What we shipped

Kling 4.0 is now available for text-to-video, image-to-video, reference-to-video, keyframes and video editing, alongside a faster Flash variant for quick drafts.

• Up to 30 seconds at 720p or 1080p (Flash tops out at 20s and 720p)
• Multi-shot prompts, up to 7 reference elements and 10 keyframes in one generation
• Start and end frames for image-to-video, native audio on keyframes mode
• Video editing from a source clip guided by a text prompt
• $0.157/s at 720p, $0.235/s at 1080p (Flash is 720p only at $0.157/s)

Flash trades resolution and the end-frame, keyframes and edit modes for speed; use it for iterations, then switch to the full model for the final render.

You can now create editable SVG files in varg, either from a text prompt or by converting an existing image into vector paths.

• Ideogram Vector turns a prompt into an editable SVG at 14 credits (low quality), 17 credits (medium, the default) or 21 credits (high) per SVG. You can add up to 5 reference images, pick from 17 aspect ratios or auto, and get up to 8 SVGs per request. To convert a JPEG, PNG or WebP image to SVG instead, pin fal:ideogram/vector/image-to-vector for 11 credits per image.

High quality with reference images costs 34 credits per SVG, regardless of how many references you add, while low and medium keep their usual price with references.

The varg plugin is now listed in the ChatGPT store. Install it and generate video, images and voice from inside any ChatGPT conversation — the same models and credits as the app.

• Install from the ChatGPT plugin store
• One sign-in, no API key — your varg account carries over
• Works alongside the Claude connector, Cursor and VS Code through MCP

You start a generation in ChatGPT; managing renders and files still happens in the varg dashboard.

Copy for agents works on presets again. It failed with "Failed to copy" from both the bulk bar and the preview window.

• Select any number of presets and copy them in one go; each arrives as its own block
• Each block carries the preset ID and the exact request fragment for every tool it works with
• The same fix covers Files items whose details load after the click

Paste the result straight into Claude Code, Cursor or any other agent. Copying several presets takes a second while their details load.

When you asked the agent what a generation cost, it used to answer in dollars — "~$0.70" — because the API names its amounts after cents and the model read those fields literally. It now answers in credits, the unit you actually pay in.

• Every amount in a tool result (estimates, balances, spend, job costs, error details) reaches the model with its key renamed: `user_price_cents` → `user_price_credits`, `cents_per_second` → `credits_per_second`, `money_unit: "cent"` → `"credit"`
• The number is unchanged — 70 cents is 70 credits — only the label changes
• `check_balance` now tells the model `"unit": "credits"`; the server instructions tell it never to convert credits to money, since what a credit costs depends on your plan
• The widget already spoke credits; it now also accepts the old `*_cents` spelling from older servers, so a mixed fleet keeps rendering

You get prices you can act on. The trade-off: a credit is not a fixed amount of money, so the agent will not tell you what a generation costs in dollars even if you ask — point it at your plan page instead.

Every varg result inside Claude and ChatGPT now renders in a single widget that matches the host's theme — light or dark — instead of the four separate viewers that came before.

• generate_image, generate_video, check_balance and the rest all open the same card
• estimate_cost and poll_job render inside the widget instead of as plain JSON
• upload_file_widget handles multiple files in one view
• the old media-viewer, balance-viewer, analytics-viewer and upload-widget URIs are still registered for hosts that pinned them

You get a calmer, more consistent surface across hosts. The trade-off: the widget is one bundle, so a host that pinned a specific legacy viewer by URI keeps working but no longer gets its own tailored layout.

You can now generate images with accurate Chinese and English text and control where text and other elements sit on the canvas with MOSS Image.

• MOSS Image: text-to-image with Chinese and English text rendering and layout control. You describe the layout in natural language, tag sequences or a JSON layout. Output runs from 512 to 2048 px per side, with up to 4 images per request. A 1024x1024 image costs 7 credits, 1024x576 (16:9) costs 4, 768x1024 (portrait) costs 5, 512x512 costs 3, and 4 square 1024x1024 images cost 23.

Price scales with output size and image count, so 4 images at 2048x2048 cost 89 credits.

varg is now a published plugin in ChatGPT. Add it once, sign in with your varg account, and ask ChatGPT for a video, an image or a voiceover in the same chat.

