的一句话就是,比预览版进步显著。DeepSeek下一张牌,我猜是19.9元Coding Plan

真正没发的那张牌

加量到底加在哪儿,官方数据最直观。

0731在Terminal Bench 2.1、DeepSWE、Toolathlon-Verified等测试中都显著超过Flash Preview,部分成绩也超过了V4-Pro-Preview。DeepSeek下一张牌,我猜是19.9元Coding Plan

不过这里有个细节必须讲清楚,Code Agent测试使用了尚未发布的DeepSeek Harness minimal mode,部分榜单还是内部测试集。

成绩里不只有模型的功劳,还有 Harness 的功劳。

这反而让我更好奇了。

模型售价没动,1M 上下文没动,公开成绩却上去了。除了 0731 本身的升级,那套还没发布的 Harness 到底贡献了多少,外面现在还没人能说清楚。

今天网上又浮现了DeepSeek Harness启动内测的消息。且官方已经明确写了Harness minimal mode将很快发布,V4-Pro正式版也会随后到来。

这也是我觉得Harness 更值得盯的地方。模型升级已经发生,售价和上下文都没有变化,真正还没亮相的,是那套决定Coding Agent 怎么跑起来的 Harness。

19.9元从哪儿来的

说到这里,我脑子里才冒出一个很俗的数字,19.9 元。

这个数字不是因为 0731 又降价了。相反,DeepSeek 没动1M上下文,也没动售价,却把更多 Agent 能力塞进了原来的成本里。

所以我期待的是它后面把 Harness、Codex适配和这套便宜的tokens,打包成一个Coding Plan Lite。

19.9 元就不是凭空猜的售价,而是我依据 DeepSeek 现在这几步动作做的一次推测。

然后每周重置3次额度,以后就是”梁神,我爱你“。DeepSeek下一张牌,我猜是19.9元Coding Plan

当然,这只是我的愿望,不是官方消息。

我不知道DeepSeek最后会不会真出19.9 元的Coding Plan。

但这个售价配上那套还没发布的Harness,确实让我开始认真期待了。

试想下如果Harness真的发布,这条路线就完整多了,模型有了,接口有了,负责把它们接进工作流的那一层也有了。

到那时候Coding Agent可能才会从偶尔试试,变成很多人愿意每天开着的工具。

万一呢。。。

Source: https://mp.weixin.qq.com/s/lCyEMmSaR_iANXQ6XwJzwA If there is a copyright issue, please contact us at 13951431913#139.com or submit a comment on the right.

Hey everyone, I'm Cyber Li.

I've been busy with projects and moving these past few days, but I finally have time to chat about DeepSeek-V4-Flash-0731.

Let’s talk about why I’m excited about this.

First, correct the orientation.

I saw the official update for DeepSeek-V4-Flash-0731 these days, and my first thought was, "Wait, the price hasn't changed, but the agent capabilities have skyrocketed??"DeepSeek's Next Move: A $19.99 Coding Plan?

First, clarify a point that is easily misunderstood.

The context length for both DeepSeek-V4-Flash-0731 and the previous preview version is 1M tokens. The current price is $0.14 per million input tokens, $0.28 for output, and $0.0028 for cache hits. The price hasn't been lowered with the 0731 update.DeepSeek's Next Move: A $19.99 Coding Plan?

The more accurate description is that they've increased the value without increasing the price.

The prices were already very competitive, especially the cache hit price. For coding agents, every step forward requires re-prompting with the platform's instructions, tool definitions, project files, and conversation history. The longer the task, the more repetitive the context becomes.

The fact that cache hits are only 1/5 of the regular input price means long tasks are much more affordable to run continuously.

However, this alone isn't what makes the 0731 update special. The most exciting part to me is that DeepSeek has significantly improved the agent's capabilities on top of the existing context and pricing.

It's starting to integrate into workflows.

What's even cooler is that it natively supports the Responses API and is specifically optimized for Codex. The official documentation explicitly states that the Codex CLI, ChatGPT desktop, and VS Code Codex extension all use the same specification, meaning you can configure it once and use it everywhere.DeepSeek's Next Move: A $19.99 Coding Plan?

