Replicate is the multimodal home, image (Flux, Recraft, SDXL), video (WAN, Veo, Sora-style), voice (ElevenLabs-style), and LLM inference all behind one API with per-second GPU billing. The Cog packaging format lets you bring your own models if the catalog doesn't have what you need. The trade is LLM-specific: Together and Groq are faster and cheaper for raw text inference. For apps that mix modalities, chat that generates images, agents that produce video, Replicate is the right shape.
- Multimodal, LLM, image, video, voice models all in one API
- Per-second GPU billing (T4 to H100), no minimum commit
- Cog packaging format for custom model hosting
- Image gen at $0.04-$0.09/output; video at $0.09-$0.25/second
- Streaming support for chat / agent UX
- Slower and more expensive than Together / Groq for pure LLM inference
- Per-second billing includes cold-start time on first invocation
- Token-based pricing only on featured LLMs, most models bill by GPU time
- Less specialized than vertical providers (Fal for image, ElevenLabs for voice)
- Custom model deployment via Cog adds operational complexity
Replicate is a model-hosting platform that runs open-source AI models behind a unified API, LLMs, image generators, video models, voice synthesis, all with per-second GPU billing and streaming support. The Cog open-source packaging format means any team can deploy a custom model with the same hosting primitives that power the catalog.
Where it fits
Replicate is the right pick when your product spans multiple modalities, a chat app that also generates images, an agent that produces video, a creator tool that needs voice synthesis alongside LLM reasoning. Maintaining one provider, one API key, and one billing surface across all of that is the wedge.
For featured models, pricing is hybrid: image generation by output ($0.04-$0.09), video by second of output ($0.09-$0.25), LLMs by token (e.g., Claude 3.7 Sonnet at $3/M input). For most other models, you’re paying per second of GPU runtime on the hardware the model needs (T4 at $0.81/hr to H100 at $5.49/hr).
Avoid Replicate when LLM inference is your primary workload, Together and Groq are faster and cheaper. Avoid for high-volume production image gen if you’re doing it constantly, dedicated endpoints elsewhere become more economic.
Pricing in practice
GPU-second billing across the lineup: T4 at $0.81/hr, L40S at $3.51/hr, A100 80GB at $5.04/hr, H100 at $5.49/hr. Featured models use hybrid pricing, Flux at $0.04/output image, WAN 2.1 video at $0.09-$0.25/second depending on resolution, token-based pricing on the LLM featured models.
The pay-for-what-you-use model means cold starts (model loading time on first request after idle) are billed at the same rate as inference. Keeping models warm via concurrency settings reduces this cost for production workloads.
How it compares
Together AI, Faster, cheaper, deeper LLM-specific serverless inference. Pick when text is the primary workload.
Fireworks AI, LLM-focused serverless with focus on speed and fine-tuning. Pick when LLM is the main workload and Together’s coverage isn’t quite right.
Modal, Python-native serverless GPU with full code control. Pick when you’re building custom inference pipelines, not consuming models.
What changed recently
Replicate has expanded video generation coverage through 2025-2026 with multiple frontier video models (WAN, Sora-style alternatives, Pika, Runway) and improved per-second billing across resolution tiers. The Flux family from Black Forest Labs continues to lead the image generation surface, with Recraft providing a designer- focused alternative. Streaming support landed across more LLM endpoints, improving chat / agent UX. The Cog packaging format continues to be the differentiator for custom model deployment without operating Kubernetes.
Sources
- Replicate Pricing, replicate.com
- Replicate Docs, replicate.com
- Cog, github.com/replicate/cog
- Replicate Explore, replicate.com
- beat · 2026-05-30
Groq raising $650M to fund the inference neocloud pivot, Axios reports
The Axios scoop pegs Groq at up to $650M from existing investors, with Disruptive and Infinitum committing to backstop the round. Smaller than 2025's Series E and with no disclosed valuation, it reads as a Groq 2.0 raise.
- comparison · 2026-05-25
Inference platforms in 2026: how Modal, Replicate, Together, and Fireworks divide the category
After Modal's $355M Series C, the GPU inference category divides cleanly across four platforms: Modal as a Python runtime, Replicate for multimodal, Together as the OSS LLM workhorse, and Fireworks as Together's performance-focused peer.
- beat · 2026-05-24
Modal closes $355M Series C at $4.65B valuation
Modal raised $355 million at a $4.65 billion post-money valuation, led by General Catalyst and Redpoint. The round, which came together in two tranches, lands seven months after a $1.1 billion Series B and follows a fivefold revenue jump to roughly $300 million ARR.