Six AI Media Platforms Compared. Only One Remembers What It Made.

Six platforms will happily generate you an image. Only one of them remembers what it did.

That sounds like a small distinction. It isn’t. If you’re running media generation as a team process instead of a party trick, it’s the whole ballgame, and it’s the thing nobody puts on a pricing page.

I went through six platforms that all get filed under “AI media generation” and priced them in the same week: Oxen.ai, fal.ai, Atlas Cloud, Replicate, Runway, and Artlist. What follows is what each one actually does, what it costs, and which one I’d bet a production workflow on.

Generation is a transaction. Production is a pipeline.

Nearly every platform here sells the same shape of thing. You send a prompt, you get a file, the transaction closes. That’s a fine way to make one image. It’s a lousy way to run a media operation.

Here’s the test I keep coming back to. Your best-performing product video was generated four months ago. Can you answer these five questions about it today, without asking the person who made it?

1. Which model and version produced it?
2. What was the exact prompt and seed?
3. Which reference assets went in?
4. If a fine-tune was involved, what data trained it?
5. Can you regenerate it, changing exactly one variable?

On five of these six platforms, the answer is no, or it depends on a spreadsheet somebody maintains by hand. Replicate deletes API prediction records after one hour. Atlas Cloud has an asset library but no lineage. Runway has folders. Artlist has a history panel inside a project. fal.ai keeps a list of your recent training runs and that’s about it.

None of that is incompetence. It’s a category decision. These are inference platforms, and an inference platform’s job is turning a request into a file fast and cheap. Judged on that, fal.ai and Atlas Cloud are genuinely excellent, and both are quicker and cheaper per image than the platform I’m about to argue for.

But the moment generation stops being one person improvising and becomes a team shipping assets on a schedule, against a brand standard, with a fine-tuned house model, that missing layer stops being a nice-to-have. You need to know what produced what. You need to reproduce it. And you need the data that trained your model to still exist in six months.

That’s the pipeline gap, and Oxen.ai is the only platform here built to close it. It’s the same shift we help teams make with the rest of their AI workflows: from one person improvising to a process that survives a sick day.

The short version: Oxen.ai is a Git-style version-control system for datasets and models, with an inference API and managed fine-tuning on top. The other five are inference or creative tools with storage attached. That architectural difference is the whole argument, and it’s also why Oxen loses on catalog size, price per image, and company maturity. Both halves are below.

The field

Six platforms, sorted by what they actually are underneath the marketing.

Platform	What it really is	Catalog	Billing	Company
Oxen.ai	Data version control + inference + managed fine-tuning	~200 models	Pay-per-use, credits don’t expire	Small, ~$6.75M raised
fal.ai	Generative-media inference + serverless GPU	1,000+ models	Pay-per-use only	Independent, ~$4.5B valuation
Atlas Cloud	Multi-model inference aggregator	449 models	Pay-per-use + coding plans	Self-reported bootstrapped
Replicate	Model marketplace + custom deploy	50,000+ models	Pay-per-use, GPU-second	Cloudflare subsidiary
Runway	Foundation-model lab + creative suite	Own models + resold	Subscription + credits	Independent, ~$5.3B valuation
Artlist	Licensed stock library + model reseller	900k assets, ~100 models	Subscription + credits	~$300M ARR

Two things worth flagging before we go further. Replicate was acquired by Cloudflare (announced November 2025, closed December 1); its catalog is being folded into Workers AI. And Artlist is mostly a stock library with a generation front-end attached, reselling other people’s models through one credit pool. Its real product is licensing certainty, not generation.

What each one does and doesn’t do

The capability matrix. This is the table I’d actually use to make a decision.

Capability	Oxen.ai	fal.ai	Atlas Cloud	Replicate	Runway	Artlist
Dataset version control	Full (Git-style)	None	None	None	None	None
Output / prompt lineage	Committed to repo	None	Asset library	1-hour retention	Folders, tags	Per-project history
Model versioning	In repo	Run history	None	Immutable hashes	Vendor-managed	Vendor-managed
Self-serve fine-tuning	Managed, priced	LoRA, flat fee	DIY on GPU pods	LoRA, GPU-second	Voice only	None
Automated evaluation	Dataset-wide API	Sandbox compare	Model Explorer	Playground	Parallel gens	Variants
Review / approval flow	Workspaces	None	None	None	Frame comments	None found
Self-host / VPC	Open-source server	Enterprise	Enterprise	Enterprise	No	No
Bring your own model	Via fine-tune	Serverless deploy	LoRA import	Cog packaging	No	No
Roles & permissions	Not published	3 roles + SSO	Full RBAC	Not published	Enterprise only	Unverified
MCP server	None found	Two	Yes	Official	Yes	No
Output indemnification	Not published	None found	None	Enterprise add-on	Enterprise only	None found

The pipeline test

A media pipeline has four stages where state has to survive between sessions. Judge the platforms on each handoff and the field thins out fast.

