Krea 3 与 Grok Imagine Image 2.0 对比：一款模型有 Elo 评分，另一款只有秒表

There is a head-to-head table you can build for K​rea 3 and Grok Imagine Image 2.0, and then there is the one that actually exists. Grok Imagine Image 2.0 has been measured: x​AI's editor, released 7 August 2026, sits at 1,110 Elo on the Artificial Analysis image-editing leaderboard — thirteenth of the ranked models, on 7,792 blind comparisons, at a listed $60 per 1,000 images. K​rea 3, released 15 October 2026, has been measured by nobody. It arrives with a vendor stopwatch — under 8 seconds per edit at native 2K for $0.06 — and no third-party number of any kind. So this is less a contest between two editors than a choice between two kinds of evidence.

那种对比的两半都是真的，而且互不抵消。Elo 告诉你的是，当人群看不到标签时会偏好什么；它无法告诉你你的标识是否在编辑中幸存下来。延迟数字告诉你的是厂商自家基础设施在厂商自家提示词上做了什么；它无法告诉你这个模型到底好不好。把两者放在一起读，才是唯一诚实的选型方式，而这也是本文余下部分要做的事。

每一项是什么，用供应商自己的话说

Grok Imagine Image 2.0 is positioned by x​AI as an editing environment rather than a one-shot generator, and its feature list reads like one: Magic Wand for region-level edits, segmentation for precise region selection, background removal, multi-image input, a Smart Resize that recomposes one image into nine aspect ratios by inventing new frame content rather than cropping, and templates for product photography, headshots, e-commerce listings, game assets and marketing posters. It is the default Quality Mode on grok.com/imagine and on mobile. The API additionally exposes aspect ratio — including 21:9 and 5:2, both new with 2.0 — resolution, and a low/medium/auto quality parameter.

K​rea 3 is narrower and stranger. It is an editor with no generator story attached: region-scoped changes in a drawn box, character consistency across a sequence, up to six references on the launch post and eight on the product page, native 2K, and a voice-note workflow in the app where you speak an instruction and get the edit back. Where G​rok ships named tools and presets for verticals, K​rea ships one mechanism — find the region, render only that — and a speed claim.

• Released — K​rea 3: 15 October 2026. Grok Imagine Image 2.0: 7 August 2026.

• Independent editing rank — K​rea 3: unranked, no entry. Grok Imagine Image 2.0: #13 at 1,110 Elo on 7,792 samples, per Artificial Analysis.

• Price — K​rea 3: $0.06 per image at 2K, $0.03 at 1K, flat. Grok Imagine Image 2.0: from $0.04 per image, billed at the quality tier actually served; the board's representative figure is $60 per 1,000.

• Speed — K​rea 3: under 8 seconds end to end at 2K, vendor-stated. Grok Imagine Image 2.0: no figure from x​AI and no independent measurement; the only seconds-per-edit number we can find for it is K​rea's own, at 19 seconds, which is a competitor's measurement on a competitor's traffic.

• Resize and reformat — K​rea 3: not a documented workflow; edits are region-scoped. Grok Imagine Image 2.0: Smart Resize across nine aspect ratios, recomposing rather than cropping.

• Multi-image input — K​rea 3: 6 or 8 depending which K​rea page you read. Grok Imagine Image 2.0: five source images per edit, per current documentation, up from the three the API capped at launch.

• 权重 — 都没有。双方均无开放权重；两者都是托管模型。

读一排十三个

面对 Elo 分数，人们容易把那个数字当作一纸判决。其实，它更有用的地方在于：它是一张标出分歧所在之处的map。 Wait — no, must not leave English. Let me redo: "它更有用的地方在于，它是一张标示出分歧所在之地的地图。" Final: 面对 Elo 分数，人们容易把那个数字当作一纸判决。其实它更有用的用法，是把它当成一张标示出分歧所在之地的地图。

On the Artificial Analysis image-editing board, Grok Imagine Image 2.0 sits at 1,110 Elo across 7,792 blind comparisons, listed at $60 per 1,000 images. For scale, the top of that same board is O​penAI's GPT Image 2.5 Sunburst at 1,198 on a stated interval of 1,187 to 1,209, its Flare sibling at 1,177, and a tie for third between Nano Banana 2.1 and M​icrosoft's MAI-Image-2.6 at 1,148. Those gaps are wide relative to the intervals — Sunburst to G​rok is 88 Elo — but they say nothing at all about latency, and G​rok's entry carries less than half of Sunburst's sample count.

Two caveats to hold. The board lists G​rok's creator as S​paceXAI, which is a housekeeping detail worth knowing before you search for the row and fail to find it. And the rank is board-specific in a way that matters here: the same model sits fifth on the text-to-image board at 1,166 Elo, and thirteenth on editing at 1,110. If you are buying an editor, the editing row is the one you are buying, and it is the lower of the two.

The sharper version of that comes from the tier it replaces. The older grok-imagine-image-quality slug still sits on the editing board at 1,091, nineteenth place, so on editing the new tier is ahead of the old one by roughly nineteen Elo, for a $0.01-per-image premium. On text-to-image the same generational step was worth well over a hundred. x​AI shipped a much better generator and a modestly better editor, and only one of those improvements would have led a launch post.

