Higgsfield's Supercomputer is hosted and closed-source. It
plans the brief, picks the model from a menu they curate, runs the
render on their servers, and bills you in credits that expire on a
timer. Ralphy is the inverse: an
MIT-licensed CLI that uses your own OpenRouter and ElevenLabs
keys, runs on your machine, and writes every generation to your disk.
You pay the model vendor directly, at list price. There's no markup
and no platform fee. Want to swap Kling for Seedance or Veo? Edit one
line in MODELS.md.
MITLicense
$0Platform fee
0dCredit expiry, ever
If you ship one reel a quarter and don't write code, a Higgsfield
subscription is faster to set up and you'll feel the difference on day
one. Add a coding agent or a campaign that runs all year and the
constraints flip. Credits forfeit. Prompts get harvested. The model
menu is whatever Higgsfield negotiated this quarter. You usually
notice around month six.
What Higgsfield Supercomputer is, accurately
Supercomputer launched on May 14, 2026. It's Higgsfield's umbrella
product, an agentic stack that wraps their existing studios into one
"describe it and ship" loop. The headline pieces:
Cinema Studio 3.5 — 3 camera bodies, 5 lens types, 5 focal lengths
(including the new 75mm), 3 apertures, plus 70+ motion presets
(dolly, crane, FPV, crash zoom, bullet-time).
Marketing Studio — 25+ hook variations, the Virality Predictor
(Hook Score, Hold Rate, brain-heatmap), and a Hermes-driven campaign
builder.
Soul ID — train a digital double from your photos so the face
holds across angles, lighting, and styles.
40+ built-in tools orchestrated by the agent — scriptwriting,
upscaling, audio mixing.
Multi-model out of the box — Seedance 2.0, Kling 3.0, Veo 3,
Sora 2, Wan 2.6, MiniMax Hailuo 02 on the video side; Nano Banana Pro,
Seedream 5.0 Lite, Flux.2 Pro, GPT Image on the still side.
Swappable LLM driver — the agent runs on GPT 5.5 Pro, Claude
Sonnet, Claude Opus 4.6, or Gemini 3.1 Pro.
Soul ID is ahead of any open-source pipeline for identity-locked faces
right now. Cinema Studio's camera vocabulary is more granular than a
prompt template you'd write by hand. The rest of this article doesn't
argue with that.
What it does argue with is what the pitch leaves out: what the credits,
the data posture, and the closed pipeline cost you after the first reel
ships.
The five things you actually own when you self-host
Hosted studios optimize for "type a brief, get a video." Ralphy is for
the boring part that comes after. Shipping reels every week. Pulling
up last quarter's scene-04 prompt because the rewrite needs to match
it. Telling the CMO what you spent on regens in Q1 without having to
email someone's support inbox.
Your keys, your rates. Every model call goes direct to OpenRouter
and ElevenLabs at list price. No platform markup, no
credit-to-model lookup table to re-derive every time a vendor moves.
A grep-able generation log. Each project writes
generations.jsonl: prompt, output URI,
latency, cost, model id, append-only. Not a "history" tab in someone
else's dashboard.
No expiring credits. No Ralphy account, so no balance that resets
monthly and no 90-day timer ticking on top-ups you bought last quarter.
Reference fidelity in your repo.ralphy ref
writes structured JSON for master shots, persona descriptors, and
camera grammar. All committed. Re-runnable next year, after the vendor
has changed three model versions.
A deterministic render pass that's just code. The composition is
HyperFrames (HTML + GSAP) or Remotion (React + headless Chromium) —
caption timing, transition curves, music ducking are all TypeScript
you can diff. Want to overlay your own A/B-test telemetry on the
render? It's a <div>.
