GPT-6 family released
GPT-6 Astra, Sol and Luna arrived in September, followed by GPT-6.1 Sol with multi-agent support in beta. We are testing them against the GPT-5.x models already in production.
We design, build and run software on OpenAI's platform: support assistants, document automation, voice agents and workflow agents, grounded in your data, measured with evals and kept within budget.
Illustrative demo with sample data.
The platform changed a lot this year. These are the updates that matter most for businesses building on it, and what we do about each.
GPT-6 Astra, Sol and Luna arrived in September, followed by GPT-6.1 Sol with multi-agent support in beta. We are testing them against the GPT-5.x models already in production.
A managed way to run agents, now with computer use. Good news for workflow automation that needs to operate real software.
Apps built on Assistants threads must move to the Responses or Agents API. We migrate them without losing conversation history.
Choose where each request is processed — useful for data-residency rules in Australia, the EU and elsewhere.
Sol, Terra and Luna tiers, with later price reductions that lower the cost of steady workloads.
Run OpenAI models from your AWS account through a Responses-compatible endpoint, alongside Azure OpenAI.
Source: OpenAI API changelog. Checked October 2026.
A demo calls the model and hopes. A production feature wraps it in checks, data and measurement. Follow one request through the system, or tap any step.
No single model is best at everything. Pick a job to see what we would start with and why.
Most teams overpay for AI because every request goes to the biggest model with a fresh prompt. Move the sliders to see how four levers shrink a typical bill.
An illustrative model of how the levers combine, not a quote. We measure your real traffic before promising a number.
Each one is built around a business result — fewer tickets, faster processing, more bookings — not around the model.
Answer orders, bookings and policy questions 24/7 from your own data, and hand over to a person with the full context.
Explore AI chatbots →Turn invoices, contracts, claims and forms into clean, validated data in your ERP or CRM — no retyping.
Explore Generative AI →Ask questions across policies, manuals and past tickets and get answers with sources your team can trust.
Explore AI development →Phone lines that book, reschedule and answer questions in natural speech, in several languages.
Explore Voice bots →Agents that read email, update the CRM, raise invoices and book meetings — asking for approval where it matters.
Explore AI agents →Inspect photos for damage or defects, read screenshots, and create on-brand product and marketing images.
Explore Computer vision →Models do not know your prices, policies or customers. RAG lets them look things up first, so answers are accurate, current and cited.
PDFs, web pages, tickets, databases and SharePoint are gathered and cleaned.
Text is split into meaningful passages that keep headings and context.
Each passage becomes a vector, so similar meaning sits close together.
The question is matched to the best passages the user is allowed to see.
The model writes the answer only from those passages, citing each one.
We do not judge an AI feature by how it feels in a meeting. A test set of real questions scores every change before it reaches your customers.
| Question | Answer given | Result |
|---|---|---|
| Refund within 30 days? | Yes, with receipt | ✓ Pass |
| Ship to Fiji? | Yes, 5–8 days | ✓ Pass |
| Cancel after dispatch? | Return instead | ✓ Pass |
| Price match competitors? | Not offered | ✓ Pass |
| Warranty on chairs? | Answer missed 5-year term | ✕ Fix |
| Change delivery address? | Yes, before dispatch | ✓ Pass |
Enterprise and regulated teams need more than a good answer. These are the controls we set up on every OpenAI project.
OpenAI does not train its models on API data by default. We confirm the settings for your organisation and projects.
For eligible use cases, requests are not stored by OpenAI at all — important for health, legal and finance.
Choose the processing region per request, or run through Azure OpenAI or Amazon Bedrock in your own cloud account.
Healthcare teams can sign a business associate agreement directly in the OpenAI dashboard. See our healthcare work.
Project-scoped keys with expiry, workload identity instead of shared secrets, and hard spend limits per project.
Every prompt, tool call and answer can be logged with PII masked, so you can explain any decision later.
OpenAI retired the Assistants API in August 2026, and Whisper follows in February 2027. We move you to the current APIs without losing history or quality.
Threads, runs and assistant objects stopped working.
One API for text, tools, files and state — the base for new builds.
Multi-step agents with handoffs, sandboxes and memory controls.
Move recordings to GPT Transcribe before the deadline.
| Old | Moves to |
|---|---|
| Assistants + Threads | Responses API with conversation state |
| Runs and run steps | Responses with tool calls, or Agents API |
| Code interpreter (Assistants) | Hosted shell / code tool in Responses |
| File search (Assistants) | File search tool with vector stores |
| Whisper | GPT Transcribe / GPT Live Transcribe |
A typical path for a focused feature. Bigger platforms run as several of these in sequence.
We map the job to be done, the data, the risks and how success will be measured.
A working prototype on your real data, with a first eval set of real questions.
Guardrails, permissions, error handling and cost limits, tested against the evals.
Production release with monitoring, dashboards and a rollback plan.
We review evals, feedback and costs monthly and adopt new models when they win.
We are not tied to one vendor. Where another model fits better, we will tell you, and we often combine several behind one router.
| Model family | Strengths | When we choose it |
|---|---|---|
| OpenAI | Broadest set of APIs: chat, agents, realtime voice, images and video | Products that need voice, agents and multimodal features from one vendor |
| Anthropic Claude | Strong long-document work, coding and careful instruction following | Analysis of long contracts and reports, coding assistants |
| Google Gemini | Very long context and deep Google Cloud integration | Teams on Google Cloud and Workspace, video understanding |
| Meta Llama | Open weights you can host yourself | Strict data control, on-premise or offline deployments |
| DeepSeek | Open models with strong reasoning at low cost | Cost-sensitive reasoning workloads you can self-host |
| Qwen | Open-weight family with strong multilingual support | Multilingual products and self-hosted options |
Use cases, data rules and integrations differ by industry. Explore how we apply AI in yours.
Straight answers to what clients ask us most.
It depends on the job. We usually test the GPT-6 tiers and a GPT-5.6 tier on your own questions, then route simple requests to a smaller model and hard ones to a larger or reasoning model. That mix typically gives better quality per dollar than any single model.
OpenAI shut the Assistants API down on 26 August 2026. We move threads, files and tools to the Responses API or Agents API, keep your conversation history, and add evals so behaviour stays the same or improves.
Not by default: OpenAI does not use API data to train its models. For sensitive workloads we add zero data retention where eligible, regional processing, PII masking, and can deploy through Azure OpenAI or Amazon Bedrock in your own cloud account.
Running costs depend on traffic, prompt length and the model mix. We estimate it per request before building, then cut it with prompt caching, model routing, the Batch API and shorter outputs, and set hard spend limits so bills never surprise you.
A focused feature such as a support assistant or document extraction usually reaches production in 5 to 8 weeks: discovery, a prototype on your data, evaluation and hardening, then launch with monitoring.
Yes. We expose your systems to the model as typed tools or MCP servers with per-user permissions, so it can look up orders, update records or book meetings, with approval steps for anything sensitive.
OpenAI is usually fastest to ship and strongest across voice, agents and images. Open models such as Llama, Qwen or DeepSeek make sense when data must stay on your own servers or volume is very high. We often combine both behind one router.
Tell us what you want the AI to do, the data it needs and who will use it. We reply within 4 business hours with honest advice.