Replicate after Cloudflare: pricing, MCP and alternatives

This page is built from Replicate's own site, docs, changelog and terms, Cloudflare's announcements and 2025 annual report, Google's Gemini API price list, the model pages of fal, WaveSpeedAI and Kie.ai, Artificial Analysis' provider benchmarks, an AI Photo Labs test, and user reports on Trustpilot, Hacker News and GitHub, all checked on October 21, 2026. Replicate's claims are labelled as its own.

What Replicate sells

Replicate runs AI models behind one API. Its docs describe "over 100 official models" that are always on, priced per output and maintained with their authors: Veo 3.1, Nano Banana Pro, Kling 3.0, Seedance 2.0, FLUX 2, GPT Image 2 and a few language models such as Claude. Beyond those, Cloudflare counts more than 50,000 community and fine-tuned models, billed per second of hardware (an A100 80GB at $5.04 an hour, an H100 at $5.49). Private models are billed for all time online, idle included. You can package your own model with Replicate's open-source Cog tool and run it on the same bill.

Since July 16, 2025 new accounts pay with prepaid credit, and there is no fixed free credit.

The Cloudflare acquisition

Cloudflare announced on November 17, 2025 that it would buy Replicate, and the deal closed on December 1, 2025. Cloudflare's annual report for 2025 puts the total purchase cash consideration at $57.4 million, $46.6 million of it recorded as goodwill. At the announcement Replicate said "the API isn't changing", and its API and per-output prices show no published change since.

Other things have moved: Replicate's status page now redirects to Cloudflare's. On April 16, 2026 Cloudflare said the Replicate team had joined its AI Platform group and that Replicate's models would come to its AI Gateway; as of October 21, 2026 that was not visibly live, and Cloudflare's August 7, 2026 post merging Workers AI and AI Gateway does not mention Replicate. Replicate's public changelog has been quiet since April 21, 2026.

Prices against the official APIs

Prices in US dollars, from each service's own model or pricing page and Google's Gemini API pricing (updated October 19, 2026):

Google's models cost more on Replicate than on Google's API: Veo 3.1 Fast 50% more, Nano Banana Pro about 12% more (25% at 4K), exactly what fal charges; Veo 3.1 at 1080p matches Google. Kling's and BytePlus's own prices could not be confirmed, so the last two rows compare the four services only. Replicate charges double what fal and WaveSpeedAI charge for Kling 3.0 (on its tier labelled "standard") and is the cheapest of the four on Seedance 2.0, though its page does not say whether audio is included. Kie.ai is lower on Google's models but does not say how it gets access. Replicate lists no platform fee.

What public tests show

Artificial Analysis, an independent benchmarking firm, runs the same prompts on several providers every day. Its three-day median end-to-end times, read on October 21, 2026:

Replicate was the fastest or near-fastest of the three on most shared models (its Wan figure is its Fast endpoint). Two caveats: its FLUX.1 [dev] p95 was 25.1 s, a sign of occasional cold starts or queueing, and it charged $25 per 1,000 FLUX.1 [dev] images against WaveSpeedAI's $12. On FLUX.2 [klein] 4B it was the cheapest at $1 per 1,000. Only open-weight models are covered.

AI Photo Labs, a comparison site whose operator is not named, timed Replicate against fal on March 30, 2026. Replicate was faster on FLUX (about 1.40 s against 3.49 s on FLUX.1 Dev) but much slower on a burst of eight SDXL requests (about 21.5 s against 3.3 s), about 17.6 s of it spent queueing. The sample is small.

A September 2026 DEV post by mrzitoun reports a 6.5 s median cold start for Replicate on an A100 against 1.8 s on Modal, without a described method. WaveSpeedAI's blog claims Replicate takes 2 to 5 minutes per video; that is a competitor's claim with no method shown.

What users report

Trustpilot shows 1.8 out of 5 from 16 reviews, 94% one star, on an unclaimed profile with no replies. Reviewers report Seedance requests blocked by a "copyright" moderation error with funds stuck in credit (October 9, 2026), prepaid credit lost after a year of non-use (September and October 2026), and unanswered support tickets.

On Hacker News, an August 17, 2026 thread asked "Is there still a functioning team behind replicate.com?" after a support request went unanswered for a week while the LoRA trainer returned HTTP 500 errors. On GitHub, issues in the JavaScript client sit without replies, including a reported billing loop on a disabled private deployment (#382, August 24, 2026). The Cog repository is still active, with v0.23.1 released on October 19, 2026.

MCP, Agent Skills and coding agents

Replicate publishes an official MCP server in two forms: a remote server at mcp.replicate.com, launched August 14, 2025, and a local npm package, replicate-mcp. Its tools cover the HTTP API: searching models, reading metadata, running predictions and checking status. Setup is documented for Claude Desktop, Claude Code, Cursor and VS Code with GitHub Copilot, and the docs name Codex CLI and Gemini CLI as compatible. An experimental "code mode" instead lets the agent write TypeScript against the Replicate SDK in a sandbox.

On April 21, 2026 Replicate published seven Agent Skills in its replicate/skills repository (find-models, compare-models, run-models, build-models, publish-models, prompt-images, prompt-videos), installed with npx skills add replicate/skills and aimed at Claude Code, OpenCode, Codex and Pi. Replicate pairs them with MCP: the server makes the calls, the skills give the guidance.

Replicate has no OpenAI- or Anthropic-compatible endpoint in its docs, so it is not a drop-in model backend for coding agents. Official clients exist for Python (a new SDK in public beta since October 2025) and JavaScript; the Vercel AI SDK has a Replicate provider for images, LiteLLM supports Replicate models, and Replicate generates llms.txt files for its docs.

Privacy, terms and who runs it

• The contracting company is Replicate, LLC, a Cloudflare subsidiary, under terms updated April 1, 2026.
• API predictions' inputs, outputs, files and logs are deleted after one hour by default; predictions made in the web interface are kept until you delete them.
• Customers own their data and outputs, subject to model licences. Replicate's licence covers using customer data to build that customer's fine-tuned models, and anonymized aggregate data to improve its services.
• Prepaid credit expires 12 months after purchase and is non-refundable. Failed runs on public models are not charged.
• California law applies, with AAA arbitration and no class actions.
• No uptime figure or self-serve SLA is published. Cloudflare's status page listed Replicate as operational on October 21, 2026, with incidents on 2 of the last 30 days and no details shown.

How it compares with fal, WaveSpeedAI and Kie.ai

fal charges the same for Google's models and half as much for Kling 3.0, and publishes a full incident history. WaveSpeedAI also halves the Kling price, accepts Alipay and WeChat Pay, and keeps credits forever. Kie.ai is cheapest on Google's models but does not say how it gets access. Teams on Cloudflare can send Replicate calls, with their own token, through Cloudflare AI Gateway for logging and control.

Who should skip it

Skip Replicate if you need quick answers from support or a clear roadmap now, or if you mainly buy Kling or Google's models, which are cheaper elsewhere. Spend any credit within a year.