Stage	What it means	Who can actually do it
1. Curate	Assemble and maintain the reference set: brand assets, approved shots, training images. It changes, and you need to know what changed.	Oxen (commits, branches, diffs). Everyone else: a folder.
2. Train	Turn that set into a house model that knows your product and your style. Cheap now. The hard part is knowing which data produced which weights.	Oxen, fal, Replicate. Atlas: DIY on rented GPU pods. Runway: voice only. Artlist: no.
3. Evaluate	Run a candidate against a fixed test set, score it, compare to last month’s model. Not eyeballing two images side by side.	Oxen (dataset-wide eval API). Everyone else: manual and visual, one prompt at a time.
4. Ship and trace	Generate at volume, then six months later reconstruct exactly how any asset was made.	Oxen (committed lineage). Replicate deletes API predictions after an hour.

What Oxen actually gives you

The core is an open-source Rust version-control engine, Apache-2.0 licensed, built for data rather than code. Familiar verbs, different guts: init, add, commit, push, branches and merges, over a Merkle tree with block-level deduplication. It’s not Git-LFS with extra steps.

Three features do the real work:

• Workspaces are server-side staging areas. An agent or a batch job stages new rows and assets all day, and a human commits only the approved batch. Their docs call it a pull request for data, which is exactly right.
• The evaluations API runs a model across every row of a dataset in your repo, tracks token usage, and commits the scored results back as a new commit. Comparing models becomes a diff between two commits instead of a vibe.
• Managed fine-tuning provisions the GPUs, trains, writes the weights back into your versioned repo, and spins the hardware down. Billed per second, with per-model rates published in their live API.

The one performance claim with a number attached: syncing a million-image ImageNet dataset took 1 hour 30 minutes on Oxen, against 3 hours for DVC on local storage and more than 20 hours for Git-LFS. That’s their own benchmark, so discount it accordingly. The Git-LFS number matches what anyone who’s tried it will tell you.

The one customer number on record: AlliumAI, via Oxen’s GPU partner Baseten, cut a Qwen-Image-Edit fine-tuning and inference workload from $46,800 to $7,530 a month. That’s an 84% reduction at roughly half the latency. One named customer with one number is thin evidence. It’s also more than any other platform here published.

What it costs

Two different pricing philosophies are running here, and confusing them is how teams get surprised.

Platform	Model	Entry	Notable mechanic
Oxen.ai	Pay-per-use	$0, credits	Credits never expire; fine-tuning billed per GPU-second
fal.ai	Pay-per-use	$0	Per output unit, or per GPU-hour for custom deploys
Atlas Cloud	Pay-per-use	$0, free credits	Credits expire after 365 days; all sales final
Replicate	Pay-per-use	$0	Public models bill active time only; private models bill idle too
Runway	Subscription + credits	$15/mo	No credit rollover on Standard or Pro; teams $69/seat
Artlist	Subscription + credits	~$19.99/mo	No rollover; credit values were rebased mid-2026

Oxen’s two pricing axes

Oxen is the one platform here that charges on two separate axes, and conflating them is how you’d misread its pricing. Storage and transfer sit on a subscription. Models and GPUs are always pay-as-you-go on top, with no subscription required. You can run the whole inference side on a free account and never pay a monthly fee.

Plan	Monthly	Repos	Storage & transfer
Explorer	Free forever	Unlimited public, 5 private (max 3 collaborators)	50GB each
Hacker	$30	Unlimited private	100GB each, expandable
Pro	$60	Unlimited private	500GB each, expandable

For a DigiSavvy-sized team that matters twice over. The free tier is genuinely usable for a first house model, and the non-profit discount is worth asking about if you work with 501(c)(3) clients, which is a lane where none of the other five platforms offer anything comparable.

I run the pay-as-you-go side myself and prefer it. Topping up credits beats another monthly line item, and it keeps the cost attached to actual work instead of a seat count.

