ComfyUI v0.38: New Nodes, and Recraft Costs 43% More

ComfyUI shipped version 0.38.0 on 29 September 2026 with 59 merged pull requests, eight of them for Partner Nodes: Tencent HY Image 3.5, Seedream 5.0 Flash, Seedance 2.5 Draft, Recraft V4.1 Flash, Quiver Arrow 2, Claude Opus 5.5, and GPT-6 Sol and Luna. It also removed the Sora nodes and dropped torchaudio from the core requirements.

The changelog leaves one thing out. In this release every Recraft generation node got 43% more expensive. We read the price badges in the node source at the v0.37.0 and v0.38.0 tags: Recraft V4 went from $0.04 an image, the same as Recraft's own API price, to $0.0572. That 1.43x factor is not new. It appears on every Claude and GPT chat model added to ComfyUI since 22 July, and v0.38 carries it to three more models.

What v0.38.0 ships

The release splits into paid Partner Nodes, which run on the provider's servers and bill Comfy credits, and local changes that cost nothing to run. The paid side is where the new models are.

Rounded, the three new image models cost $0.01 (Recraft V4.1 Flash), $0.026 (Seedream 5.0 Flash) and $0.034 (HY Image 3.5 at 2K) per image.

The Seedance 2.5 Draft flow is the most useful workflow change here. You run a cheap 480p draft, check it, and pass its draft_task_id to the "Seedance 2.5 Draft to Final Video" node, which renders the 1080p version of that exact clip. The node tooltip warns that you must set the draft node's seed control to fixed, or the next queue generates a new draft instead of finishing the one you approved. Drafts stay renderable for 7 days.

The HY Image 3.5 Edit node accepts @Image1 to @Image5 references in the prompt. Comfy's guide positions it for multilingual text rendering, photorealism, and keeping people and products consistent across edits.

Recraft got 43% more expensive inside ComfyUI

The Recraft V4.1 Flash pull request (#16501) repriced every Recraft node. Its checklist ticks "Need pricing update" and "Metronome rate cards updated", so billing changed as well as the badge. The table below compares Recraft's direct API price with the ComfyUI badge before and after.

Of the 11 Recraft nodes with a flat price, 10 rose exactly 1.43x: text to image, image to image, inpainting, text to vector, vectorize, replace background, remove background, crisp upscale, creative upscale and V4 Create Style. The one that moved the other way is the legacy Create Style node, which dropped from $0.04 to $0.00715.

In practice, 100 Recraft V4 images cost $4.00 through Recraft's API and $4.00 through ComfyUI last week. From v0.38 they cost $5.72 through ComfyUI. Recraft V4.1 Flash is still the cheapest image model in the release at about a cent an image, but that cent is $0.003 more than Recraft charges directly.

The 1.43x pattern across Claude and GPT nodes

The same factor explains the new LLM prices. Anthropic lists Claude Opus 5.5 at $4 input and $20 output per million tokens. The ComfyUI Claude node shows $5.72 and $28.60. OpenAI lists GPT-6 Sol at $2 and $10; the node shows $2.86 and $14.30.

Of the 14 LLM price rows we could match to a published vendor price, 8 sit at 1.43x and 6 at list. The split is by date. The git history shows the 1.43x rows arrived with the Opus 4.8, Fable 5 and Sonnet 5 commit on 22 July 2026. Every Claude and GPT chat model added since then carries the same factor, while GPT-5.6, added on 16 July, and everything older still bill at list. Claude Sonnet 5.5, merged to the main branch on 30 September after v0.38 was tagged, falls into the Sonnet 5 row and shows $2.86 and $14.30 against a $2 and $10 list. The Arrow 2 price formula in nodes_quiver.py multiplies Quiver's token rates by 1.43 in plain sight.

Comfy does not publish a margin. Its Partner Node pricing page lists prices in credits only, and the Comfy Router pricing page gives the conversion: $1 buys 211 credits. The credit figures match the badges to the cent, so Recraft V4 is 12.07 credits and Opus 5.5 input is 1.207 credits per thousand tokens. There is nothing hidden here, but you only see the multiple if you put the two price lists side by side.

For a single image, 43% is small change. It adds up in a prompt-rewriting chain: an Opus 5.5 node that reads 50,000 tokens and writes 10,000 on each queue costs $0.40 a run with your own Anthropic key and $0.572 through the Partner Node. Over 1,000 runs, that is $400 against $572.

What is free: the local changes

Most of the 59 pull requests touch local inference, and none of them bill credits.

