xAI's Imagine Image 2.0 Is on the API at $0.04 to $0.08 an Image

Second on the board in August, and on xAI's API now

A media buyer drops a leaderboard screenshot into the team channel with one line under it. Should we switch? The screenshot is real, and so is the board. When we published this post on August 8, the honest answer took about ninety seconds, because nobody on that team could call the model in second place. They can call it now, so the answer takes a little longer.

We corrected this post on November 3, 2026, after xAI's API changed the answer. We didn't test the API for this post. The API facts come from xAI's own docs (they now carry the name SpaceXSI), read that day, and the launch-day numbers keep their August dates.

On August 7, 2026, xAI shipped Imagine Image 2.0 into Grok's web and mobile apps, and its fast variant debuted second on both of Arena's image boards. On Image Edit it scored 1439 Elo, behind OpenAI's gpt-image-2 (medium) at 1463. On Text-to-Image it scored 1320, behind gpt-image-2 at 1380. That's two silver medals in one day.

Then there's the part that decides whether any of it reaches your ad account. On launch day, xAI said access was coming soon and gave no date, and its developer price list showed two image models, neither of which was this one. That's changed. xAI's docs, read November 3, list grok-imagine-image-2.0 on the API at $0.04 to $0.08 per image, by resolution and quality. On both of those August 7 boards, gpt-image-2 was first, and GPT Image 2 is the model the Novoads image-to-ad and product-to-ad flows already run. So the short answer to "does this change what I generate ad creative with" has gone from no to maybe, and a test on your own product shots settles it. The long answer is the useful one, because it tells you what to check.

What actually landed on August 7

Three details in this launch carry most of the weight, and the aggregate coverage flattens all three. They're worth ten minutes now, because they'll matter again at the next model launch.

The entry that ranked was not the model in the app

xAI launched Imagine Image 2.0 as a new Quality Mode on grok.com/imagine and in the Grok iOS and Android apps. The row at number two on Arena that day was labeled grok-imagine-image-2.0 (low), the fast tier of that family. The Quality Mode that consumers actually got wasn't the row that ranked.

OpenAI's leading row that day was gpt-image-2 (medium), which is also a tier rather than a ceiling. So the comparison people were celebrating was one speed tier of a new model against one middle tier of an incumbent. That's a legitimate result, and a narrower one than "second best in the world." Any headline that said Imagine Image 2.0 ranked second, with no qualifier, quietly promoted a variant into a product.

Preliminary is not a formality

Both Grok rows carried Arena's Preliminary flag that day, and the vote counts explain why.

Board, as of Aug 7, 2026	First	Second	Votes, first vs second
Image Edit	gpt-image-2 (medium), 1463	grok-imagine-image-2.0 (low), 1439	184,189 vs 5,931
Text-to-Image	gpt-image-2 (medium), 1380	grok-imagine-image-2.0 (low), 1320	69,194 vs 2,722

A model with 5,931 votes and a model with 184,189 votes weren't being measured with the same instrument. The confidence intervals said so out loud. On Image Edit, it was plus or minus 8 on the challenger and plus or minus 4 on the leader. On Text-to-Image, it was plus or minus 12 against plus or minus 5. A day-one placement is a first impression rather than a verdict, and xAI itself date-stamped the claim instead of stating it flatly.

On one board, second wasn't separated from third

Text-to-Image was the thinner of the two results. The second row read 1320, plus or minus 12. The third row, Reve 2.1, read 1302, plus or minus 8. Those ranges overlapped. Statistically, you couldn't yet tell second from third on that board, which meant the order could change without the model changing at all.

None of this makes the release unimpressive. But the number was provisional, and provisional numbers are exactly the ones that get quoted for months after they've stopped being true.

What you can call on xAI's API now

Leaderboard stories are easy to write and easy to read. Availability stories are neither, which is why the fact in this launch that matters most to your decision is the one that travels least.

