Pricing update - 24 August 2026
We are making three changes to how generations are counted. Together they mean more models to choose from, and lower cost for typical workloads.
A lot more models. We have added models from a wider set of providers, including several that perform close to the frontier at a fraction of the cost.
More granular rates for every model. Rates now follow what each model costs to run, rather than a flat charge per image or a rounded whole generation. That gives you more control over what a piece of work costs.
New agentic pricing for text models. A single per-message rate no longer describes what Scribble does. Text work is now priced across seven size buckets, so a quick edit costs a fraction of a generation and a full campaign brief is priced for what it actually involves.
Overview
In most cases, a generation is every time you interact with an AI model to generate something. Knowing how many generations you have used matters, because most subscriptions include a set number of them.
How you trigger a generation does not change what it costs. You might prompt a model directly, or an agent might generate on your behalf as part of a larger task. Either way, the rules below apply.
Where a cost is a fraction, it is rounded up to one decimal place. Small pieces of work cost a fraction of a generation, not a whole one.
Below are specific details pertaining to:
Agentic & Text Work
Example Agentic Brief Costs
Image Generation
Video Generation
How the Models Compare
Feed Variations
AI Voiceovers
Scores
Agentic & Text Work
When Pencil’s AI does work for you with a text model — a chat turn with Scribble, a step in a workflow, or an editor action — the amount of work involved can vary a lot. A quick copy tweak might take one small model call. A full campaign brief might take dozens of large ones.
To keep pricing fair, each unit of work is priced by how much work was actually done. After it completes, it is placed into one of seven size buckets, from XS to XXXL, based on two things: how many AI model calls it made, and how much information (tokens) those calls processed. The bucket takes the higher of the two.
The price also depends on the AI model you choose.
Size buckets
Bucket | AI model calls | Total tokens processed |
XS | ≤1 | <50K |
S | 2–3 | 50–300K |
M | 4–10 | 300K–1M |
L | 11–30 | 1–3M |
XL | 31–100 | 3–10M |
XXL | 101–300 | 10–30M |
XXXL | >300 | >30M |
Note: the bucket is whichever tier is higher. Work that makes 5 calls (M) but processes 2M tokens (L) lands in L.
Generations per unit of work, by model and bucket
Model name | XS | S | M | L | XL | XXL | XXXL |
Anthropic: 9 models | 0.4 | 6.9 | 14.2 | 34.5 | 63.3 | 138.0 | 299.0 |
Fable 5 | 0.7 | 13.3 | 27.2 | 66.4 | 121.8 | 265.6 | 575.4 |
Claude Opus 5 | 0.6 | 10.6 | 21.6 | 52.7 | 96.5 | 210.5 | 456.0 |
Claude Opus 4.8 | 0.5 | 9.3 | 19.0 | 46.3 | 84.9 | 185.2 | 401.2 |
Claude Opus 4.6 | 0.5 | 9.3 | 19.0 | 46.3 | 84.9 | 185.2 | 401.2 |