• One click from varg.ai/mcp or the ChatGPT plugin page — no developer connector, no URL to paste
• Sign-in is a browser approval, not an API key
• Everything you make lands in your varg library and uses the same credits as the app
• Every generation is priced before it runs, so ChatGPT can quote it and wait for a yes

A connected ChatGPT can spend your credits. It cannot create API keys or change account settings, and you can disconnect it any time under Settings → Connected apps.

A new image model joins the catalog. FLUX 3 Image Klein is the lighter FLUX 3 — same family, faster and cheaper per picture, at the cost of some detail on dense prompts.

• 7 credits ($0.07) for a standard 1K image
• Available in the dashboard catalog and at /v2/image
• SDK and curl examples on the models page

Pick it for speed and volume; reach for full FLUX 3 when the prompt is complex or the output is hero material.

ElevenLabs v4 is available on varg. Direct the delivery of any line with inline tags like [whispers] and [shouts], and the model follows them.

• Pick eleven_v4 from the Voice menu in the Playground, or add ?model=eleven_v4 to the URL
• Tags work in the same prompt field as the text, no separate control
• The Home carousel has a new card linking straight to the Playground with the model preselected

v4 is priced higher than v3. Use v3 when you do not need the tag-directed delivery; it stays the default for plain narration.

varg.ai/api is a public page where you can browse every model on the platform before signing up. Open it, see what a model costs and what it can do, and copy a ready cURL, JS or Python call.

• 157 models across video, image, voice and music, with prices in credits
• Featured reels autoplay while in view, so you see real output, not a stock thumbnail
• Tabs for Video, Image, Voice and Toolkit, plus modality chips and sort by Featured, Cheapest or A-Z
• Capability tiles for render, video_analyze, FFmpeg, webhooks, result cache and MCP
• Code samples for Seedance 2.5, Nano Banana 2 and ElevenLabs v3, built from the Playground's own snippet builders

A signed-out visitor who opens /devs now lands here instead of a login wall. Deeper developer pages still ask you to sign in.

Two new speech models from ElevenLabs are available: Eleven v4 and Eleven v4 Turbo. Both accept audio tags — inline directions like [excited], [whispers] or [laughs] — that the model interprets as emotional performance, not as literal text.

- Eleven v4 — #2 on the Artificial Analysis Speech Arena, 90+ languages, voice clones from 10 seconds of audio, professional voice clones return - Eleven v4 Turbo — #1 on the arena, ~100 ms inference latency for real-time use, half the price of v4 - Both support the same voices and the same with-timestamps endpoint as v3 - v4 costs $0.08 per 1,000 characters; Turbo costs $0.04

Use them anywhere you use speech today — the Playground, the API, the SDK. Audio tags are the trade-off: they add expressiveness but make output less predictable, so test a line before sending it to a client.

Seedance 2.5, 2.0 and 2.0 mini now go to PiAPI first, with ByteDance's official API right behind it. Prompts can now be up to 20,000 characters. A prompt longer than PiAPI's 4,000 goes straight to a provider that takes it.

• Default order: PiAPI, then ByteDance, then fal. Seedream 5 pro stays on ByteDance first, then fal
• Seedance 2.5 on PiAPI from 19.2¢/s at 480p, 44.6¢/s at 720p and 101.7¢/s at 1080p
• Seedance 2.0 from 11.6¢/s and 2.0 mini from 7.4¢/s at 480p
• Long prompts: 4,001 to 20,000 characters run on ByteDance or fal. Over 20,000 the request is rejected, and the error lists each provider's limit
• ByteDance price follows what the generation actually used. The estimate is the ceiling: a 4 s 480p Seedance 2.5 clip cost 55¢ against a 62¢ estimate
• To stay on one provider, name it: byteplus:seedance_2_5 or piapi:seedance_2_5

PiAPI costs more than ByteDance did as the default: Seedance 2.5 at 480p is 19.2¢/s instead of 15.6¢/s. In return, reference photos of real people work without extra steps. If a job falls back to another provider, you pay that provider's price, which can be above the estimate you saw.

A video job no longer jumps back to "queued" when a provider refuses it. The status only moves forward, queued then running then done, and a new stage field says what the job is doing right now: submitting, uploading references, rendering, retrying on another provider, or saving the result.