What xAI's price list actually sells today

We read xAI's developer pricing page on November 3, 2026 (the page says it was last updated October 10). It lists two Imagine API image models. grok-imagine-image runs $0.02 per image at both 1K and 2K. grok-imagine-image-2.0 is priced by resolution and quality, in US dollars per image:

Resolution	Low	Medium
1K	$0.04	$0.06
1.5K	$0.05	$0.07
2K	$0.06	$0.08

Each input image adds $0.01 on grok-imagine-image-2.0 and $0.002 on grok-imagine-image. If you leave the quality setting out, xAI's generation docs (last updated October 8) say the default is auto, which currently uses low for new images and medium for edits. So set it yourself when the budget matters.

The row that's no longer there is grok-imagine-image-quality. An archive copy of the page from August 10 lists it at $0.05 per image at 1K and $0.07 at 2K, with no 2.0 row yet. xAI's migration guide, last updated September 2, says that model is retired effective November 2, 2026, and that requests to it are served by grok-imagine-image-2.0 at low. The guide says they're billed at the low rate, which it puts $0.01 per image below the old model at every resolution, and that grok-imagine-image isn't affected.

Where the callable models rank

xAI's migration guide says grok-imagine-image-2.0 at low "ranks #2 on the Arena Image Edit leaderboard and #3 on Text-to-Image." That's xAI's wording, last updated September 2. We couldn't read Arena's live boards on November 3, so we can't give you that day's order. The newest copies we could read are archive copies from October 8, stamped October 6 and October 7.

On both boards, those copies list:

• gpt-image-2.5-sunburst first
• gpt-image-2.5-flare second
• gpt-image-2 (medium) third
• a row labeled grok-imagine-image-2.0 (canvas) fourth

grok-imagine-image, the cheaper model, is 22nd on Image Edit and 25th on Text-to-Image in those same copies.

On launch day the split was wider. In an archive copy of Arena's overview from August 9, the top row for a model a developer could buy that day, a dated grok-imagine-image-quality (20260519), was 7th on Image Edit at 1390. On Text-to-Image, grok-imagine-image-quality was at 14. It had 1228 Elo on 48,786 votes.

So that week, the headline number and the models you could buy were two different populations, roughly a hundred Elo points apart on Text-to-Image. We think that split has closed, because the low setting that ranked is one you can now ask for by name.

What coming soon turned into

On launch day, xAI said API access for Imagine Image 2.0 was coming soon. That was a real commitment, but you couldn't plan around it, because four things a creative pipeline needs were all missing. As of our November 3 read, xAI's docs supply three of them, and the fourth is still open:

• A model ID to write into a config, grok-imagine-image-2.0, next to grok-imagine-image.
• A published developer price, so a month of variations is a number rather than a guess. It's the table above.
• A rate limit, so you can design a batch up front instead of discovering the limit later. xAI's rate-limits page, last updated October 5, gives 6 requests per second at the default tier and says to contact sales for higher Imagine limits.
• A date, so you can schedule a migration against something, is the one we still can't give you. The pages we read don't state the day the API opened. An archive copy of xAI's launch post from August 11 still says "API access is coming soon," and the copy from August 12 says "Image 2.0 is available in the API as grok-imagine-image-2.0."

That was the difference between a model that exists and a model you can build on. An ad team uses an image model by calling it forty times on a Tuesday, in the ratio the media plan asked for, against a product photo that has to stay recognizable. Everything in that sentence needs an endpoint, and as of November 3 this model has one.

An Elo is not an ad-creative score

The API is here, with published prices, but the leaderboard still doesn't answer the question an ad team is actually asking. We're keeping that apart from the availability problem, because it's still there now that the API is open.

What the vote actually measures

An Arena score aggregates blind pairwise human preferences on prompts the voters chose to write. It's a genuinely good instrument for general taste, and that's why the boards are worth reading at all. It's also a popularity measure over a mix of prompts that has nothing to do with your category.

