Agents, Cowork and deep research draw on extra, usage-based allowance on top of your subscription. Describe the job - see roughly what it costs. Every figure on this page is in New Zealand dollars, and when the meter starts differs by vendor.
Everything on this page is an estimate built from the published rates below. Change any number and the page recalculates. Your changes are saved in this browser only.
When the meter starts. This is the difference that catches people out, and it matters more than the per-token rates. On ChatGPT Business and Claude Team or Enterprise, your seat comes with included usage - a weekly or rolling allowance - and the costs on this page only begin once that allowance is spent. Plenty of people never reach it. Microsoft Copilot Cowork works the other way: the Copilot licence buys access, and the work itself is billed in Copilot Credits from the very first task, with no included allowance. Your admin may have pre-purchased credits, but those are still drawn down task by task. So the Copilot column is what you pay from day one, while the other two are what you pay after you have had your fill. One exception worth knowing: a Codex-only ChatGPT seat also bills from first use.
Notes: Claude extra usage is billed at API rates. The everyday model here is Claude Opus 5.5 at US$4 in / US$20 out - Anthropic's own guidance is that most workloads start with Opus 5.5, and it is the model built for the long-running agentic work Cowork does. The most powerful is Fable 5.1 at US$10 / US$50. OpenAI still publishes no official dollars-per-credit figure. The ~US$0.04 used here is the rate at which its credit prices match its own API prices exactly, so it's well grounded - but check your own workspace billing page before you rely on it. For ChatGPT, the everyday model is GPT‑6.1 Sol at 50 / 2.5 / 250 credits per 1M read / cached / written - OpenAI's recommended default in ChatGPT Work, at the same read and write rates as GPT‑6 Sol. The most powerful is GPT‑6 Astra at 250 / 25 / 1,250, five times the price. GPT‑6.1 Sol runs in ChatGPT Work and Codex but not in ChatGPT for Word, Excel or PowerPoint, where the cheaper choices are GPT‑5.6 Terra (50 / 5 / 300) and GPT‑5.6 Sol (100 / 10 / 500, a promotional rate held at least until 21 November 2026). Workspace Agents bill only at the GPT‑5.6 rate. On Enterprise and Edu, an admin has to switch GPT‑6.1 Sol on before staff can use it. Microsoft Copilot Credits are US$0.01 each pay-as-you-go, and Microsoft's Cowork planning bands are light 100-300, medium 300-700 and heavy over 700 credits - it calls these rough planning estimates. Prices are converted at the 30-day average USD/NZD rate (1.77 at this check) rather than the day's rate, so the page doesn't jump around with the currency. That is up from 1.74 in October, which adds about 2% to every price for reasons that have nothing to do with AI. The two columns no longer match. ChatGPT's everyday tier is half the price of Claude's for ordinary reading and writing, because the models sit at different points on each vendor's range. At the top the headline rates are identical, US$10 / US$50 either way, but cached reading splits them: Fable 5.1 charges US$0.25 per 1M against Astra's US$1.00, so repeat work over the same material is cheaper on Claude. Real jobs also differ in how many tokens each model uses, which this page doesn't model. Ranges are deliberately wide: the same agent job can vary several-fold between runs.