• attempts on every job lists each provider tried, with a short reason and a code for the ones that failed (content_privacy, content_policy, quota, rate_limit, timeout, expired, prompt_too_long, provider_error)
• Reference uploads to ByteDance show their progress, "uploading references to ByteDance (1/2)"
• A BytePlus task that sits in its queue for an hour moves to the next provider, instead of after six
• A job interrupted by a server restart mid-submit is picked up again within about 10 minutes, instead of failing after 7 hours
• Cancelling a job before it starts rendering is always free

stage and attempts are in GET /v2/jobs/:id, /status and your webhook payload. Code that waits for "queued", "submitting" or "running" to end keeps working; "submitting" is no longer sent as a status. The raw provider error is not exposed per attempt, only the reason code; the job's own error still carries the full text when it fails.

You can now outpaint an image to a new shape, either one of 9 preset aspect ratios or a custom canvas, while the original pixels stay at full resolution.

• Bria FIBO-Edit 1.5 Expand: extends one image to a preset aspect ratio such as 16:9 or 9:16, or to a custom canvas up to 25 MP. You can add an optional prompt to describe the new area or leave it out. Each image costs 5 credits, whatever the ratio or canvas size.

Seedance and Seedream now run on ByteDance's official API by default, and Seedance is about 10% cheaper. If ByteDance can't finish a generation, it moves on to the next provider on its own, and your job completes there instead of failing.

• Seedance 2.5 from 15.6¢/s at 480p, 36.4¢/s at 720p, 83.2¢/s at 1080p
• Seedance 2.0 from 9.5¢/s at 480p, plus 4K
• Seedance 2.0 mini from 6.6¢/s at 480p
• Seedream 5 pro at 8¢ per image, the same as before
• Fallback order: ByteDance, then PiAPI, then fal. You pay the price of the provider that delivers, and failed attempts cost nothing
• To keep one provider, name it: byteplus:seedance_2_5 or piapi:seedance_2_5

ByteDance's official API turns down reference photos of real people. With a canonical model name, those jobs go to PiAPI automatically at PiAPI's price. A job that has already started rendering on ByteDance can no longer be cancelled: it finishes and is charged as usual.

You can now remove the background from an image and get back a real alpha channel, with the original pixels untouched. Add a short prompt to add, drop or isolate objects in the cut.

• Bria Guided Remove Background: cuts the subject out of one image, optionally steered by a prompt such as 'only the chair'. It costs a flat 4 credits per image, with or without a prompt, and supports autocrop through provider options.

It works on one image per request.

Krea 3, Krea's newest image model, is available through the API and the SDK as krea_3. Use krea to always get the latest Krea model; today that is Krea 3. It is free while Krea runs its early access.

• Text-to-image and edits on one model: attach up to 8 images to steer the result or to edit one of them
• Pick thinking_mode fast, high or ultra; fast returns a 1K image in about 5 seconds
• 1K or 2K output, 10 aspect ratios from 21:9 to 9:16
• Pass any Krea 3 parameter through provider_options.krea: seed, negative_prompt, exact width and height

One image per request; Krea has dropped batch generation. When early access ends, Krea 3 starts costing credits at Krea's list price.

Seedance 2.5 now renders at the resolution you ask for. 480p and 1080p requests were generated and billed as 720p; they now come out at 480p and 1080p and are priced for what you get.

• Image-to-video with a single image no longer fails with "omni_reference_task_type requires omni_reference mode"
• 21:9 and 1:1 work on the standard tier instead of falling back to 16:9
• Standard seedance-2.5: 18 / 41 / 93 credits per second at 480p / 720p / 1080p (was 33 / 67 at 480p / 720p)
• seedance-2.5-less-restriction: 20 / 45 / 102 credits per second (was 18 / 41 / 93)

The standard tier costs roughly half what it did. The less-restriction tier costs about 10% more than before.

Dress a person in a garment, or put a product in their hand, starting from photos you already have. Each edit costs 5 credits.

• Bria FIBO-Edit 1.5 virtual try-on: one person photo and 1-3 garment images
• Bria FIBO-Edit 1.5 product holding: one person photo and 1-3 product shots, and the person ends up holding the product
• Commercially safe output from Bria's licensed training data

The first file you send is always the person; the garments or products follow it. A text prompt is optional and steers the edit.

You can now describe a piece of music in up to 1,024 characters and get back an instrumental track. Output is instrumental only, with no vocals.

• Mureka 9.5 generates an instrumental track from your text description for 25 credits per request.
• Mureka 9 and Mureka 8 cost 8 credits per request. Select them through provider_options.
• Mureka 7.6 costs 6 credits per request. Select it through provider_options.