And G​rok's independent reputation to date is uneven in a specific way: in a six-prompt, single-seed comparison against gpt-image-2 we covered previously, x​AI's model led clearly on photorealism and identity retention — correct portrait anatomy, pore-level skin detail, natural catchlights — and lost on production-critical behaviour, rendering Traditional Chinese characters when asked for Simplified in a film-poster headline, and returning a photorealistic image with only a light painterly pass when a watercolour style fusion was requested. That is one small test, not a body of evidence, and we said so at the time. But the failure it found — multilingual text compliance — is the kind that takes a localisation pipeline down, and no Elo average will ever surface it.

K​rea 3 has no equivalent record in either direction. It is four days old, has no board entry, and the only third-party signal attached to it is a customer testimonial: Pixelcut's co-founder and CEO, quoted by K​rea, saying his team ran early checkpoints in front of paying users and that lighting on product shots was the standout. Useful, unverifiable at our end, and labelled here as a partner statement.

比价格差距更重要的功能差距

从纸面上看，价格很接近：六美分对上一个“$0.04 起”，而后者会随着实际提供的质量档位上涨。差别不在于那几美分，而在于每个模型愿意对你的画面做什么。

Grok Imagine Image 2.0 will reformat. Smart Resize invents new content to turn one image into nine aspect ratios rather than cropping, which is exactly what an ad team needs when the same hero has to run at 1:1, 9:16 and 2:1. It will also composite up to five sources into one edit. K​rea 3 will not do either, as documented — its pitch is the opposite discipline. Draw a box, describe the change, and everything outside the box stays where it was, with precision defined as the share of pixels outside the edit area left unchanged, compared at quarter resolution after alignment.

Those are complementary, not competing, and the cost of choosing wrongly is asymmetric. A marketing team that buys K​rea 3 for a reformatting workflow gets a model that is excellent at a job it was not asked to do. A catalogue team that buys Grok Imagine Image 2.0 for label-accurate product edits gets a very capable editor with no published containment metric — and possible multilingual text risk on top, per that one test.

按目录规模的费用，附带两个注意事项

For K​rea 3 the arithmetic is trivial and the risk is that it is the only arithmetic available. 100,000 edits at $0.06 is $6,000, and K​rea's own FAQ says the launch price is the list price, not a subsidy, and that it expects the price to fall rather than rise. That is a vendor statement about the future, so a budget built on it is a bet.

For Grok Imagine Image 2.0 the shape is different, and there is a deadline hiding in it. The model bills from $0.04 per image at the quality tier actually served, and the default when you omit the parameter is auto — which currently serves low for generation and medium for editing, so the same request can bill differently depending on which path the service picks. Pinning the tier makes the bill predictable. Meanwhile the older grok-imagine-image-quality slug, at $0.05 per image, retires on 2 November 2026 after a 60-day notice; from that date it keeps resolving but requests are served by Grok Imagine Image 2.0 at low quality. Teams that change nothing see a price drop and a silent change in output quality, which is the worst of both — unless you read the model field in your responses. That retirement schedule is x​AI's own documentation, not something anyone has independently verified.

两者放在哪里，以及如何握住它们

这两个模型，本月都不该是你直接接入面向客户的流水线的选择。其中一个有排行榜名次，名下还有一个已退役的 slug，以及一桩有记录的多语言文本故障；另一个有价格、有秒表，却没有任何外部证据。对两者而言，合理的做法都是在有兜底方案的前提下进行试用。

That is the specific thing routing buys you here, and it is worth being precise about what we do and do not cover. OrcaRouter provides one OpenAI-compatible endpoint across a few hundred models with provider list prices passed through at no markup — a vendor price cut reaches your bill the same day — plus failover that sends a provider error or rate limit to a fallback rather than into your batch. On availability: our catalogue carries x​AI's grok-imagine-image line and its model page is live, but the newest Grok Imagine Image 2.0 tier is a separate entry, so check the current model list rather than assuming the 2.0 quality mode is routable today. K​rea 3 is not routed by us at all. What the endpoint does cover is the rest of the comparison set — GPT Image 2 and the image models you would benchmark either of these against — so the test you should be running is one configuration, not three integrations.

证据一边倒时，该如何决定？

如果你今天需要一款拥有公开、由大众评测得出的质量定位的编辑器，那就是 Grok Imagine Image 2.0；你应该事先知道，它排名第十一，而不是第一；除非你将其固定，否则它的账单会随质量层级变化；而且它的 slug 级迁移是有日期限制的。请阅读你响应中的 model 字段，并在 2026 年 11 月 2 日之前迁移即将退役的质量 slug。

If you need flat pricing, contained edits and the shortest wall-clock time anyone in this category has published, that is K​rea 3 — with the understanding that you are the first person testing it, that its reference count is documented two ways, and that its identifier appears as K​rea 3 Beta in K​rea's own plan tables.

决定胜负的不是模型本身，而是这两种失效模式中，哪一种是你承受得起的：一次编辑要花四十秒，却会在本地化上给你"惊喜"；还是另一次编辑八秒就能返回结果，但除了它的制造者，从未被任何人评判过。拿你自己的四十张图片把两者都跑一遍，让答案属于你自己，而不是任何一家供应商。

对于一个失效模式尚未被记录的模型来说，有用的对照物是别处已经出错的记录：这一类模型的已记录 AI 事故与智能体故障，都收录在OrcaRouter AI 事故档案库中，其中每条条目都带有自己的来源、严重程度、置信度评级和争议标记。