Claude Opus 4.7 | 0.4 | 7.7 | 15.7 | 38.2 | 70.1 | 152.8 | 331.0 |
Claude Sonnet 4.6 | 0.3 | 4.1 | 8.3 | 20.2 | 37.0 | 80.7 | 174.7 |
Claude Sonnet 4.5 | 0.3 | 4.1 | 8.3 | 20.2 | 37.0 | 80.7 | 174.7 |
Claude Sonnet 5 | 0.2 | 2.7 | 5.5 | 13.5 | 24.7 | 53.8 | 116.5 |
Claude Haiku 4.5 | 0.1 | 1.4 | 2.9 | 7.0 | 12.8 | 27.8 | 60.3 |
OpenAI: 10 models | 0.1 | 2.0 | 4.1 | 10.0 | 18.2 | 39.7 | 85.9 |
GPT-5.4 | 0.2 | 3.9 | 7.9 | 19.1 | 35.1 | 76.4 | 165.6 |
GPT-5.6 Sol | 0.2 | 3.3 | 6.6 | 16.2 | 29.6 | 64.6 | 139.8 |
GPT-5.5 | 0.2 | 3.3 | 6.6 | 16.2 | 29.6 | 64.6 | 139.8 |
GPT-5.2 | 0.2 | 2.7 | 5.4 | 13.2 | 24.1 | 52.6 | 113.9 |
GPT-5.3 codex | 0.2 | 2.7 | 5.4 | 13.2 | 24.1 | 52.6 | 113.9 |
GPT-5.6 Terra | 0.1 | 1.7 | 3.3 | 8.1 | 14.8 | 32.3 | 69.9 |
GPT-5.4 mini | 0.1 | 1.1 | 2.3 | 5.5 | 10.1 | 21.9 | 47.5 |
GPT-5.1 | 0.1 | 1.1 | 2.2 | 5.4 | 9.9 | 21.5 | 46.5 |
GPT-5.4 nano | 0.1 | 0.4 | 0.7 | 1.7 | 3.1 | 6.8 | 14.6 |
GPT-5.6 Luna | 0.1 | 0.2 | 0.4 | 0.9 | 1.5 | 3.3 | 7.0 |
Zhipu: 1 model | 0.1 | 1.2 | 2.3 | 5.6 | 10.3 | 22.4 | 48.4 |
GLM-5.2 | 0.1 | 1.2 | 2.3 | 5.6 | 10.3 | 22.4 | 48.4 |
Google: 9 models | 0.1 | 1.2 | 2.3 | 5.6 | 10.2 | 22.2 | 48.1 |
Gemini 3.1 Pro Preview | 0.2 | 2.7 | 5.5 | 13.3 | 24.3 | 52.9 | 114.5 |
Gemini 3.5 Flash | 0.2 | 2.2 | 4.4 | 10.7 | 19.5 | 42.5 | 92.0 |
Gemini 2.5 Pro TTS | 0.1 | 1.9 | 3.9 | 9.5 | 17.4 | 38.0 | 82.2 |
Gemini 2.5 Pro | 0.1 | 1.2 | 2.5 | 6.0 | 10.9 | 23.7 | 51.2 |
Gemini 2.5 Flash | 0.1 | 0.9 | 1.9 | 4.5 | 8.3 | 17.9 | 38.8 |
Gemini 3.6 Flash | 0.1 | 0.6 | 1.3 | 3.0 | 5.5 | 12.0 | 25.9 |
Gemini 3 Flash | 0.1 | 0.5 | 0.9 | 2.1 | 3.7 | 8.1 | 17.5 |
Gemini 2.5 Flash Lite | 0.1 | 0.2 | 0.4 | 1.0 | 1.7 | 3.7 | 8.0 |
Gemini 3.1 Flash Lite | 0.1 | 0.1 | 0.2 | 0.4 | 0.6 | 1.3 | 2.7 |
DeepSeek: 1 model | 0.1 | 1.1 | 2.1 | 5.1 | 9.2 | 20.1 | 43.5 |
DeepSeek V4 Pro | 0.1 | 1.1 | 2.1 | 5.1 | 9.2 | 20.1 | 43.5 |
Moonshot: 1 model | 0.1 | 0.8 | 1.6 | 3.9 | 7.1 | 15.3 | 33.2 |
Kimi K2.6 | 0.1 | 0.8 | 1.6 | 3.9 | 7.1 | 15.3 | 33.2 |
Alibaba: 1 model | 0.1 | 0.4 | 0.7 | 1.7 | 3.0 | 6.5 | 14.0 |
Qwen 3.6 Plus | 0.1 | 0.4 | 0.7 | 1.7 | 3.0 | 6.5 | 14.0 |
MiniMax: 3 models | 0.1 | 0.2 | 0.4 | 0.9 | 1.7 | 3.6 | 7.8 |
MiniMax M2.7-highspeed | 0.1 | 0.3 | 0.6 | 1.4 | 2.5 | 5.3 | 11.5 |
MiniMax M3 | 0.1 | 0.2 | 0.3 | 0.7 | 1.3 | 2.8 | 5.9 |
MiniMax M2.7 | 0.1 | 0.2 | 0.3 | 0.7 | 1.3 | 2.8 | 5.9 |
Average across all 35 models | 0.2 | 2.8 | 5.6 | 13.7 | 25.1 | 54.7 | 118.4 |
Note: provider rows show the average across that provider’s models. After any turn or workflow step completes, open its cost menu (alongside content provenance) to see the bucket, the generations charged, and the models used. You will not be charged for failed generations.
Example Agentic Brief Costs
These examples show what typical briefs cost end to end, using GPT-5.5 for the agentic work plus media generations at the standard rates below. Your actual cost depends on the models chosen and the size of the work.
