Insights

The $20 AI seat is being pulled apart

A year ago, office AI cost about $20 a month per person. You paid GitHub Copilot or Google One, got an assistant, and the bill was predictable. That model is breaking.

On 1 June 2026, GitHub Copilot switched from flat-rate requests to token-based billing. Every plan now includes a monthly pool of AI Credits: 1,500 for Pro ($10 USD), 7,000 for Pro+ ($39 USD), 20,000 for Max ($100 USD). Users on GitHub’s community forums reported burning through 57 percent of their monthly Pro allowance in a single CLI session. A greeting message consumed 9.39 credits (GitHub Community Discussion #197089, June 2026).

Google moved in the opposite direction. At I/O 2026, it cut Google AI Plus from A$12.99 to A$4.99 while doubling storage to 400 GB (Google AI plans page, verified October 2026). Gemini 3 Pro access at that price undercuts every coding assistant on the market.

Ollama launched per-token cloud pricing on 31 August 2026, replacing its old GPU-time model. Pro costs US$20/month with US$60 of included credits. Running DeepSeek V4 Pro costs US$1.32 per million input tokens and US$3.96 per million output tokens. GPT-OSS 120B is US$0.15 input, US$0.60 output (Ollama pricing page, verified 20 October 2026). Credits do not roll over. No hard cap; you keep going at published rates once the allowance runs out.

LM Studio’s Bionic+ costs US$20/month and adds US-hosted open models (Kimi K3, DeepSeek V4 Flash, GLM 5.3) with discounted bulk tokens. Cloud pricing runs from US$0.13 per million input tokens for DeepSeek V4 Flash to US$3.00 for Kimi K3 (LM Studio pricing page, verified October 2026). Local inference through the desktop app remains free.

Three pricing models, one decision

Assistant tier (US$10-40/month): GitHub Copilot and Google AI. Cheap, familiar UI, broad model access. Copilot’s credit pool means heavy users will overspend unless someone watches the meter. Google’s AI Plus comes with Google’s data terms. Neither tier lets you choose which model handles most queries.

Agent tier (US$20-100/month): Ollama Cloud and LM Studio Bionic. You pick the model, pay per token. More predictable cost per task because you control which model runs. These are US-hosted. For a business under Australian data residency obligations (the Privacy Act, DISP requirements, or state procurement standards), sending operational data to a US endpoint may not work.

Self-hosted (hardware + electricity): A used NVIDIA RTX 4070 costs A$800-1,100 on the AU secondary market. Electricity adds A$8-15/month at Perth rates. Data stays on the hardware, no per-token bill. Trade-off: smaller models do less. A 7B parameter model will summarise and draft but will not match GPT-OSS 120B or Kimi K3 on reasoning.

Where the money goes

For a 10-person Perth business using AI for email drafting and document summaries:

TierAnnual per seatData stays onshore
GitHub Copilot Pro~A$180No
Google AI Plus~A$60No
Ollama Cloud Pro~A$370No
LM Studio Bionic+~A$370No
Self-hosted (amortised yr 1)~A$180Yes

(AUD at USD/AUD ~1.55, before tax)

Google AI Plus costs least. Self-hosted is the only option that keeps data onshore. Which matters depends on the work: someone triaging invoices does not need Kimi K3. Compliance reporting probably wants a model above 30B parameters.

That flat-rate AI seat from mid-2025 is gone. Each vendor now prices on different logic: credits, tokens, or hardware. For an Australian SME, the question is not which tool is cheapest. It is which billing model you can predict and which data jurisdiction you can defend.

Sources