OpenAI introduced GPT-6.1 Sol on 29 September, pitching stronger coding and office-work capabilities at standard API prices of US$2 per million input tokens and US$10 per million output tokens. For businesses considering AI agents, the question is whether those rates translate into cheaper work that people can actually use.
The company describes the model as offering “Near-Astra intelligence for a fifth of the price”. That comparison is with its more expensive GPT-6 Astra model. OpenAI says Sol matches Astra on its reported DeepSWE software-engineering evaluation at roughly one-fifth of the cost, but those results are vendor-reported benchmarks rather than a guarantee for a particular business.
The comparison with the previous Sol model is different. GPT-6 Sol already lists the same US$2 input and US$10 output rates. GPT-6.1 therefore combines claimed capability gains with unchanged headline token prices, while cutting the cached-input rate from US$0.20 to US$0.10 per million tokens. Describing the launch as a general price cut would miss that distinction.
The published model specifications also set limits around those headline rates. Requests exceeding 272,000 input tokens attract higher charges, and processing tiers can change pricing. Tokens are the pieces of text a model processes or generates; a business pays for usage across a workflow, not simply for giving an agent a task.
An illustrative workload of one million ordinary input tokens and 100,000 billable output tokens would cost US$3 at the standard rates. That calculation assumes each request stays below the long-context threshold and excludes cache writes, tools and taxes. It is a token-cost example, not a forecast of what an hour of coding or a completed customer-service job will cost.
Repeated context offers another potential saving for agents working through related tasks. OpenAI’s caching documentation explains that reusing a matching prompt prefix can reduce subsequent input charges, but writing that content to the cache also costs money. A coding workflow that repeatedly uses the same instructions could benefit; constantly changing inputs may produce a different result.
The wider bill includes services beyond the language model itself. OpenAI’s pricing schedule separately charges for tools such as web search, file search and code-execution containers. Businesses also need to account for integration and human review when assessing whether an agent saves money. A cheaper attempt offers little value if staff repeatedly have to repair its output.
Australian access and Australian-dollar pricing are separate considerations for local buyers. Australia appears on OpenAI’s supported countries list, while its billing guidance says API-credit purchases remain limited to US dollars. The rates above are therefore not an A$ retail quote; account-specific GST treatment was not established in the materials reviewed.
For an Australian software team or operations department, the useful comparison is the cost of an accepted result. A trial could track completed tasks, review time, retries and total charges against the existing workflow. If Sol delivers comparable results with less spending or correction, lower model costs could make more routine work economical to automate; the benchmark claim alone cannot establish that saving.
