Working with the OpenAI API?

Adding AI to your product? We build it properly.

Most teams can get a demo talking to GPT in an afternoon. Shipping that same demo into a real product, under real users, without a runaway bill or a leaked prompt, is a different piece of work. That second step is the part this team does.

  • Since 2015Fixing technology
  • FreeEstimates
  • US teamReal engineers

openai.log

still failing

What it is

OpenAI API at a glance

The OpenAI API gives programmatic access to OpenAI's models from your own code. The endpoints include chat completions covering the GPT family, embeddings through text-embedding-3-small and large, image generation via DALL-E and GPT-4o image, audio with Whisper for transcription and TTS for voice output, a moderation endpoint, fine-tuning, and the Assistants API which wraps threads, runs and tools.

There is also a Realtime API for low-latency voice conversations and a Batch API that processes jobs asynchronously at half the per-token price. Billing is per million tokens, with input and output tokens priced separately on every model. SOC 2 compliance, enterprise data processing agreements and a zero-retention opt-in are available for workloads that need them.

Type
AI model API
First released
2020 as the GPT-3 beta
License
Proprietary SaaS
Maintained by
OpenAI
Typically used for
Chatbots, summarisation, classification, data extraction, embeddings and semantic search, voice, image generation, code assistance

The honest version

OpenAI API pros and cons

No technology is all upside. This is where OpenAI API earns its place and where it tends to cause trouble.

Where OpenAI API is strong

  • The strongest frontier modelsGPT-4o, the o1 reasoning line and the GPT-5 family are at or near the top of most benchmarks that matter for real product work.
  • A mature API surfaceChat completions, embeddings, Assistants, Realtime, Batch and structured outputs through JSON schema all share one account and one SDK.
  • Streaming and tools built inServer-sent event streaming, function calling and vision plus audio in the same endpoint are supported without a second service in front.
  • SDKs for every stackOfficial libraries for Python, Node, .NET, Go and Java keep auth, retries and streaming consistent across whichever language the product already runs.
  • Easy to start withA free tier and a playground mean a team can prototype an idea the same afternoon it is proposed, before any procurement conversation.
  • Assistants threads hold stateThe Assistants API abstracts conversation history, tool calls and file search, which saves writing a lot of glue code for a first version.

Where OpenAI API bites

  • Token costs add up fastLong system prompts, retrieved context, tool definitions and verbose outputs each cost money on every call, and the bill is not obvious until it arrives.
  • Prompt injection is a real riskAny time user input or scraped content reaches the model, an attacker can try to hijack it. Chat completions do not solve this on their own.
  • Rate limits vary by tier and modelA new account gets small per-minute token and request quotas, and limits differ between models in ways that catch load tests by surprise.
  • Models get deprecated quicklyOpenAI retires specific model versions on months of notice, not years, so integrations need a plan for rolling forward on their own schedule.
  • Enterprise terms differ from ChatGPTData retention, training opt-out and compliance on the API are governed by separate agreements from the consumer product, and they need reading.
  • Structured outputs still need validationJSON schema mode is reliable but not absolute. Production code still needs to parse, validate and fall back when the response does not fit.
  • Latency is uneven at peakResponse times vary with load and model, and a demo that felt instant at 10am can feel sluggish under traffic in the evening.

Sound familiar?

Common OpenAI API problems we fix

The OpenAI API problems that bring people to us, and what you have once each one is fixed.

monthly spend alert: $3,200 (budget $800)

The OpenAI bill is out of control

Usually a prompt that keeps growing, a retrieval step that pulls far too much context, or a loop that retries silently on every error.

Costs cut, with numbers to show it

prompt injection detected in support reply

User input is reaching the model unchecked

Scraped pages, uploaded files or chat input carry instructions that override the system prompt and change what the assistant does.

Inputs scoped and sanitised

RateLimitError: You exceeded your current quota

The integration keeps hitting rate limits

Per-minute token and request limits on the tier are being hit at peak, often made worse by retries that stack on top of each other.