Brief type | Typical bucket | Agentic work (gens) | Media gens | Total (gens) | Per asset (gens) |
Copy & messaging refresh | S | 3.3 | — | 3.3 | 0.4 |
Format adaptation / resize | M | 6.6 | — | 6.6 | 0.9 |
New static concepts | L | 16.2 | 19.2 | 35.4 | 4.5 |
Video cutdowns | L | 16.2 | — | 16.2 | 5.4 |
Short film | L | 16.2 | 48.0 | 64.2 | 64.2 |
Full campaign pack | XL | 29.6 | 19.2 | 48.8 | 4.1 |
Average across these six briefs | — | 14.7 | 14.4 | 29.1 | 13.2 |
Note: media assumes Nano Banana Pro for stills and Google Veo 3.1 for video — the highest-rated model in each case, and also the most expensive. Your model choice moves these totals a long way: the same 6-second film generated on Kling AI VIDEO 2.5 Turbo costs 8.4 media generations rather than 48.0, taking the brief from 64.2 to 24.6. Adapting, resizing or cutting down an existing master does not generate new media, so those briefs carry no media generations — only the agentic work is charged.
Image Generation
Image generation is priced per image, at a rate set by the model you choose. Every image generated counts, whatever its resolution or quality.
Model Name | Generations |
|
Adobe: 2 models | 3.2 | / image |
Adobe Firefly 5 | 3.2 | / image |
Adobe Firefly V4 Standard | 3.2 | / image |
Getty: 1 model | 2.1 | / image |
Getty Images (Generative) | 2.1 | / image |
OpenAI: 2 models | 1.1 | / image |
OpenAI GPT Image 1.5 | 1.1 | / image |
OpenAI GPT Image 2 | 1.0 | / image |
Google: 8 models | 0.7 | / image |
Nano Banana Pro (Gemini 3 Pro Image) | 2.4 | / image |
Nano Banana 2 (Gemini 3.1 Flash Image) | 0.7 | / image |
Google Imagen 4 Ultra | 0.6 | / image |
Imagen 4 Standard | 0.4 | / image |
Nano Banana (Gemini 2.5 Flash Image) | 0.4 | / image |
Nano Banana 2 Lite | 0.4 | / image |
Google Imagen 3 | 0.3 | / image |
Imagen 4 Fast | 0.2 | / image |
Stability: 2 models | 0.5 | / image |
Stable Diffusion 3 / 3.5 Large | 0.7 | / image |
Stable Diffusion (SDXL / Core) | 0.3 | / image |
ByteDance: 3 models | 0.5 | / image |
Seedream V5 Pro | 0.7 | / image |
Seedream 4.5 | 0.4 | / image |
Seedream V5 Lite | 0.4 | / image |
xAI: 2 models | 0.4 | / image |
Grok Imagine Image v2.0 | 0.6 | / image |
Grok Imagine Image | 0.2 | / image |
Bria: 1 model | 0.4 | / image |
Bria 3.2 | 0.4 | / image |
Alibaba: 2 models | 0.4 | / image |
Qwen Image 2 | 0.4 | / image |
Qwen Image 3 | 0.3 | / image |
Black Forest Labs: 1 model | 0.3 | / image |
Flux 2 Pro | 0.3 | / image |
Kuaishou: 4 models | 0.3 | / image |
Kling AI IMAGE 2 | 0.3 | / image |
Kling AI IMAGE 2.1 | 0.3 | / image |
Kling AI IMAGE 3.0 | 0.3 | / image |
Kling AI IMAGE 3.0 Omni | 0.3 | / image |
MiniMax: 1 model | 0.1 | / image |
MiniMax image-01 | 0.1 | / image |
Average across all 29 models | 0.8 | / image |
Note: post generation actions will have separate cost, e.g. 1 edit = 1 generation on that model’s rate.
Video Generation
Video generation is priced per second of generated video, at a rate set by the model you choose, but with an important rule:
Where the cost is a fraction, the generations are rounded up to one decimal place.