MiniMax H3 now costs what fal charges for it: 5 seconds at 720p is 32 credits instead of 137, and the 2K default drops from 137 to 69. New models arrive alongside it.

• MiniMax H3: 480p to 4K, first and last frame, a soundtrack on image-to-video, image, video and audio references
• MiniMax H3 Max: a 1080p tier and any length from 0.92 to 15 s, plus new tools on the same model: relight, recast, insert, extend (also on Turbo), 3D to video, camera controls, lip sync, five style presets, and LoRA on H3
• Vidu Q4: 3-16 s with native audio from a first frame, or from up to 12 reference images and 3 voice clips, 540p to 4K
• Seedream 5 Flash: fast, low-cost Seedream 5.0 text-to-image and edit with up to 10 references
• Recraft V4.1 Flash: 2 credits per image, with palette and background colour control
• Creatify Boreal: product and presenter videos with synced speech, 1-20 s
• PixVerse VibeMV: a music video for a whole song, 10 s to 6 min, 15 style presets
• ID-V2V and ID-V2V Relight: restyle or relight a video from one edited frame, keeping identity, expressions and motion
• Pixelcut looping video: a product photo into a seamless 5-15 s loop

Models priced by output length (VibeMV, ID-V2V, and H3 Max relight, recast and lip sync) reserve the longest possible output and charge the real length when the job finishes. Vidu Q4 has no text-only mode: give it an image or a reference.

The sidebar is now grouped by what you came to do, and the API has its own page in the app.

• Pinned at the top: Home, Agent and All projects, the page most of you open first
• Create: Video, Image, Voice, Canvas, Templates
• Build: All models, MCP, API and Docs side by side
• Library: Files (formerly Resources), Shared with me, Analytics
• Resources: What's new, Academy, Pricing, folded by default

The new API page has the base URL, a three-step curl quickstart and your API keys in one place. You create, rename and revoke keys there, the same keys as in Settings.

Blog, SDK, Support, Discord and the rest of the site are in a footer at the bottom of Home, Projects, Templates and the other browsing pages. It stays out of the Agent, Playground, Canvas and Files.

The groups remember being folded under new names, so any group you had folded opens once and waits for you to fold it again.

Home is now where you start: what's new at the top, the agent composer right under it, then your recent projects and templates. Type a brief there and the agent opens a new project with it.

• What's new — new models and features as a row of cards that moves on every 6 seconds and stops while your pointer is on it
• Try an example — five ready briefs under the composer, each with its price in credits before you run it
• All projects — one click from the top of the sidebar, right under Agent
• Folding sidebar — Main, Workspace, Learn, Projects and Shared folders each fold, and stay the way you left them in this browser
• Templates — one card per template, with titles that no longer run into the next card

The sidebar remembers its folded groups per browser, not per account, so a second laptop starts with everything open.

Google and OpenAI are shutting down several model versions. The model names you call keep working, now on new routes. Only Sora 2 is gone.

• Veo 3.1, Fast and Lite — Google retires the preview versions on October 22. veo_3_1, veo_3_1_fast and veo_3_1_lite now run through fal at the same prices, except Veo 3.1 Fast with sound: about 16 cents a second instead of 11 to 13. New on all three: first and last frame, image to video on Lite, and audio: false for a cheaper silent clip
• Omni Flash — omni_flash now runs Gemini Omni Flash 1.1, the version Google keeps. Same price at 720p, no 8-second minimum, and 360p to 4K are available
• GPT Image 1 — OpenAI retires it on October 23. gpt_image_1 now runs GPT Image 2.5 Flare at the same price per quality tier; edits cost a cent more
• Sora 2 — OpenAI shut it down on September 24. sora_2, sora_2_pro and sora_2_remix now return model_not_found right away instead of failing after you start a job

Options you passed in provider_options.google or provider_options.openai for these models no longer apply. Their equivalents now go in provider_options.fal.

Nineteen new models are in Tools and the API, across video, image, speech and music. The headline ones: Wan 3.0, the top-ranked text-to-video model; Nano Banana 2.1, added the day it came out; and ElevenLabs v4, the top-ranked voice model.

