**DRAFT — pending editorial expansion.** This article is a working draft published as scaffolding for the NINtec content programme. The current version covers the substantive perspective in compressed form; the published version will expand each section to the 2,000+ word depth the topic warrants. Editorial review is required before promotion.
Teams considering whether to migrate from OpenAI to Claude rarely fail because Claude cannot do the job. They fail because the migration is run as a model swap when it is actually a port. This playbook is what we have learned running OpenAI-to-Claude migrations across fintech, healthcare, logistics, and SaaS clients.
Phase 0: Decide with data, not anecdote
The first step is not the migration; it is the decision. Run a head-to-head eval on your specific workload before any migration commitment. Discovery produces eval data on task accuracy, latency, cost per outcome, and edge-case behaviour. Decide whether to migrate based on the data; do not migrate because of vendor preference or generic benchmarks. We have completed Discovery engagements where the recommendation was 'do not migrate' — and those clients kept their existing OpenAI deployments.
Phase 1: Prompt re-engineering
Direct prompt translation between providers loses 5–15% quality. Claude rewards different patterns than GPT — XML-structured prompts with named tags, explicit thinking-block reasoning, structured-output schemas with concrete examples. Production migration includes prompt re-engineering, not direct translation. Each prompt is rewritten, eval-tested, and locked at the new eval-bar before promotion.
Phase 2: Eval parity
Existing OpenAI evals are extended with Claude-specific test cases. The migration ships only when Claude meets or exceeds the existing eval-bar on the workload. This is non-negotiable; migrations that skip parity validation produce regressions that show up in production weeks later.
Phase 3: Telemetry and cost re-mapping
Token counting differs between providers. Cost dashboards need updating to reflect the new economics — including prompt caching opportunities (Anthropic's caching is unusually efficient and frequently surfaces 30–70% cost reductions on repeated-context workloads). Per-feature, per-tenant cost telemetry is rebuilt against the new provider.
Phase 4: Tool-use translation
Function calling (OpenAI) maps to tool use (Anthropic) conceptually but structurally differs — JSON schema mapping, parallel-call semantics, error handling. We have a translation playbook for the common cases plus manual review for edge cases. Production tool-use loops are tested end-to-end before cutover.
Phase 5: Dual-run cutover
Both providers run warm during the migration window. Workloads cut over one at a time behind feature flags. Rollback is always one flag flip away. Most migrations run in dual mode for at least 4 weeks before the rollback path is decommissioned. This is the discipline that keeps the migration safe; the alternative — flag-flip and pray — has caused costly incidents in our portfolio history.
Phase 6: Steady-state cleanup
OpenAI dependencies are removed only after Claude has run alone in production for 4–8 weeks without regression. Cost telemetry shows projected savings realised. The team migrates institutional knowledge — runbooks, documentation, training — to the Claude-centric posture.
When to migrate, when not to
Migrate when: long-context workloads dominate (Claude wins), prompt caching opportunities are large (Claude wins on cost), enterprise contract terms favour Anthropic, codebase-aware tooling is in scope (Claude Code is more developed than alternatives). Do not migrate when: the workload is undifferentiated, your team's existing OpenAI fluency is deep, your existing OpenAI / Azure procurement is established and switching would create more friction than capability gain. Honest model choice rests on the workload, not the vendor.
NINtec's migration practice
Typical migration runs 8–16 weeks end-to-end. NINtec's OpenAI-to-Claude migration practice has shipped this across 10+ clients across regulated and non-regulated industries. The playbook is documented; the engineering is repeatable; the dual-run discipline is non-negotiable.