Model Name | Generations |
|
Adobe: 1 model | 5.0 | / sec |
Adobe Firefly Video (Video1 Standard) | 5.0 | / sec |
ByteDance: 4 models | 3.7 | / sec |
Seedance 2.0 | 6.5 | / sec |
Seedance 2.5 | 4.5 | / sec |
Seedance 2.0 Fast | 2.3 | / sec |
Seedance 2.0 Mini | 1.5 | / sec |
OpenAI: 2 models | 3.0 | / sec |
Sora 2 Pro | 5.0 | / sec |
Sora 2 | 1.0 | / sec |
Black Forest Labs: 1 model | 2.8 | / sec |
FLUX 3 | 2.8 | / sec |
Alibaba: 1 model | 2.7 | / sec |
Happy Horse | 2.7 | / sec |
Stability: 1 model | 2.0 | / sec |
Stable Video (image-to-video) | 2.0 | / sec |
Google: 4 models | 1.9 | / sec |
Google Veo 3.1 | 4.0 | / sec |
Google Veo 3.1 Fast | 1.5 | / sec |
Gemini Omni Video (Omni Flash) | 1.0 | / sec |
Google Veo 3.1 Lite | 0.8 | / sec |
Kuaishou: 4 models | 1.6 | / sec |
Kling AI VIDEO 2.1 | 2.8 | / sec |
Kling AI VIDEO 3.0 | 1.4 | / sec |
Kling AI VIDEO 3.0 Omni | 1.4 | / sec |
Kling AI VIDEO 2.5 Turbo | 0.7 | / sec |
xAI: 2 models | 1.6 | / sec |
Grok Imagine Video v1.5 | 2.4 | / sec |
Grok Imagine Video | 0.7 | / sec |
Runway: 4 models | 1.3 | / sec |
Runway Aleph 2.0 | 2.8 | / sec |
Runway Gen-4.5 | 1.2 | / sec |
Runway Gen-3 Alpha Turbo | 0.5 | / sec |
Runway Gen-4 Turbo | 0.5 | / sec |
Lightricks: 2 models | 1.3 | / sec |
LTX 2.5 | 1.7 | / sec |
LTX 2.3 | 0.8 | / sec |
MiniMax: 1 model | 1.0 | / sec |
MiniMax-H3 | 1.0 | / sec |
Average across all 27 models | 2.2 | / sec |
Examples:
A 5-second Runway Gen-4 Turbo output = 5 × 0.5 = 2.5 generations
A 3-second Google Veo 3.1 output = 3 × 4.0 = 12.0 generations
A 10-second Gemini Omni Video output = 10 × 1.0 = 10.0 generations
Note: all video rates above are for 1080p output, except Gemini Omni Video, which generates at 720p and has no 1080p mode. Where a model publishes several resolutions we show the 1080p tier, or the nearest tier the vendor publishes, so the rates are comparable model to model. Generating at 720p costs less on most models and at 4K costs more; the cost menu on any completed job shows the rate that was actually applied.
Note: you will not be charged for failed generations. Kling models must be enabled by a Workspace Admin before use. Gemini Omni Flash is in preview from Google, so quality and behaviour may change.
Priority Capacity
Models listed in the platform as [Model name] [Priority Capacity] run on throughput we reserve with the provider in advance. They return at lower latency and stay available during peak demand periods, when shared capacity can queue. There is nothing to reserve or arrange: pick the Priority Capacity version from the model list whenever you want it, and the standard version whenever you do not. It costs one extra generation per unit.
Model | Standard | Priority Capacity | Unit |
Charged per image |
|
|
|
Adobe Firefly 5 | 3.2 | 4.2 | / image |
Nano Banana Pro (Gemini 3 Pro Image) | 2.4 | 3.4 | / image |
Nano Banana 2 (Gemini 3.1 Flash Image) | 0.7 | 1.7 | / image |
Nano Banana 2 Lite | 0.4 | 1.4 | / image |
Bria 3.2 | 0.4 | 1.4 | / image |
Charged per second |
|
|
|
Google Veo 3.1 | 4.0 | 5.0 | / sec |
Topaz AI Video Enhance | 1.6 | 2.6 | / sec |
Google Veo 3.1 Fast | 1.5 | 2.5 | / sec |
Runway Gen-4.5 | 1.2 | 2.2 | / sec |
Note: Priority Capacity can be chosen per job and dropped again on the next one. There is no commitment, no lead time and no minimum. Every other model runs on shared capacity, which is the default and is priced as published. Available on the models listed here, which cover Google video and image, Adobe Firefly, Runway, Topaz and Bria.
How the Models Compare
Rates differ a great deal between models, and so does output quality. These charts put the two together: how many generations a model charges against its standing on the public Arena leaderboards, which are decided by blind head-to-head votes from the community. Cheaper is to the left, better is higher, so the best value sits toward the top left.
How to read the line graphs: the dashed line joins the best-quality model available at each price, so anything sitting below it is beaten on both counts by something cheaper. A triangle means fewer than 5,000 votes have been cast on that model, so its position is less settled.
Note: the three Arena boards are scored independently, so an Elo of 1450 in video and 1450 in text are not comparable to each other. Each chart must be read on its own. Quality scores are from the public Arena leaderboards, read on 24 August 2026, and are not produced by Pencil.
AI Voiceovers, Sound & Music
Speech, dubbing, sound effects, transcription and music are charged per model, the same way images and video are. Rates below are grouped by the unit the work is measured in.