• Video — wan_3 (up to 30 seconds at 1080p with sound, first and last frame, up to 10 reference images), omni_flash_1_1 (Gemini Omni Flash 1.1, from 360p drafts to 4K, plus video edit), grok_imagine_1_5 and grok_imagine_1_5_lite (Lite from about 2 cents a second with sound), minimax_h3_max_turbo (close to H3 Max quality at half the price) and heygen_video_1 (HeyGen's own video model, up to 2K)
• Image — nano_banana_2_1, grok_imagine_image_2, muse_image (Meta, $0.02 per image), qwen_image_3 (dense text in 12 languages), and krea_2, krea_2_medium and krea_2_medium_turbo, where attached images become style references
• Speech — eleven_v4 and eleven_v4_turbo with audio tags and your cloned voices, and gemini_3_8_flash_tts and gemini_3_8_flash_lite_tts with 130 languages and two-speaker dialogue
• Music — lyria_3_5 for full songs up to about 3 minutes with vocals ($0.11 per track), and eleven_music_v2_5 for tracks up to 10 minutes

Shared fields map where a model supports them: aspect_ratio, resolution, duration, audio and files. Anything model-specific goes in provider_options. Not every model takes every field: Lyria 3.5 sets its length from the prompt, not from duration.

Ideogram 4.5 is now in Tools and the API as ideogram_v4_5. Use it when the words in the image have to be right: posters, logos, signs, packaging, social cards.

• Text to image — low $0.04, medium $0.07 (default) or high $0.24 per image, up to 8 per request. Size never changes the price
• Editing — attach an image and the same model edits it: up to 4 reference images, an optional mask (black is edited, white is kept), and a high-precision mode that restores every pixel you did not ask to change
• Very low quality for edits at $0.02 per image, for fast drafts
• Quality, mask and precision go in provider_options.fal (quality, mask_url, edit_precision)

aspect_ratio maps to 1:1, 4:3, 3:4, 16:9 and 9:16. Other ratios need one of Ideogram's exact sizes in provider_options.fal.image_size, and masked edits always keep the source size.

Team analytics now shows every creator in the "Where it went" breakdown, not just your own.

• Team members with read or write access see all people, personal API keys and folders
• Previously, non-owners saw only their own rows; everyone else collapsed into "Other creators"
• Totals were always correct — only the per-person split was hidden

The owner-only restriction was a leftover from when team management was owner-only. Any team member can now see who spent what.

A new Analytics tab shows what you spent, what you made, and who made it — for your personal account and for any team.

• Headline spend vs the previous period, with a chart split by capability or by creator
• A 12-month calendar of every file you generated, with streaks and your busiest day
• "Where it went" by model, capability, creator, folder, source and job status — click a row to filter the whole page
• The charges or runs behind every number, with a run's cost explained
• Range presets, filters in the URL, CSV export, and "copy as API request"
• Credits only, no dollar amounts

The same numbers are available through the API at /v2/analytics and from an MCP client, so the dashboard, a curl and an agent agree. Attribution follows the root of a job tree, so the generations inside a render read as yours, not as "internal".

Starting several generations at once could show "You have no credits left" while your balance was fine. Each running job holds its estimate until it finishes, so a burst of ten could use up the free balance for a few seconds. The card treated that as an empty account and offered a plan.

• The card now says what is true: how many credits are held, by how many running jobs, and how many are free right now
• Retry runs the same job again once the holds clear, no new request needed
• "Not enough credits" with a plan appears only when the credits are actually gone
• Approving a paused pipeline step gets the same treatment: "Approve again" when credits are held, a plan when they are not

A retry can still hit the same wall if the other jobs have not finished yet; the card tells you the numbers each time rather than guessing.

The cost approval card now checks your available credits, not your total. Credits held for a job that is still running no longer count as spendable, so the card and the API agree on whether you can afford the next generation.

• Billing shows an On hold line whenever credits are reserved for running jobs, and says they come back when the job finishes
• The approval card reads Available instead of Balance while something is on hold, and the insufficient-credits message says how many are held
• Approve waits until your balance has loaded instead of assuming it is enough
• Holds left behind by paused multi-step jobs are settled: completed steps are charged, the rest is returned

Before this, a job that had just started could leave you looking at "enough credits" in the app and an insufficient-credits error from the API. The number you see is now the number you can spend.

Multi-step pipelines (Seedance Loop, Loop 2.5, Long extend) now bill one step at a time. When the first step finishes and the job pauses for your review, you are charged for that step only and the rest of the hold goes back to your balance — nothing is frozen while you decide, even if you take days.

