A new place in the model line-up
OpenAI released GPT-6.1 Sol on 29 September, describing it as a model for complex coding and professional work at a lower cost than GPT-6 Astra. The launch creates a middle option for teams whose tasks stretch beyond simple, high-volume processing but do not require the most expensive flagship model for every request. Its significance is less about a single benchmark than about the purchasing decision it changes: organisations can now test a newer Sol against Astra on difficult real work, with a potentially substantial price difference.
The company’s model page describes near-Astra performance, but that phrase is vendor positioning rather than a guarantee for a particular application. A coding assistant, document workflow or browser agent may stress different capabilities. Buyers should compare the two models on representative tasks, including failure handling, tool use and the cost of retries. A cheaper first attempt is not necessarily a cheaper completed job if it needs more intervention.
Published specifications and rates
For prompts up to 272,000 input tokens, OpenAI lists standard prices of US$2 per million input tokens, US$0.10 per million cached input tokens, US$2.50 for cache writes and US$10 per million output tokens. The model page lists a 1,050,000-token context window and a maximum of 128,000 output tokens. Those limits enable long tasks, but they should not be read as a recommendation to fill the entire window indiscriminately. Long prompts can increase latency, cost and the burden of checking whether the model used the right evidence.
The pricing page adds important qualifications. Prompts above the 272,000-token threshold attract higher rates for the full request, while faster processing costs more and Batch or Flex can cost less. That makes prompt length and service tier part of deployment design, not merely billing details. Teams with large repositories or document sets should measure how much context each job actually needs, then compare total completed-task cost across models and tiers.
Tooling and operational choices
GPT-6.1 Sol is available through the Responses and Chat Completions endpoints, but OpenAI directs developers to Responses for tool calling. The model page says Chat Completions is supported without tool calling. This matters to anyone upgrading an established agent pipeline: changing the model identifier alone may not preserve the same tool behaviour if the application uses an older endpoint or assumes previous reasoning settings.
Reasoning effort supports low, medium, high, xhigh and max, with medium the default. The none and minimal settings are not supported. OpenAI also lists web search, file search, computer use, hosted shell and other tools when the model is used through Responses. Teams should confirm which tools are genuinely enabled in their project and test any permission boundaries. A model’s capability list is not an authorisation policy for production systems.
Multi-agent and residency deserve separate tests
The 29 September changelog identifies multi-agent support as a beta. It lets a model delegate work to subagents within a Responses request, potentially helping with tasks that can be divided into research, implementation and checking. Beta status makes evaluation especially important: a distributed workflow can increase parallelism, but also creates more intermediate outputs to audit and more ways for assumptions to drift. A useful trial records which subtask was delegated, what came back, and whether the final answer reconciled conflicting findings.
OpenAI says the model supports US and EU data residency, although fast mode is unavailable with EU residency. Residency and processing tier therefore need to be checked together rather than as separate tick-boxes. Australian organisations with contractual or regulatory requirements should verify their own data-flow obligations before choosing a region. The published availability statement does not itself answer every question about logs, tools, connected services or where external data is retrieved.
What to evaluate next
A disciplined trial would sample jobs at several complexity levels, compare Sol with Astra and the prior Sol release, and score the actual deliverable rather than a short model response. Track accepted changes, human review time, tool-call failures, time to completion and the token bill. For document-heavy work, include tasks that test whether citations refer to the right passages; for coding, include tests and security review rather than relying on code that merely looks plausible.
The release gives developers a compelling new price-performance candidate, not a universal migration instruction. The strongest case will come from workloads where Sol reaches the required quality reliably while costing less across the entire workflow. The official changelog provides the 29 September launch date; the model documentation provides the current technical and pricing details. Both should be revisited before a production commitment because limits, availability and commercial terms can change. A staged rollout with clear rollback criteria can expose hidden regressions before they affect customers. Teams should record the model identifier, reasoning effort, processing tier and prompt version for every comparison so later results can be reproduced.