What the GPUs actually cost

Fine-tuning and dedicated inference bill per second of GPU time, and idle instances shut themselves down after 15 minutes. That auto-shutdown is the detail that keeps a forgotten training job from quietly eating a month’s budget.

Instance	GPU	VRAM	Per hour
A10G	1× A10G	24 GiB	$1.65
H100 MIG	1× H100 (40GB)	40 GiB	$1.95
H100	1× H100	80 GiB	$4.87
H200	1× H200	137 GiB	$9.98
4×H100	4× H100	320 GiB	$15.00
2×H200	2× H200	282 GiB	$19.50
8×H100	8× H100	640 GiB	$38.99

Oxen also splits fine-tuning into full fine-tunes and LoRA, and tells you per model which one it supports. Some, like Krea 2 Raw and MiniMax H3, are LoRA-only. Others, including the smaller Gemma 4 models, the Qwen 3.5 family and LTX-2.3 Pro, support both. Nobody else here publishes that distinction on a pricing page.

Same model, different platforms

The cleanest comparison is pricing identical work. These are published rates for the same underlying models, pulled the same week.

Workload	Oxen.ai	fal.ai	Atlas Cloud	Replicate	Runway
FLUX.1 [dev] image	$0.025	n/a	n/a	$0.025	n/a
Qwen Image	$0.025	$0.02/MP	$0.028	n/a	n/a
Nano Banana 2 image	$0.059–0.196	$0.06–0.16	$0.028–0.08	n/a	~$0.07
Kling Pro, per video second	$0.07	$0.07	$0.06	n/a	n/a
Veo 3, per video second	$0.20	$0.40	$0.05–0.40	n/a	n/a
LoRA training run	~$4.87/GPU-hr	$2 flat (FLUX)	DIY on GPU pods	~$1.85–2/run	Not offered

Read this before you budget: per-generation price is the least important number in that table for most teams. A studio doing 5,000 images a month on FLUX.1 [dev] spends about $125 either way. The costs that actually bite are the retry you paid for twice because nobody recorded the working prompt, and the fine-tune you rebuilt from scratch because the training set was in someone’s Dropbox. Both are pipeline costs. Neither appears on any pricing page.

The subscription platforms, in plain numbers

Tier	Monthly	Annual (per month)	What you get
Runway Standard	$15	$12	625 credits, roughly 52 seconds of flagship video
Runway Pro	$35	$28	2,250 credits, 1 brand kit, 1 voice clone
Runway Max	$95	$76	9,500 credits, rollover, HDR export
Runway Team	$69/seat	n/a	Pooled credits, no SSO or roles
Artlist AI Starter	$19.99	$11.99	16,500 credits
Artlist AI Core	$39.99	$23.99	40,000 credits
Artlist AI Creator	$69.99	$41.67	80,000 credits

That Runway Standard line deserves a second look. Fifteen dollars a month buys roughly 52 seconds of Gen-4.5 video. Not 52 clips. Fifty-two seconds, for the whole month, and a generation that finishes but misses still spends the credits (Runway refunds outright errors). The most common complaint about both subscription platforms is exactly this gap between the sticker price and the work it covers.

The case against Oxen

A recommendation you can’t argue with isn’t a recommendation, it’s an ad. So here’s where Oxen genuinely loses.

It’s a small company and you’d be betting on it

Oxen Labs, Inc. has about $6.75M in disclosed funding (an SEC filing from December 2024), a headcount somewhere between eight and a few dozen, zero G2 reviews, and one named customer. Compare that to fal at a reported $4.5 billion valuation, Runway at $5.3 billion, and Replicate now sitting inside Cloudflare. If a platform is going to be your system of record for production media, vendor durability is a real line item, and Oxen scores worst here by a distance. I’ve written before about why I’m suspicious of platforms that can change the deal on you, and a media stack is no different.

A script can’t read their pricing page

Their pricing page is clear and complete once you’re looking at it in a browser. It’s the only one here that publishes per-GPU hourly rates and splits full fine-tunes from LoRA per model. But the marketing site sits behind a bot checkpoint that refused every automated request across two days, so that pricing table isn’t reachable by a script. (Their docs are fine.) If you’re the kind of team that tracks vendor pricing programmatically, budget for doing that by hand.

The catalog is a fifth of fal’s

Around 200 models against fal’s 1,000-plus and Replicate’s 50,000. The headline models are all there, Flux, Qwen, Wan, Veo, Kling, Seedance, Sora, so most teams won’t notice. If your workflow depends on a specific niche checkpoint, check before you commit.

