The most common question on a discovery call after pricing: "Which AI should we be using?" The honest answer in Q3 2026 is that the three frontier models — ChatGPT, Claude and Gemini — are close enough that the right pick depends entirely on the job. Here's how we actually choose for client builds.
The TL;DR
- Need reliable long-context reasoning, careful writing, or production code? Default to Claude.
- Need the deepest tool/plugin ecosystem or image generation in-line? Default to ChatGPT.
- Live inside Google Workspace and need huge context windows? Default to Gemini.
- Building for an enterprise that's standardised on Microsoft 365? Default to Copilot (which runs on GPT under the hood).
- Need on-prem or cost-sensitive at scale? Default to open-source (Llama, Mistral, Qwen).
Almost every real build uses more than one. The choice isn't religious — it's situational.
Where each one wins
ChatGPT (OpenAI)
The broadest utility model. Best plugin and tool ecosystem on the market. Strong multimodal capabilities — DALL·E 3 inside the same interface, voice mode, vision. The GPT Store and Custom GPTs make it the easiest place to ship a quick internal assistant without writing code.
Wins on:
- Tool calling and plugin orchestration
- Image generation alongside text
- Voice-mode conversational interfaces
- "I want to ship this in an afternoon" internal use cases (Custom GPTs)
- The largest pool of developers who already know the API
Loses on:
- Long-context reasoning when documents exceed ~100K tokens
- Carefully nuanced business writing (Claude tends to read more thoughtfully)
- Coherence over very long conversations
Claude (Anthropic)
The model we default to for nuanced business work. 200K-token context window, strong reasoning, careful and considered prose, excellent code generation. The "Projects" feature makes it the easiest place to build a knowledge-grounded internal assistant without an integration.
Wins on:
- Long-document reasoning (200K context handled well, not just nominally)
- Code generation and review — particularly with complex requirements
- Business writing where tone and judgement matter
- Following instructions precisely and avoiding the helpful-but-wrong trap
- Computer use and agentic workflows (more reliable in 2026)
Loses on:
- Image generation (no native model)
- Plugin/tool ecosystem (smaller than OpenAI's)
- Voice interfaces (no native voice mode)
Gemini (Google)
Google's flagship, with deep integration into Workspace. The 2M-token context window is genuinely useful for analysing entire codebases, long videos, or huge document collections in one shot.
Wins on:
- Massive context (2M tokens — by far the largest)
- Tight integration with Gmail, Docs, Sheets, Drive
- Multimodal: strong on video and audio understanding
- Pricing for high-volume use cases
Loses on:
- Polished consumer experience compared to ChatGPT and Claude
- Plugin / developer ecosystem (catching up but still behind)
- Consistency — outputs can vary more than the other two
What about Microsoft Copilot?
Copilot for Microsoft 365 is essentially GPT-4-family wrapped in Microsoft's enterprise controls and embedded across Word, Excel, Outlook and Teams. If your organisation has already standardised on M365, Copilot is the path of least resistance for personal-productivity AI. We rarely recommend it as the foundation for a custom build, because you're paying for the M365 surface area you don't need.
The enterprise tier difference (and why it matters)
The single biggest reason Australian businesses get this choice wrong is conflating consumer ChatGPT with the enterprise APIs.
The free or consumer-tier products can — and sometimes do — use your data to improve models. The enterprise tiers (OpenAI Enterprise, Anthropic, Google Vertex, Azure OpenAI) contractually do not train on your data. They also offer data residency options, audit logs, SSO, and admin controls that the consumer products don't.
For any client work — and especially anything touching customer or employee data — we always build on enterprise tiers. Cost difference is modest. Compliance difference is enormous.
Pricing comparison at typical SMB scale
API pricing is in constant flux and has trended down 5–10× year-over-year since 2023. Per-million-token pricing as of Q3 2026 (USD, rounded to the nearest dollar — convert at current rates):
| Model | Input ($/M tokens) | Output ($/M tokens) | Best for |
|---|---|---|---|
| GPT-4-class | $2–$5 | $10–$15 | Tool use, multimodal |
| Claude (top tier) | $3–$5 | $15 | Reasoning, writing, code |
| Gemini (Pro tier) | $1–$3 | $5–$10 | High-volume, long-context |
| Open-source (Llama 3.x via host) | $0.20–$0.80 | $0.40–$1.20 | Volume tasks, cost-sensitive |
For a typical SMB build (voice agent or workflow automation), monthly API spend is usually under $200/month — sometimes well under. The cost calculation that matters isn't the model's price per token; it's whether the build delivers an outcome worth more than the all-in cost.
Which to pick for…
Customer support
Claude for triage and reply drafting (judgement, tone, knowing when to escalate). GPT for image-heavy queries (returns, defects). Open-source for high-volume classification (intent routing).
Sales
Claude for account research and personalised outreach drafting. GPT for image-rich outbound (LinkedIn). Either for call summarisation and CRM updates.
Content marketing
Claude as the default writer. GPT for image generation alongside. Gemini if your team works in Google Docs and wants in-editor assistance.
Code and engineering
Claude as the default — particularly for code review, refactoring and complex feature work. GPT for cases where tool integrations are central. Open-source for high-volume code-completion at scale.
Internal Q&A over private documents
Claude for the reasoning quality. The vector store and retrieval layer matters more than the model choice at this level — pick a model your engineers know well.
Voice agents
GPT for the inference layer, ElevenLabs for the voice synthesis. Claude is workable but the OpenAI Realtime API has the lower-latency voice path right now.
The honest meta-answer
The frontier models are converging. In Q3 2026, the difference between Claude, GPT and Gemini on any individual task is rarely more than 10–15% in real-world quality. The choice that matters more is how well your build:
- Grounds the model in your actual data (RAG)
- Filters and validates outputs before they reach users
- Hands off to humans when confidence drops
- Logs every call for audit and improvement
- Lets you swap models without rewriting the application
A vendor-neutral build that can swap from Claude to GPT to Gemini in a config change is worth far more than picking the perfectly optimal model today. Models will change. The integration layer is what compounds.
Key takeaways
- Claude is our default for reasoning, writing and code. ChatGPT is our default for tools, plugins and multimodal. Gemini is our default for huge-context and Workspace-native work.
- Always build on enterprise tiers — not the consumer products — for any work touching customer or employee data.
- API costs are small. Outcome matters; pricing comparisons rarely matter at SMB scale.
- Build vendor-neutral. The right model today may not be the right one in six months.
- The model choice matters less than the surrounding integration: retrieval, guardrails, hand-offs, logs.
Want help putting any of this into practice?
30-minute discovery call. We'll map your highest-leverage AI opportunities to revenue or time saved — no obligation.