Key facts
| Migration shape | Base URL and model-name change for OpenAI-style clients |
| Endpoint inventory | Chat, streaming, JSON mode, function calling and embeddings are live |
| Models | 30+ models behind the same endpoint |
| Pricing | Flat monthly self-serve plans with unlimited fair-use usage |
| Free tier | plugsky-micro and plugsky-lite, no card required |
| Trial | 14-day full-access trial for stronger models |
| Rollback | Revert the base URL and key; no schema changes |
TL;DR
- Audit endpoints before touching code; migrate by capability, not by service.
- Replay recorded traffic so evaluation reflects production.
- Canary a small traffic share before full cutover.
- Keep rollback to a configuration change, not a release.
- Move text workloads first; leave media and platform endpoints for later.
How it works, step by step
- Inventory every OpenAI endpoint and feature in use, including retries and error handling.
- Create a Plugsky key on the free plan and map each model name to a replacement.
- Point staging at the Plugsky endpoint and run your existing test suite.
- Replay recorded production requests and compare outputs, latency and errors.
- Canary a small production slice and monitor quality and spend dashboards.
- Complete cutover, keep rollback documented, and review after one full traffic cycle.
Try it yourself
Open the OpenAI migration checker →
Stage 0: inventory and compatibility audit
Migration failures usually come from hidden dependencies, not from the happy path. List every endpoint, SDK helper and framework integration in your codebase, then mark each as live or coming soon on the target platform. Plugsky has chat, streaming, JSON mode, function calling, embeddings, RAG and agents live; audio, images, moderation, files, batch, fine-tuning, assistants and responses are coming soon.
- Search for SDK imports and raw HTTP calls to the old base URL.
- List model names and their roles, from classification to reasoning.
- Note timeouts, retries and fallback behaviour you rely on.
- Record which services share keys and rate limits.
Stages 1 and 2: replay and canary
Replay recorded traffic against the new endpoint before you change anything live. Diff the outputs with a rubric rather than eyeballing samples: correctness, format stability, refusal behaviour and latency percentiles. This catches prompt sensitivity that unit tests miss.
Then canary. Send a small share of production traffic to Plugsky behind a feature flag and watch error rates, time-to-first-token and quality signals. The free plan and the 14-day full-access trial make this stage inexpensive, and flat monthly self-serve pricing means a canary does not distort a token budget.
Stage 3: cutover, monitoring and rollback
The cutover itself should be a configuration change: base URL, key and model mapping. Keep the old values one flag away so on-call engineers can revert in seconds. After cutover, review model mix, error rates and spend weekly until behaviour is boring.
Rollback is the same one-line change in reverse because the API stays OpenAI-compatible. See the live pricing page for current plans, and plan a follow-up review once you have a full billing cycle of data.
Honest comparison
| Phase | Staged migration to Plugsky | Big-bang rewrite | Staying on OpenAI |
|---|---|---|---|
| Compatibility check | Schema audit in staging | Discovered in production | Not needed |
| Evaluation | Replayed traffic plus rubric | Limited | Not needed |
| Cutover | Canary then ramp | Single switch | Not needed |
| Rollback | Config change | Redeploy release | Not applicable |
| Cost shape after | Flat monthly self-serve | Unchanged | Usage-based |
Frequently asked questions
Do I need to rewrite my application?
No, if it uses the OpenAI SDK or an OpenAI-compatible framework. The migration is a base URL, key and model-name change plus evaluation.
How long does an OpenAI API migration take?
Most teams finish a staging migration in days and a production cutover in one to two weeks, with most time spent on evaluation rather than code changes.
Is there a free way to test the migration?
Yes. plugsky-micro and plugsky-lite are free with no credit card, and a 14-day full-access trial opens stronger models for hard prompts.
How is pricing structured?
Self-serve plans are flat monthly with unlimited fair-use usage and no per-token billing, which makes canary periods easy to reason about. See the live pricing page.
What if I need audio or image endpoints?
Those are coming soon on Plugsky. Keep them on OpenAI for now and migrate the text, tool and embedding workloads first.
How fast is rollback?
As fast as a configuration change. Revert the base URL and key and the old provider serves traffic again.