No published roles, SSO, or audit logs

Atlas Cloud publishes a full permissions model with five predefined roles. fal.ai documents three roles plus enforceable single sign-on and SOC 2. Oxen’s answer to compliance is a private cloud or on-prem deployment arranged through sales. That’s a fine answer for a large enterprise and no answer at all for a twelve-person team that needs to cut off a departing contractor on a Friday.

No MCP server, and unverifiable terms

Atlas Cloud, fal, Replicate and Runway all ship one. Oxen doesn’t, which in 2026 is a real gap in how engineers expect to reach a platform from inside their editor. And output ownership, commercial-use rights and indemnification all live on that same blocked domain, so get them in writing before client work depends on them. To be fair, indemnification is weak or absent across this whole category outside enterprise tiers.

When to skip Oxen outright: if you’re generating at volume with no training and no reproducibility requirement, its advantages are inert and you’re paying an attention tax for nothing. Use Atlas Cloud for price or fal.ai for speed and tooling. If your team is creative rather than technical, Runway’s suite beats a command line every single day.

Who should pick what

Pick Oxen.ai if you’re building a repeatable media operation

A house model trained on your own assets, a team producing on a schedule, and a requirement to reproduce or audit any asset later. Also the right call if you have data-residency constraints, since the server is open source and self-hostable. It’s the only platform here where the dataset, the model and the outputs live in one versioned place. If you’re not sure media is even the right thing to automate first, start with the automations I’d set up for any small business.

Pick fal.ai if you’re shipping generation inside a product

Best developer experience of the group, the deepest catalog that’s still curated, real serverless deploy for custom models, streaming endpoints, and the strongest team controls of any pay-per-use platform here. Flat $2 FLUX LoRA training with commercial rights included is the simplest training story in the category.

Pick Atlas Cloud if cost per generation decides it

Often cheaper than the model owner’s own pricing, with 449 models at the time of writing, proper permissions, an MCP server and unusually clear billing docs. The tradeoff is explicit: no managed training, no versioning, and you’re trusting an aggregator to keep pace with dozens of upstream vendors.

Pick Replicate if you need an obscure model

50,000 models because anyone can publish, with the quality variance that implies. Immutable hash-pinned model versions are the best reproducibility primitive here after Oxen’s. Watch the one-hour prediction retention and the idle billing on private models.

Pick Runway if your team are editors, not engineers

The only one here training its own frontier video models, wrapped in a production suite with frame-accurate commenting, brand kits, and an Adobe Premiere panel that puts generation on the timeline. Buy it as a creative suite. The credit math is unforgiving and it’s no longer the outright quality leader, but nothing else here is a production tool in the same sense.

Pick Artlist if licensing certainty matters more than capability

You’re buying a cleared catalog with a broad commercial license and generation attached, which is a legitimate thing to want when a client asks who owns the footage. As a generation platform it’s the weakest here: no training, no versioning, and credit costs that have already been rebased once.

What I couldn’t verify

Every number above came from a vendor page, a live API, or a vendor document on the date shown. These specific items didn’t survive verification, and you should confirm them before anything contractual depends on them.

Item	Status	Why it matters
Oxen terms and output rights	Unverified	Terms, privacy policy and a Trust Center are all linked from their site, but weren’t machine-readable for this review. Confirm ownership and indemnification in writing.
Oxen founding date	Disputed	Sources say 2020 or 2022. The open-source repo was created August 2022.
Artlist pricing tiers	Directional	Live pages refused automated requests. Figures come from an archived snapshot plus conflicting reviewer citations, and a mid-2026 credit rebasing makes older numbers unreliable.
Runway Team annual price	Inferred	Monthly seat price confirmed; the annual equivalent wasn’t published in a readable form.
Atlas Cloud SOC 2 / HIPAA	Claimed	Prominent site-wide claim; no certificate or auditor letter was obtainable.
All user-sentiment items	Anecdotal	Reddit, Trustpilot and Hacker News complaints are directional signal from small samples, not reliability data.

One more thing worth knowing. Runway is a named defendant in ongoing copyright litigation over training data, including a class action filed in February 2026. Artlist’s position is the mirror image, since its whole pitch is a cleared, licensed catalog. If provenance risk is live for your clients, that contrast should weigh more than any feature in the matrix above.

Pricing in this category changes monthly. Everything here was verified between 19 and 22 September 2026, and I’d re-check any number before you commit budget to it.