OpenAI consulting that audits, optimizes, ships.

OpenAI consulting is the engineering practice of getting features built on the OpenAI API - GPT-4, GPT-4o, the o-series, Assistants, function calling, structured outputs, batch, fine-tuning - from a working prototype into a production system with cost, latency, and quality under control. The honest version of the category audits the existing deployment first; the dressed-up version sells more API calls. We are the first kind. The proof is Sentinel, our live AI analytics product on Shopify.

Default to Chat Completions plus your own state. The Assistants API earns its place when you need built-in retrieval over uploaded files, code interpreter, or persistent threads you don't want to manage - and you have accepted the latency and the lock-in. We have shipped on both. Most teams arriving at us with Assistants in production would have been better served by Chat Completions plus a Postgres table, and we will say so on the fit call before we accept the engagement.

Write the eval first, pick the model second. GPT-4o is the daily driver for general reasoning at fair cost. The o-series earns its premium when the task is genuine multi-step reasoning, math, code, or planning - and most prompts that look like reasoning problems are actually retrieval or structured-output problems wearing a costume. GPT-4o-mini handles high-throughput classification and simple extraction at one-tenth the cost. We have shipped systems where 80% of calls go to mini, 18% to GPT-4o, and 2% to o-series for the hard ones. Run the eval. Pick by the balance between cost and quality.

Strict mode with structured outputs, JSON-schema validation on the way in, retry-with-correction loops on schema failure, and an eval that scores tool-selection accuracy independently from argument accuracy. Most function-calling reliability failures are not the model picking the wrong tool - they are the model picking the right tool with wrong arguments because the schema was permissive enough to allow it. Tighten the schema. Then tighten the eval.

Three levers, in order of impact. Route by eval - send the cheap calls to GPT-4o-mini, the hard ones to GPT-4o or o-series, and re-run the routing decision quarterly. Use the Batch API for anything not user-facing - half the price for a 24-hour SLA, which fits more workloads than teams realize. Cache aggressively with prompt caching (storing repeated prompt prefixes to cut API costs) on long system prompts. We have cut OpenAI bills by 40-70% on the engagements where the team had been on a single tier for everything.

Almost always: prompt first, then RAG (retrieval-augmented generation), then fine-tune only if the eval says you have to. OpenAI fine-tuning earns its place when the format is rigid, the task is narrow, and you have at least a few hundred high-quality labeled examples. Most teams arriving at us are budgeting to fine-tune three months too early. We will tell you on the fit call whether your problem actually needs it, or whether a tighter prompt and a better structured-output schema would do the same job for one-tenth the engineering cost.

When the eval says so. We do not recommend platform migrations on vibes. We have moved teams off OpenAI to Anthropic when long-context reasoning or instruction-following on safety-sensitive tasks won the eval, and we have moved teams off OpenAI to a fine-tuned open-source model when latency and cost were the deciding factors. We have also kept teams on OpenAI when the migration math did not pay off. If you are already standardized on Claude, you are on the wrong page - see our anthropic consulting practice.

An OpenAI strategy sprint is two weeks, no extensions. A production proof-of-concept is fourteen days from kickoff to a feature serving real traffic behind a feature flag. A full implementation is six to twelve weeks depending on integration surface and data hygiene. We refuse engagements that don't fit a two-week window at the unit level - if the work cannot be sliced into 14-day deliverables, we have not finished scoping it.

Strategy sprints run $25–45k. Production proofs-of-concept run $50–150k. Full implementations start at $150k and scale with integration depth. Embedded team augmentation is a monthly retainer. Pricing is uniform across our consulting practice - we charge by engagement shape, not by which API you are calling.

Yes, mutual NDA before any technical conversation. We do not work for clients with conflicting active engagements in the same competitive set during a quarter - a rule we enforce on ourselves more strictly than most clients ask us to.