Music: ElevenLabs MusicGenre + tempo prompt, no licensing chase
✓
✗
Soul ID character consistencyHiggsfield-only; ahead of OSS today
partial
✓
Cinema Studio camera presets70+ motion presets, focal-length controls
✗
✓
Swap models without a releaseEdit MODELS.md, restart
✓
✗
Pricing model
Ralphy
Platform fee$0
Token markup✗
Monthly credits roll overn/a (no credits)
Top-up expiryn/a
Refund windown/a
Prompts used as training data✗
Team / seat fees✗
Higgsfield Supercomputer
Platform fee$15-99/mo, billed annually
Token markup✓
Monthly credits roll over✗
Top-up expiry90 days
Refund window7 days, zero credits used, 6% fee
Prompts used as training data✓
Team / seat fees✓
What a reel actually costs (real OpenRouter prices, pulled 2026-05-26)
An earlier draft of this page claimed Ralphy was cheaper than
Higgsfield by an order of magnitude. That was wrong, and the table
below is the corrected math.
The real story is dull. Ralphy charges you what OpenRouter charges
OpenRouter's direct API users — the same number, no platform fee, no
credit conversion, no expiry timer. The binary just wires the
providers together.
Numbers pulled live from the OpenRouter model pages:
Model
OpenRouter rate
5-second clip
kwaivgi/kling-v3.0-pro
$0.168 / sec
$0.84
google/veo-3.1
$0.40 / sec
$2.00
google/veo-3.1-fast (4K)
$0.10 / sec
$0.50
bytedance/seedance-2.0
$7 / M tokens (W·H·d·24/1024)
~$0.76 (720p) / ~$1.70 (1080p)
openai/gpt-5.4-image-2
$8 in + $15 out / M tokens
~$0.05 / image (1024²)
google/nano-banana-pro
$2 in + $12 out / M tokens
~$0.04 / image
Live OpenRouter pricing · pulled 2026-05-26 · pass-through to Ralphy
A canonical 30-second vertical reel on Ralphy — six 5-second scenes,
six ref frames, one VO, one music bed — costs roughly:
Line item
Cheap mix (veo-fast)
Default mix (kling-pro)
Premium mix (veo-3.1)
6 × video clips (5s each)
~$3.00
~$5.04
~$12.00
6 × reference images
~$0.24
~$0.30
~$0.30
Scenario / scoring LLM
~$0.10
~$0.20
~$0.30
ElevenLabs VO + Music
subscription
subscription
subscription
Render compute
$0
$0
$0
Total OpenRouter pass-through
~$3.34
~$5.54
~$12.60
Per-reel cost at model cost price · 30s reel, 9:16 · ElevenLabs sub (~$22/mo) amortized separately
Default mix lands at $5-6 per reel. The veo-3.1-fast cheap mix
(the only 4K model in the catalog) drops to ~$3.30. Heavy regens,
premium models, or 60-second reels push it up. The cost isn't
estimated; ralphy render writes each call's actual price into
generations.jsonl as the render happens. You can cat it.
The same Kling reel on Higgsfield runs about 7 credits on Plus. With
1,000 credits a month, that's room for ~142 Kling reels. Face-value
math is $0.27 per reel, if you burn every credit. At 10 reels a month
you use 70 of them and the other 930 evaporate at month-end. Effective
cost is $3.90 per reel shipped, and you can't change which video model
that buys you without changing your subscription tier.
At low-to-mid volume the Higgsfield headline number is cheaper. That's
true. It only stays true if you burn every credit before it expires,
which most teams don't. The face-value arithmetic ignores the forfeit,
the markup, and the fact that you can't change models without changing
tiers.
Where Ralphy actually wins is around the invoice, not on it:
Vendor list price, no markup. Whatever OpenRouter charges is the
line item. We keep zero.
Nothing forfeits. Render nothing this month and the bill is
zero. The ~$36 of unused Plus credits that vanish at month-end on
Higgsfield don't exist here.
No vendor lock. Edit one entry in MODELS.md, restart. The
pipeline doesn't care which vendor's weights run the clip. On
Higgsfield the menu changes when Higgsfield re-negotiates it.
A budget your agent enforces. Set a per-reel ceiling in
MODELS.md; the agent refuses to regen past it. The hosted
equivalent is the credit bar going red.
Ralphy is open-source and pluggable. Higgsfield is a closed studio.
Higgsfield is a closed studio. Forty-plus tools, one fixed agent, one
fixed render path, a model menu they curate. You take what the product
team picked. Ralphy is a harness. MIT-licensed code, every provider
call funneled through two small files
(cli/lib/providers/media.ts and
cli/lib/providers/llm.ts), and a rendering layer that's
just HTML or React. Fork it, vendor it into your repo, or rewrite any
piece you don't like.