• The approval card shows the price per step, not just the total: First generation 210, Extend 389
• The quote reads the real length of your reference video, also for clips pasted as a URL. A 5 s reference is quoted at ~600 credits, not the old 16 s ceiling of 1,596
• The step card shows what you were actually charged, markup included, so the number matches your balance
• Reject after a step and you pay for that step only. If the second step fails, the first stays paid and nothing more is taken

The one new thing to know: Approve re-reserves the remaining steps. If your balance no longer covers them you get a clear message and a Choose a plan button; the job stays paused, so top up and approve again.

Approving several cost quotes in a row no longer breaks the chat. Click Approve on every card you want; the clicks are collected and sent as one run, and the thread stays where it is instead of jumping to each new message.

• A multi-step pipeline (loop_extend, long_extend) parked for your review is never re-run on its own. The unrequested step-1 regenerations some of you saw on 20 September came from the server mistaking a paused job for a dead worker; it now leaves paused jobs alone.
• Approving the same quote twice — a lost connection, a second tab, a click after an error — submits one job, not two.
• Approve while a run is still streaming and the card queues behind it, with Undo until it goes out.
• A momentary auth hiccup no longer sends you to the login page mid-batch.

One approve-turn still submits its jobs one at a time, so a batch of ten takes a few seconds longer than a single one.

One Approve now means one job. Approving a cost estimate used to occasionally launch the same generation several times when the agent decided to "continue the batch" in the same reply, and every copy was billed. That is fixed: an approval is spent exactly once, and any extra call the agent makes on that turn does nothing.

• The estimate card keeps its state after a page reload — Approved stays approved, a failed submit shows the error instead of offering Approve again
• The Approve message shows a thumbnail of what you approved; once the job lands, clicking it jumps to the node on the canvas
• A request the API rejects (missing prompt, unknown model) comes back to the agent as an error to fix, not as a card with no price

Approvals in chats from before this release may still show an Approve button after a reload when several estimates were pending at once; pressing it will not launch a duplicate.

You can now play YouTube previews directly in What's new. A preview starts only after you click it, so the card never starts sound on its own; closing the card stops playback.

The card now shows media previews and limits announcements to the last seven days. Older releases remain available on /changelog.

Image generation now defaults to GPT Image 2.5, OpenAI's latest image model. Every image the agent generates without an explicit model pick goes through GPT Image 2.5 via fal's Flare endpoint.

• Same model handles text-to-image and image editing — pass a reference image and the API routes to the edit endpoint automatically
• Flare is 50% lower latency than GPT Image 2.0, with the same token pricing
• Quality tiers from auto to max, 1-4 images per call, aspect ratios from 1:1 to 9:16
• nano-banana-2, flux-schnell and other models are still available — specify them explicitly when you need them

The trade-off is cost: GPT Image 2.5 runs ~21 credits per image at default quality, where nano-banana-2 was 5. The quality jump is worth it for most use cases; when you need fast iterations or cheaper batches, pin a lighter model.

Attach a video — or paste a link to one — and ask what happens in it. The chat agent now watches the footage itself: scenes, motion, speech and on-screen text, not just the file name. Chat runs on Gemini 3.8 Flash, which reads video natively.

• Video attachments up to 100MB go straight to the model in the conversation
• Videos referenced by URL get a structured report — summary, scene timeline with timestamps, speech transcript, on-screen text — shown as a card you can expand
• The same analysis is available in the API: POST /v2/video/analyze. One analysis costs 6 cents; repeating it on the same video is cached and free
• While the agent thinks, its reasoning now streams live instead of a silent pause

The report treats the video as data: anything inside the footage that looks like an instruction to an AI is flagged, never followed. Videos over 100MB are referenced by URL only.

You can now generate and edit images with native OpenAI GPT Image models through the varg API.

• Models: GPT Image 1, 1.5, 2, 2.5 Flare, 2.5 Sunburst, GPT Image 1 Mini and ChatGPT Image Latest
• Inputs: text prompts, reference images and optional masks
• Outputs: one PNG, JPEG or WebP image per job, with quality and size controls

Native OpenAI requests give you direct model behaviour and image-editing controls. They are metered per output image, so higher quality settings cost more.

GPT Image 2.5 from OpenAI is now available for text-to-image and image editing.