What it doesn't measure is most of what an ad reviewer checks:

• whether the serum label stayed legible and said what the real label says
• whether the packaging shade survived the render
• whether the headline reads at thumb speed on a phone in daylight
• whether the fortieth asset in a campaign still matches the first

Those are the things that decide whether a creative ships. We made the same distinction when Meta pitched Muse Image at advertisers, and again when Microsoft's MAI Image 2.5 Pro arrived with hero-imagery language and a one-megapixel ceiling. The announcement and the constraint that stops you on a Thursday are usually written on different pages.

The typography claim is the one to test yourself

xAI positions the model on instruction-following accuracy, clean typography and layout in complex visuals, and consistency across multiple generations. Those are exactly the right things to claim for ad work, and exactly the claims a leaderboard Elo can't confirm.

Text rendering is where image models diverge most sharply and where marketing language is least reliable. Ideogram 4 built its whole release around in-image text, with explicit bounding-box layout and palette control, and Seedream 5.0 Pro leans on multilingual text rendering for teams shipping one offer into several markets. Those are structural commitments you can inspect, and a preference score isn't one.

Pin every mark, or the model writes its own

The failure mode we hit most often has nothing to do with resolution. Hand an image model a product without explicit instructions and it'll invent brand copy. The inventions are always plausible:

• a claim on the label that reads like marketing but that nobody approved
• a tagline the brand has never used
• a certification badge that doesn't exist
• a variant name lifted from the general shape of the category

It all looks correct at a glance, which is what makes it dangerous, and it survives right up until legal reads the ad.

A better model won't fix this. The fix is to pin every mark in the prompt, then transcribe what came back from your own render rather than from the gallery. The same discipline keeps an AI actor consistent across ad variations, and it's why the prompt craft around a model tends to matter more than its rank, right up until the model absorbs the workaround and turns it into a feature. A new number two on a public board doesn't, on its own, change how anyone works.

The model that topped both boards on August 7 is already drawing your ads

The quiet result in this story was that gpt-image-2 (medium) held first place on both boards on the same day, on far more votes than the challenger. The top spot has moved since. In the October 8 archive copies, OpenAI's GPT Image 2.5 Sunburst and Flare rank above it, and both are in the Novoads catalog too.

What GPT Image 2 costs inside Novoads

You don't have to go and evaluate that model, because the Novoads image-to-ad and product-to-ad flows already run it, at fal medium quality and a flat 1 credit per image. There's one number and one balance, with no separate provider key and no per-token rate to reverse-engineer before you can say what a batch of variations costs. Nano Banana Pro, Seedream 5 Pro and Seedream 5 Lite are in the same catalog for when a job calls for a different engine.

The economics matter more than the ranking here. Creative testing is a volume game, and testing creative properly means generating far more candidates than you'll ever run. A flat per-image price turns that into multiplication you can do in your head.

Do the multiplication once and the difference gets concrete. Say a week of image testing means four concepts with five variations each, in two placement ratios. That comes to 40 renders. At 1 credit an image, that's 40 credits, and you knew the number before you started. Those come out of the 200 credits our Starter plan grants each month. Now price the same week against an endpoint that bills per output token, and the same question turns into a forecast you can't settle until the invoice arrives. That's more than a rounding difference between vendors, because you're comparing a budget with a guess.

xAI's price list lets you do that sum up front too. At 1K and low, those 40 renders cost $1.60, plus $0.01 for each input image you send. That's less than 40 credits on our Starter plan. At 7.5 cents a credit, those come to about $3. It's a different model, though, and the integration is yours to build and keep running.

Why the ratio list matters more than the Elo

Our image-to-ad flow offers six aspect ratios: 4:5, 9:16, 1:1, 16:9, 2:3 and 21:9. That list is a more practical spec than any Elo on either board, because a placement won't take a creative in the wrong shape at all. Check your ratio requirements per platform before you generate rather than after, whichever model you end up using.