That difference compounds in four places hosted studios can't follow
you to:
No vendor lock on models. OpenRouter alone fronts ~200 models
across OpenAI, Google, Anthropic, Black Forest Labs, Kwaivgi,
ByteDance, MiniMax, and a dozen smaller labs. Want to swap Kling for
Seedance for Veo for whatever ships next month? Edit one entry in
MODELS.md, restart. The pipeline doesn't care which vendor's
weights run the clip. Higgsfield's menu is fixed at their
negotiation cycle.
Pluggable providers. Want to add an internal eval endpoint, a
private Stable Diffusion finetune, or a custom captions backend?
Write a function. The MIT license means you can fork, modify, ship,
and never owe upstream anything.
Custom hooks and your data. Wire your own brand database, your
A/B-test retention curves, your previous campaign's hook scores into
the scenario prompt. Ralphy reads brand, persona, ref
resources from disk — your files, your schema, your data, never
shipped off-machine.
Postmortem-driven learning. Every finished project can write a
postmortem/ directory with lessons learned, the model swaps that
worked, the prompts that didn't, and a per-reel cost rollup. The
agent reads them when it scopes the next brief. The longer you run
it, the more it knows about your register specifically, not the
median across Higgsfield's customer base.
Hosted platforms can't ship this. The whole pitch for Supercomputer is
"we pick the tools, you don't think about them." A harness makes the
opposite trade. You or your agent make the calls, the tool stays out
of the way, and the choices stack into something portable that's
actually yours.
What Higgsfield Supercomputer actually does better
Some things they ship today that Ralphy genuinely doesn't match.
A web UI a designer can click in. Ralphy is CLI-first; the assumed
driver is a coding agent. Higgsfield ships a clickable stack that
works without a terminal.
Soul ID character consistency. Training a digital double from a few
reference photos and getting consistent facial geometry across angles
and lighting is real, polished work. Ralphy's reference pipeline uses
checked-in master-shot images and --ref on each generate call, which
works but is less plug-and-play.
Cinema Studio's camera language. The combination of camera body, lens,
focal length, aperture, and 70+ named motion presets is more granular
than a prompt template. If your work depends on cinematic camera
grammar and you don't want to write it longhand, Higgsfield gets you
closer faster.
One-pass audio plus video. Higgsfield renders sync'd lipsync, ambient
beds, and music in a single agent call. Ralphy does this through Kling
--audio for English VO and an ElevenLabs overlay in the editor pass,
which is more steps.
Zero setup. No bun, no env vars, no ralphy doctor. If your team
doesn't write code, that gap is the whole conversation.
If those matter more than the audit log, the credit math, and the data
posture, Higgsfield is the right call. This article isn't trying to
talk you out of it.
What "AI-native" looks like in practice
The teams that get the most out of Ralphy share a few habits. None of
them are Ralphy-specific. They're just how AI-native teams operate in
general.
01
Treat prompts like code
Prompts live in git. Ralphy drops prompts.json and
STORYBOARD.md into every project, both committed. Your
prompt history sits next to your application code instead of in a
SaaS dashboard somewhere.
02
Let an agent drive the CLI
The bundled agent reads AGENTS.md, picks the right playbook for
the brief, runs the matching ralphy verbs, reads the postmortems
from previous projects, and writes the logs. The human reviews the
diff and the mp4. Same loop as any pull request.
03
Retrofit, don't replace
Ralphy doesn't replace Figma, Notion, or Linear. It slots in
between the brief and the render. The rest of your stack stays
where it is, and migration happens one project at a time.
If that's the loop you're already in, or where you're heading by Q4,
you don't need a hosted studio. You need a CLI you can pin to a
specific commit and an agent that knows when to refuse to render.
Being able to grep your own prompt history in Slack is a nice bonus.
Recommendations by team type
For engineering-led product teams
Ralphy —
Direct provider access, version-controlled prompts, and an
MIT-licensed binary you can pin to whatever commit you trust. The
CLI is built to be driven by Claude, Cursor, or Codex, so your
code-gen agent and your video-gen agent share the same interface.