• Flare is the default — fast, high-quality, with transparent background support and up to 4 images per request
• Sunburst is the premium tier — tighter edit control for production-ready creative work, at the cost of longer generation times
• Image editing accepts up to 16 reference images with optional masking
• Quality tiers from low to max let you trade detail and latency for cost

Both variants are live now. Use the model picker or POST /v2/image with model "gpt_image_2_5". Sunburst is reachable by pinning the exact model key when you need its extra precision.

Video walkthroughs are now built into varg. The first one takes a one-line prompt to a finished, rendered video using Claude over MCP — no editor, no timeline, no local setup.

• Find it under Academy in the sidebar
• Also public at varg.ai/academy, so you can send it to someone before they sign up
• New lessons land here as we record them

Alongside it, this feed. Releases now show up in the corner of the dashboard when there is something you have not seen, and each one links to the thing it is talking about.

A new Agent tab makes starting a project instant. callTool is now the default generation path: you describe what you want in plain language, the agent picks the model, assembles the pipeline and runs it. renderVideo is the exception, not the rule.

• Claude Sonnet 5 drives adaptive thinking behind the scenes
• Native audio from the video model takes priority over a separate TTS step, so sound is baked in
• The agent chains image, video, speech and render on its own, through one universal entry point
• Custom team-scoped pipelines you have already defined get picked up automatically

Type a sentence, get a finished video. No model picker, no parameter tuning, no pipeline wiring. You do give up manual control over every micro-decision; when you want it back, define a custom pipeline and let the agent fill in the content.

The preset catalogue is a curated, synced library of reusable parameter bundles, so models produce better results without you hunting down vendor-specific IDs. Every preset is typed:

• Voices — 434 ElevenLabs voices and 187 styles, tagged by gender, accent, tone and use case
• Styles — Higgsfield motion styles and Recraft image styles, tagged by era and visual category
• Avatars — HeyGen avatars with a paired avatar_id and voice_id
• Characters — reference-image presets whose image lands in the request's files array, so the same face appears in every scene

A recommendation endpoint takes natural-language intent — "deep male voice for narration", "cinematic 1980s style" — and returns ranked presets with a confidence score and a readable explanation of the match. The Presets tab adds type sub-tabs, a preview modal and a cached proxy for instant playback.

The API merges preset parameters into the right field paths through mapping rules, then validates the merged body against the target model's schema. If a model does not accept a preset you get a clear rejection instead of a billable failure that silently ignored your choice. Raw IDs still work — presets fill gaps, they do not fence you in.

The entire marketing site moved into the app. varg.ai is now one Next.js application serving both the landing page and the dashboard.

• Hover-to-play video wall with up to 18 autoplaying clips
• Pinterest-style masonry layout with a view toggle and duration badges
• Featured-templates flag with an admin toggle
• Auth-aware header: a CTA when logged out, credits and an account switcher when logged in
• Full SEO template system with dropdown nav, announcement pill and a column footer
• The mobile blocker is gone — the full product works on a phone

Open varg.ai and 18 living videos play immediately: no clicks, no "watch demo" button, no signup wall between a visitor and the product. The old standalone marketing repo is frozen and kept as a rollback — one deploy, one domain, one codebase to maintain.

Every generation now shows its full cost breakdown before you commit a single credit. A quote card appears before render submission, a cost badge sits on every media card, and a tooltip breaks each child generation down by type. Job-tree totals roll the whole pipeline up, so you see what the entire thing costs — not just the final stitch.

• Quote card before submission, cost badge on every media card
• Per-child cost breakdown by generation type in the tooltip
• Job-tree totals for the whole pipeline, not just the last step
• A Default/Bypass toggle: approve every action on sensitive work, or let trusted pipelines run hands-off

The quote is computed inline, in the same flow as the generation. No separate pricing page, no calculator, no guessing — and no surprise on the invoice.

Three major model drops in one month, all behind a single API key.

• FLUX 3 and FLUX 3 Draft (BFL) — text-to-video, image-to-video and video-to-video, plus a cheap draft mode for iteration
• Seedance 2.5 — the first featured model in the catalogue, prioritised in the system prompt
• MiniMax H3 Max — t2v, i2v and r2v (reference-to-video) on fal

A unified resolution field landed the same week with per-provider mapping, including proper Kling 4K routing: set resolution to "4k" and the API routes and prices it correctly for each provider. Model IDs are canonical underscore-normalised across the whole stack, and pricing now shows the minimum ("from N credits") instead of the worst case.

Zero integration time — the models are already in the catalogue, already routed, already priced. Point an existing request at the new model ID and it works.