The cool thing is, it natively supports the Responses API and is specifically optimized for Codex. The official docs make it clear: the Codex CLI, the ChatGPT desktop app, and the VS Code Codex extension all share the same specification — meaning you can set it up once and use it everywhere.DeepSeek's Next Move: A $19.99 Coding Plan?

There are already people in the public community who have connected the 0731 version to the Codex CLI and successfully generated runnable games in a single session. While some praise its effectiveness, others point out that the animations and completion are still rough. To be fair, these individual examples don't prove it's superior to all competitors, but they do show that this model is capable of real-world performance, not just good-looking benchmarks.

OpenCode has also quickly integrated the new DeepSeek V4 Flash into both their paid and free versions, with public feedback consistently noting that it is significantly better than the preview version.DeepSeek's Next Move: A $19.99 Coding Plan?

The big reveal is yet to come. The official data shows that 0731 has outperformed the Flash Preview on tests like Terminal Bench 2.1, DeepSWE, and Toolathlon-Verified, sometimes even surpassing V4-Pro-Preview.DeepSeek's Next Move: A $19.99 Coding Plan?

OpenCode quickly added the new version of DeepSeek V4 Flash to both its paid and free plans. The most consistent feedback in the public reviews is that the performance is significantly better than the preview version.DeepSeek's Next Move: A $19.99 Coding Plan?

This actually makes me more curious. If the model cost stayed the same, the context length stayed the same, and the scores went up, then the unreleased harness must have contributed significantly to that improvement.

Today, news of the DeepSeek Harness beta test has surfaced. The official statement confirms that the harness minimal mode will be released soon, followed by the V4-Pro official version.

The 0731 release outperformed Flash Preview in several tests including Terminal Bench 2.1, DeepSWE, and Toolathlon-Verified. In some cases, it even surpassed V4-Pro-Preview.DeepSeek's Next Move: A $19.99 Coding Plan?

Thinking back to the "19.9 yuan" idea, that isn't because the 0731 price dropped. Instead, DeepSeek is packing more features into the existing cost structure. I'm anticipating that once the harness is out, they will bundle it with the low-cost token pricing into a "Coding Plan Lite."

The "19.9 yuan" isn't a random guess; it's a logical extrapolation based on DeepSeek's current trajectory. I expect them to offer a weekly subscription model where you get three resets per week.

Of course, this is speculation, not official news. I don't know if DeepSeek will actually launch a 19.9 yuan coding plan. But the potential combination of this price point with a robust harness makes me genuinely excited about the future possibilities.

Once the harness is released, the picture becomes complete: you have the model, the interface, and the workflow manager all working together, potentially making this a daily-use tool for many developers.

Today, rumors surfaced again about the DeepSeek Harness entering beta. The official announcement explicitly states that the Harness minimal mode will be released soon, followed by V4-Pro.

This is why I think the Harness is the most important thing to watch. The model upgrade is already done, the price and context limits haven’t changed — but the system that determines how the coding agent actually runs is still missing.

19.9元从哪儿来的

Speaking of which, a very specific number popped into my head: 19.9 RMB.

This isn’t because 0731 got cheaper. In fact, DeepSeek hasn’t touched the 1M context limit or the pricing. They’ve just added more agent capabilities under the same cost.

So I’m hoping they’ll bundle the Harness, the Codex integration, and these cost-effective tokens into a “Coding Plan Lite”.

The 19.9 RMB price is my best guess based on DeepSeek’s current moves.

And if they’d throw in “weekly reset of 3 quotas”, I’d literally be shouting “Beam, I love you!”DeepSeek's Next Move: A $19.99 Coding Plan?

Obviously, this is my wishful thinking — not official news.

I don’t know if DeepSeek will actually offer a 19.9 RMB Coding Plan.

But combining that price with the upcoming Harness does make me excited.

Imagine if the Harness releases — then the whole picture is complete: you have the model, the API, and the orchestration layer.

That’s when a coding agent could become a daily-use tool, not just a weekend experiment.

万一呢。。。

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