Model | Generations | Unit |
Charged per minute of audio |
|
|
ElevenLabs 5 models | 1.0 | / min of audio |
ElevenLabs AI Dubbing | 2.1 | / min of audio |
ElevenLabs Music | 1.5 | / min of audio |
ElevenLabs Voice Isolation | 0.8 | / min of audio |
ElevenLabs SFX | 0.5 | / min of audio |
ElevenLabs Scribe v2 (STT) | 0.1 | / min of audio |
Google 2 models | 0.3 | / min of audio |
Gemini TTS 2.5 (2.5 Pro TTS) | 0.3 | / min of audio |
Gemini TTS 3.1 | 0.3 | / min of audio |
Charged per 1,000 characters of script |
|
|
ElevenLabs 3 models | 0.5 | / 1,000 characters |
ElevenLabs Eleven v3 | 0.6 | / 1,000 characters |
ElevenLabs Multilingual v2 | 0.6 | / 1,000 characters |
ElevenLabs Flash v2.5 | 0.3 | / 1,000 characters |
Google 1 model | 0.3 | / 1,000 characters |
Google Chirp 3 HD | 0.3 | / 1,000 characters |
Charged per generated clip |
|
|
Google 2 models | 0.5 | / 30-sec clip |
Lyria 2 | 0.6 | / 30-sec clip |
Lyria 3 | 0.4 | / 30-sec clip |
Average across all 13 models | 0.7 |
|
Note: per-minute rates are charged on the duration of the rendered audio, rounded up to the nearest tenth of a generation. Speech synthesis charged per clip covers one rendered voiceover. Transcription is charged on the length of the source audio, not the transcript. Music models are being brought online through the autumn; the rate shown is the one they will launch on.
Upscaling, Restoration & Avatars
These models operate on an asset that already exists rather than generating a new one. Upscale and restore are priced per megapixel for stills and per second for footage; avatar and lipsync are priced per second of output.
Model Name | Generations |
|
Bria 1 model | 1.4 |
|
Bria Videos (VRMBG 2.0) | 1.4 | / sec |
Topaz Labs 8 models | 0.9 |
|
Topaz AI Video Enhance | 1.6 | / sec |
Topaz Proteus Natural | 1.6 | / sec |
Topaz Starlight Precise 2.5 | 1.6 | / sec |
Topaz High Fidelity V2 | 0.4 | / 24 MP |
Topaz Recovery V2 | 0.4 | / 4 MP |
Topaz Redefine | 0.4 | / 4 MP |
Topaz Standard MAX | 0.4 | / 4 MP |
Topaz Wonder | 0.4 | / 4 MP |
HeyGen 3 models | 0.6 |
|
HeyGen Avatar 5 Digital Twin | 0.7 | / sec |
HeyGen V3 Lipsync Precision | 0.7 | / sec |
HeyGen V3 Lipsync Speed | 0.4 | / sec |
MiniMax 1 model | 0.4 |
|
MiniMax-H3-Regeneration (768P to 2K) | 0.4 | / sec |
Average across all 13 models | 0.8 |
|
Note: a megapixel rate is charged on the output size, not the input, so upscaling a small image to a large one costs more than restoring one at its original size.
Feed Variations
Feed work is charged on what is generated, not on how many formats or cells a job touches. Versioning one creative across many formats stays a single generation, which makes it the cheapest production work on the platform.
Action | Generations | How it is counted |
Feed variation | 1.0 | 1 gen per variation generated |
One creative, many formats | 1.0 | 1 gen regardless of format count |
Feed Sync | 0.1 | 1 gen per ten variations |
AI tool used inside a feed | 1.0 | 1 gen per cell actioned |
After Effects import | 0.1 | 1 gen per 10 sec of rendered output |
Note: Feed Sync is charged at a tenth of the rate of individually generated variations, so routing bulk variation work through Feed Sync is materially cheaper. After Effects imports are priced on the rendered duration of the output, not the length of the project file.
Scores
Default Pencil scores are included and consume nothing. Custom Scores set to Automatic run whenever a work item meeting their criteria is saved, and each run is charged. Scoring agents that call an external data provider cost more per request.
Score | Generations | How it is counted |
Custom Score set to Automatic | 1.0 | 1 gen per run |
Default Pencil scores | 0 | No generations consumed |
GWI Insight Agent | 4.0 | 4 gens per request |
Share of Model score | 2.0 | 2 gens per request |
Note: scores are charged on top of the work item itself. A generated image meeting the criteria of three Custom Scores set to Automatic consumes four generations in total — one for the image, one for each score. Setting a Custom Score to run manually rather than automatically puts you in control of when it is charged.
Tracking Generations
You can view your remaining generations in the counter at the bottom right of your workspace, per-user usage in any workspace’s Manage Team menu, and per-turn cost in the cost menu on any Scribble turn or workflow step.