For brand / creative teams without engineering support
Higgsfield —
Web UI, polished templates, Soul ID for character consistency, no
terminal anywhere. You give up the audit log, the data posture, and
the ability to pin a model version, but you get a dashboard a
designer can use on day one.
For ai-native marketing teams shipping ≥10 reels/week
Ralphy —
At volume the conversation isn't "Plus or Ultra", it's how many
top-up packs you'll burn through and how many will expire before
you use them. Pass-through pricing plus a per-reel budget the
agent enforces skips the entire credit-arithmetic loop.
For compliance-sensitive industries (finance, health, gov)
Ralphy —
Generations stay on your disk. Prompts are not training data.
Reference imagery, prompts, and rendered outputs never sit in a
third-party studio's S3 bucket. No central account that can be
banned, frozen, or have policy changes applied retroactively.
For creators who depend on Soul ID consistency
Higgsfield —
Honest answer: if your work hinges on locking a specific face
across hundreds of generations, Higgsfield's Soul ID is ahead of
what an open-source pipeline gets you today. Ralphy's ref
primitives are catching up, but they're not there yet for fashion
or talent-driven work.
For single-operator solopreneurs
Either —
At <10 reels a month, the credit constraints are mostly
invisible (you'll use what you bought, or you won't notice the
forfeit). Pick whichever interface matches your day-to-day. Solo
on Claude Code? Ralphy. Solo in a browser? Higgsfield.
Frequently asked questions
Ralphy is an open-source CLI you run locally, calling OpenRouter
and ElevenLabs with your own keys. Higgsfield Supercomputer is a
hosted agentic stack that owns the pipeline (studios, models,
storage, the agent driving it) and bills you in credits. You're
trading a polished web UI and Soul ID for an audit log, version
control, pass-through pricing, and a data posture where your
prompts aren't training material.
If you already treat prompts as code and AI agents as coworkers,
Ralphy is the version of this tooling that fits the rest of your
stack. You give it two API keys, it gives you a harness back. Every
model call lands in a grep-able JSONL log. The rendering layer is
code your agent can edit. Each shipped reel writes a postmortem the
next project gets to read.
Your reels cost what the models cost — that's the whole pricing
pitch. There's no platform to host. There's no account, so credits
can't sit there expiring on you. Your prompts live in a git repo
where you control them, not in someone else's training set. And the
provider layer is two files, so swapping any model for any other is
an afternoon's work rather than a vendor escalation.
Higgsfield's Supercomputer is hosted and closed-source. It plans the brief, picks the model from a menu they curate, runs the render on their servers, and bills you in credits that expire on a timer. Ralphy is the inverse: an MIT-licensed CLI that uses your own OpenRouter and ElevenLabs keys, runs on your machine, and writes every generation to your disk. You pay the model vendor directly, at list price. There's no markup and no platform fee. Want to swap Kling for Seedance or Veo? Edit one line in MODELS.md.
If you ship one reel a quarter and don't write code, a Higgsfield subscription is faster to set up and you'll feel the difference on day one. Add a coding agent or a campaign that runs all year and the constraints flip. Credits forfeit. Prompts get harvested. The model menu is whatever Higgsfield negotiated this quarter. You usually notice around month six.
What Higgsfield Supercomputer is, accurately
Supercomputer launched on May 14, 2026. It's Higgsfield's umbrella product, an agentic stack that wraps their existing studios into one "describe it and ship" loop. The headline pieces:
Soul ID is ahead of any open-source pipeline for identity-locked faces right now. Cinema Studio's camera vocabulary is more granular than a prompt template you'd write by hand. The rest of this article doesn't argue with that.
What it does argue with is what the pitch leaves out: what the credits, the data posture, and the closed pipeline cost you after the first reel ships.
The five things you actually own when you self-host
Hosted studios optimize for "type a brief, get a video." Ralphy is for the boring part that comes after. Shipping reels every week. Pulling up last quarter's scene-04 prompt because the rewrite needs to match it. Telling the CMO what you spent on regens in Q1 without having to email someone's support inbox.