Backoff, batching and the right tier

The model gpt-4-turbo-preview does not exist

A model we depended on was deprecated

Hard-coded model names, no fallback and no plan for the next version mean the integration simply stops when a model is retired.

Pinned, with a tested upgrade path

last developer: unreachable

A previous developer built the AI feature

Nobody is sure which keys are in use, what the system prompt does, or why some answers are wrong, and the original author is gone.

Taken over, documented, in your account

Assistants API: run stuck in queued state

The chatbot sometimes just hangs

Threads that never complete, tool calls that are not handled, or streaming connections that drop and are not recovered by the client.

Reliable, with timeouts and retries

embedding vector dimension mismatch (1536 vs 3072)

Semantic search returns the wrong things

Embeddings generated with mixed models, no re-indexing after a change, or a vector store misaligned with the model in use.

One model, one index, measured recall

PII sent to OpenAI without agreement

Compliance is not sure the data can go to OpenAI

Customer records, health information or regulated data being sent to the API with no data processing agreement and no retention controls.

Agreements in place, retention opted out

What we do

OpenAI API help, from first feature to ongoing upkeep

Call us for one feature, or hand us the whole integration.

Add AI to an existing product
Chatbots, copilots, summarisation, extraction and search bolted onto what you already ship, without a rewrite.
Cost and performance
Prompt and context trimming, Batch API where it fits, caching and model selection, with the monthly bill measured before and after.
Rescue and takeover
An integration a previous developer built, now in your account, documented, and working as it was meant to.
Safety and prompt injection
Input scoping, output validation, moderation and logging so user-controlled text cannot quietly hijack the assistant.
Retrieval and embeddings
Picking the right embedding model, the right vector store, and a chunking strategy that answers questions on your own data.
Managed Tech
Monitoring, model upgrades, spend alerts and a team to call, on an ongoing basis.

How it works

From first message to fixed

  1. Look Free estimate

    Tell us what's going wrong. We look at it, then tell you plainly what's wrong and what it will cost.

  2. Fix

    We fix the cause, not the symptom. If something is down, getting it back up comes first.

  3. Harden

    Backups, monitoring, updates and access, so the same problem can't quietly come back.

  4. Hand back or stay on

    You get a plain-language write-up. Or we stay on for hosting, upkeep and new work.

Often found together

Running OpenAI API alongside something else?

We work on the whole stack, so the problem can be wherever it actually is.

Questions

OpenAI API questions

Something else? Call+1 737 727 7327

Can you add an AI chatbot or copilot to our existing app?

Yes, and most of the plumbing is usually already there. Most products already have the data, the auth and the UI the feature needs. We hook the OpenAI API into what is there, scope what the model can see and do, and ship it as a normal part of the application rather than a separate experiment.

Why is our OpenAI bill so high?

The common causes are a system prompt that has grown over time, retrieval that returns far more context than it needs, verbose responses that nobody caps, and retries that run silently in the background. We instrument the calls, show you where the tokens are going, and reduce the ones that are not earning their keep.

How do we stop prompt injection in our OpenAI integration?

By treating any user-controlled input, including uploaded files and scraped pages, as untrusted. That means scoping what the model is allowed to do with tools, validating structured outputs, keeping secrets out of the prompt, and logging unusual responses rather than relying on the system prompt alone.

Can you take over an OpenAI integration a previous developer built?

Yes, and takeovers are a large share of this work. We move the keys into your own OpenAI account, document the prompts and tools, fix the obvious gaps, and give you a clear picture of what the integration does and where its costs come from.

Should we use the OpenAI Assistants API or just chat completions?

Chat completions give you more control and are easier to reason about for most product features. The Assistants API is useful when you want OpenAI to manage conversation state, file search and tool orchestration for you. We will pick the one that fits your case rather than the newest one.

How much does OpenAI API help cost?

Estimates are always free, with no obligation. Describe the feature or the problem and you get a price before any work begins.

Free estimate

OpenAI integration giving you trouble?

Tell us what you're running and what's going wrong. A person reads it and comes back with an estimate.

  • No charge to find out what it will take
  • If it's urgent, say so

Replies come from a person. By sending, you authorize us to email you back. No marketing.