What would actually change the answer

None of the above is a verdict on the model. Our August verdict was on the timing, and the timing has changed. We wrote down three specific things that would move it, so the decision wouldn't get made on vibes, and the first has arrived.

The three signals to watch

• A model ID and a published price on xAI's own pricing page. This one has arrived. grok-imagine-image-2.0 has its row, next to grok-imagine-image.
• The vote counts crossing into the tens of thousands, with the Preliminary flag gone. At that point the placement is a measurement rather than a first impression, and it may well be higher or lower than 1439. In the October 8 archive copies, a grok-imagine-image-2.0 (20260801) row has crossed that line on Image Edit, at 1425 on 50,427 votes. Those copies don't say whether that row is the API's low setting, so we'd read it with care.
• A benchmark on your own product, rather than the voters' prompts. Use ten of your real SKUs, your real labels and your real headline, in the ratio you actually run. That's the only test that answers your question.

What to do this week instead

Open Grok Imagine and look at the output if you're curious. It costs you ten minutes, and it's a genuinely strong model. If your pipeline runs on an API, you can also put your own product shots through xAI's endpoint before you move anything. A creative operation is a throughput problem, and you can't get throughput without an endpoint.

Compare that with the other kind of launch news. Ad-platform changes tend to arrive with a switch-over date attached, which is what lets you plan around them. Our note on the ChatGPT Ads matching and carousel changes is the kind of story you can put in a calendar. A model with no endpoint, which is what this one was on launch day, is a story you can only react to.

If you want something useful to change this week, change the input rather than the model. When AI creative is weak, it's usually because nobody wrote an angle, and not because the renderer lost a benchmark. That's the quality gap that gets blamed on models roughly every time a new one launches.

Which document answers which question

Read the leaderboard when you want to know:

• which labs are currently competitive at all
• whether a family improved between versions
• which models are worth an afternoon of hands-on testing

Read the vendor's price list when you want to know:

• what you can call tomorrow morning
• what a month of variations will cost
• how many requests a batch is allowed to make
• whether you can schedule a migration at all

You can turn only one of those two documents into a media plan, and it isn't the one that made the headlines in August.

How Novoads solves the model-churn problem

There's a new frontier image model roughly every two weeks now. Chasing each one has a real cost, and almost none of it is judgment. It's another account, another key, another billing unit and another integration to maintain when the model gets deprecated.

One balance, and someone else's integration problem

In Novoads, you upload a product image, and the image-to-ad and product-to-ad flows generate the creative on GPT Image 2 at medium quality for 1 credit an image, in whichever of the six ratios the placement wants. When a model earns a slot, we add it to the picker on the same balance. When it doesn't, nothing changes on your side. Novoads doesn't sell Grok Imagine, so as of November 3 Imagine Image 2.0 isn't in that picker, and you'd call it through xAI's API yourself.

Where the still goes next

Then comes the step a still can't take on its own. Write or auto-generate a script, pick an AI actor whose age, gender and accent match the audience, and the approved image becomes the opening frame of a vertical clip with synthetic voice, lip-sync and captions. Our guide to making product videos with AI walks through that loop end to end, and the same handoff drives the route from one key frame to a finished animated spot.

Novoads starts at $15/month (Starter, 200 credits per month). Plus is $49/month (1,000 credits), and Ultra starts at $129/month (3,000 credits) for more volume. Cancel anytime.

Ship on what you can call

Imagine Image 2.0 is a serious release from a serious lab, and the low variant landing second on two boards on day one was a real result that deserved the attention it got. As of November 3, it also has a model ID, a developer price and a rate limit, per xAI's docs. So for anyone running ads, it's gone from a preview of a decision to a decision you can test.

This pattern is going to repeat, because the leaderboard is the part of a launch that's easy to publish and availability is the part that's easy to skip. So keep the two questions apart. A benchmark tells you what a model can do in someone else's prompt, and a price list tells you what you can build on Tuesday. When those two disagree, the price list decides what ships.