<div>.Ship reels on a cadence.
Two API keys, zero account. ralphy doctor verifies the install in five seconds.
The straight comparison
Same job (vertical reel that doesn't look like AI vomit), two operating models.
Platform & control
Reproducibility and audit
Model coverage
Both stacks ship the same frontier model menu. The difference is the billing path that wraps it.
Pricing model
What a reel actually costs (real OpenRouter prices, pulled 2026-05-26)
An earlier draft of this page claimed Ralphy was cheaper than Higgsfield by an order of magnitude. That was wrong, and the table below is the corrected math.
The real story is dull. Ralphy charges you what OpenRouter charges OpenRouter's direct API users — the same number, no platform fee, no credit conversion, no expiry timer. The binary just wires the providers together.
Numbers pulled live from the OpenRouter model pages:
A canonical 30-second vertical reel on Ralphy — six 5-second scenes, six ref frames, one VO, one music bed — costs roughly:
Default mix lands at $5-6 per reel. The veo-3.1-fast cheap mix (the only 4K model in the catalog) drops to ~$3.30. Heavy regens, premium models, or 60-second reels push it up. The cost isn't estimated;
ralphy renderwrites each call's actual price into generations.jsonl as the render happens. You cancatit.The same Kling reel on Higgsfield runs about 7 credits on Plus. With 1,000 credits a month, that's room for ~142 Kling reels. Face-value math is $0.27 per reel, if you burn every credit. At 10 reels a month you use 70 of them and the other 930 evaporate at month-end. Effective cost is $3.90 per reel shipped, and you can't change which video model that buys you without changing your subscription tier.
At low-to-mid volume the Higgsfield headline number is cheaper. That's true. It only stays true if you burn every credit before it expires, which most teams don't. The face-value arithmetic ignores the forfeit, the markup, and the fact that you can't change models without changing tiers.
Where Ralphy actually wins is around the invoice, not on it:
MODELS.md, restart. The pipeline doesn't care which vendor's weights run the clip. On Higgsfield the menu changes when Higgsfield re-negotiates it.MODELS.md; the agent refuses to regen past it. The hosted equivalent is the credit bar going red.Ralphy is open-source and pluggable. Higgsfield is a closed studio.
Higgsfield is a closed studio. Forty-plus tools, one fixed agent, one fixed render path, a model menu they curate. You take what the product team picked. Ralphy is a harness. MIT-licensed code, every provider call funneled through two small files (cli/lib/providers/media.ts and cli/lib/providers/llm.ts), and a rendering layer that's just HTML or React. Fork it, vendor it into your repo, or rewrite any piece you don't like.
That difference compounds in four places hosted studios can't follow you to:
MODELS.md, restart. The pipeline doesn't care which vendor's weights run the clip. Higgsfield's menu is fixed at their negotiation cycle.brand,persona,refresources from disk — your files, your schema, your data, never shipped off-machine.postmortem/directory with lessons learned, the model swaps that worked, the prompts that didn't, and a per-reel cost rollup. The agent reads them when it scopes the next brief. The longer you run it, the more it knows about your register specifically, not the median across Higgsfield's customer base.Hosted platforms can't ship this. The whole pitch for Supercomputer is "we pick the tools, you don't think about them." A harness makes the opposite trade. You or your agent make the calls, the tool stays out of the way, and the choices stack into something portable that's actually yours.
What Higgsfield Supercomputer actually does better
Some things they ship today that Ralphy genuinely doesn't match.
A web UI a designer can click in. Ralphy is CLI-first; the assumed driver is a coding agent. Higgsfield ships a clickable stack that works without a terminal.
Soul ID character consistency. Training a digital double from a few reference photos and getting consistent facial geometry across angles and lighting is real, polished work. Ralphy's reference pipeline uses checked-in master-shot images and
--refon each generate call, which works but is less plug-and-play.Cinema Studio's camera language. The combination of camera body, lens, focal length, aperture, and 70+ named motion presets is more granular than a prompt template. If your work depends on cinematic camera grammar and you don't want to write it longhand, Higgsfield gets you closer faster.