• Qwen-Image 2.1: compiled transformer blocks, a tiny VAE for fast previews, union Fun ControlNet support and fp16 activation clamping. Check the Qwen-Image 2.1 research license before client work.
• Ming-Image: native support for the 6B design model we measured for VRAM, plus a fix for saving and loading it.
• MiniMax-H3: Fun-ControlNet-Union 2.0 support, an offload crash fix, and VAE tiles blended against their neighbours, which follows the 2x faster H3 VAE.
• HDR: the Convert Color Space node gains ARRI LogC3 and ACEScct, and a fix for negative values in HDR output.
• Quantization and speed: w6a8 weights, CUDA graphs for ACE-Step 1.5, faster long-context Qwen 3.5 and 3.8 text generation, Flux family optimizations, and missing AMD RDNA2 architectures added.
• Text generation: the TextGenerate node gets a system prompt input and a separate thinking output.

The torchaudio removal (#16457) is the one to watch if you use custom nodes. The pull request says torchaudio "seems to be abandoned by upstream (no stable cu132 build)" and that portable and desktop builds will keep installing it "for a bit longer". A fresh manual install from requirements.txt no longer pulls it, so a custom audio node that imports torchaudio will fail until you add it back with pip.

The Sora nodes are gone

Pull request #16609 deleted nodes_sora.py, and Comfy's pricing page now marks the OpenAI Sora Video node as retired. Any saved workflow that uses it will load with a missing node in v0.38. Seedance 2.5 and 2.0, Kling, Runway, Veo and Wan video nodes remain in the release.

How to decide: Partner Node or your own key

The multiple does not make Partner Nodes a bad deal. You get one credit balance, no provider accounts, and no API keys in your workflow files. The question is volume.

1. Open the node and read the badge. It is in the node header. Compare it with the vendor's price page for the same model.
2. Estimate a month of runs. Multiply the badge by your realistic queue count, then do the same with the list price. The gap is 30% of the badge (1 minus 1/1.43).
3. Under about $20 a month, keep the Partner Node. The saving from a direct key is smaller than the time spent managing another account.
4. For heavy LLM chains, use a direct key. A custom node that calls Anthropic or OpenAI with your key pays list price. Keep the key in an environment variable, not in the workflow JSON you share.
5. Check older model rows. Opus 4.7, Sonnet 4.6 and GPT-4.1 still bill at list inside ComfyUI, so for tasks they handle well they avoid the multiple entirely.
6. Re-read badges after every update. This change arrived inside a feature pull request, not a pricing announcement.

How we checked

We cloned ComfyUI at the v0.37.0 and v0.38.0 tags and parsed the price_badge expressions in comfy_api_nodes/. We compared them with Anthropic's, OpenAI's and Recraft's published price pages on 30 September, and cross-checked four rows against the credit figures in Comfy's docs at 211 credits per dollar. All four matched. We did not run paid generations, so the figures are what ComfyUI shows and documents, not measured invoices. The script and output are kept with this article.

Frequently asked questions

What is new in ComfyUI v0.38.0?

Partner Nodes for Tencent HY Image 3.5, Seedream 5.0 Flash, Seedance 2.5 Draft, Recraft V4.1 Flash, Quiver Arrow 2, Claude Opus 5.5, and GPT-6 Sol and Luna. Locally, it adds Ming-Image support, Qwen-Image 2.1 speedups, w6a8 quantization and HDR LogC3 and ACEScct conversion. It removes the Sora nodes and the torchaudio requirement.

Did Recraft raise its prices?

No. Recraft's own API still lists V4 at $0.04 and V4.1 Flash at $0.007. The increase is on the ComfyUI side: the v0.38 Recraft nodes show 1.43x Recraft's price, where v0.37 showed Recraft's price.

Why do ComfyUI Partner Nodes cost more than the provider's API?

Comfy does not state a reason or a margin. What the source shows is a 1.43x factor on every Claude and GPT chat model added since 22 July 2026, and now on Recraft. Older model rows still bill at list.

How many credits is a dollar on Comfy?

211 credits per $1, per Comfy's own pricing docs. Recraft V4 costs 12.07 credits per image and HY Image 3.5 costs 7.24 credits at 1K or 2K.

Will my old Sora workflows still run in v0.38?

No. The Sora nodes were deleted, so those workflows load with missing nodes. Rebuild them on another video node.

Will removing torchaudio break my custom nodes?

Only on a fresh manual install. Portable and desktop builds still include torchaudio for now. If a custom audio node fails with an import error, run pip install torchaudio in ComfyUI's Python environment.