One-pass audio plus video. Higgsfield renders sync'd lipsync, ambient beds, and music in a single agent call. Ralphy does this through Kling
--audiofor English VO and an ElevenLabs overlay in the editor pass, which is more steps.Zero setup. No
bun, no env vars, noralphy doctor. If your team doesn't write code, that gap is the whole conversation.If those matter more than the audit log, the credit math, and the data posture, Higgsfield is the right call. This article isn't trying to talk you out of it.
What "AI-native" looks like in practice
The teams that get the most out of Ralphy share a few habits. None of them are Ralphy-specific. They're just how AI-native teams operate in general.
Treat prompts like code
Prompts live in git. Ralphy drops prompts.json and STORYBOARD.md into every project, both committed. Your prompt history sits next to your application code instead of in a SaaS dashboard somewhere.
Let an agent drive the CLI
The bundled agent reads
AGENTS.md, picks the right playbook for the brief, runs the matchingralphyverbs, reads the postmortems from previous projects, and writes the logs. The human reviews the diff and the mp4. Same loop as any pull request.Retrofit, don't replace
Ralphy doesn't replace Figma, Notion, or Linear. It slots in between the brief and the render. The rest of your stack stays where it is, and migration happens one project at a time.
If that's the loop you're already in, or where you're heading by Q4, you don't need a hosted studio. You need a CLI you can pin to a specific commit and an agent that knows when to refuse to render. Being able to grep your own prompt history in Slack is a nice bonus.
Recommendations by team type
For engineering-led product teams
Direct provider access, version-controlled prompts, and an MIT-licensed binary you can pin to whatever commit you trust. The CLI is built to be driven by Claude, Cursor, or Codex, so your code-gen agent and your video-gen agent share the same interface.
For brand / creative teams without engineering support
Web UI, polished templates, Soul ID for character consistency, no terminal anywhere. You give up the audit log, the data posture, and the ability to pin a model version, but you get a dashboard a designer can use on day one.
For ai-native marketing teams shipping ≥10 reels/week
At volume the conversation isn't "Plus or Ultra", it's how many top-up packs you'll burn through and how many will expire before you use them. Pass-through pricing plus a per-reel budget the agent enforces skips the entire credit-arithmetic loop.
For compliance-sensitive industries (finance, health, gov)
Generations stay on your disk. Prompts are not training data. Reference imagery, prompts, and rendered outputs never sit in a third-party studio's S3 bucket. No central account that can be banned, frozen, or have policy changes applied retroactively.
For creators who depend on Soul ID consistency
Honest answer: if your work hinges on locking a specific face across hundreds of generations, Higgsfield's Soul ID is ahead of what an open-source pipeline gets you today. Ralphy's
refprimitives are catching up, but they're not there yet for fashion or talent-driven work.For single-operator solopreneurs
At <10 reels a month, the credit constraints are mostly invisible (you'll use what you bought, or you won't notice the forfeit). Pick whichever interface matches your day-to-day. Solo on Claude Code? Ralphy. Solo in a browser? Higgsfield.
Frequently asked questions
Ralphy is an open-source CLI you run locally, calling OpenRouter and ElevenLabs with your own keys. Higgsfield Supercomputer is a hosted agentic stack that owns the pipeline (studios, models, storage, the agent driving it) and bills you in credits. You're trading a polished web UI and Soul ID for an audit log, version control, pass-through pricing, and a data posture where your prompts aren't training material.
Try Ralphy in eight minutes.
brew install alecs5am/tap/ralphy && ralphy new 'vertical reel about <your product>' — cold start to rendered mp4.
If you already treat prompts as code and AI agents as coworkers, Ralphy is the version of this tooling that fits the rest of your stack. You give it two API keys, it gives you a harness back. Every model call lands in a grep-able JSONL log. The rendering layer is code your agent can edit. Each shipped reel writes a postmortem the next project gets to read.
Your reels cost what the models cost — that's the whole pricing pitch. There's no platform to host. There's no account, so credits can't sit there expiring on you. Your prompts live in a git repo where you control them, not in someone else's training set. And the provider layer is two files, so swapping any model for any other is an afternoon's work rather than a vendor escalation.