GPT-6.1 Sol combines a hosted API model with Work and Codex access. The useful budgeting distinction is standard input versus cached input, while availability differs from Chat.
Access and rate cards are different questions
OpenAI lists gpt-6.1-sol for API access and $2 input, $0.10 cached input and $10 output per million tokens. The announcement describes availability in ChatGPT Work and Codex for paid and organizational tiers, while explicitly saying it is not yet in Chat. Its September 29 system-card addendum describes safeguards and evaluations. This is a hosted model; the announcement does not supply downloadable local weights.
Calculate the uncached baseline first
Our worked example uses 100,000 input and 10,000 output tokens. At the published standard rates, that is $0.30 before tools or other charges. If the entire input were eligible and billed as cached input, the arithmetic would be $0.11. These are scenarios, not a promise that a prompt is cached. Build the forecast from the provider usage records after testing, because a changed prefix or different request shape may alter actual reuse.
Repeated agent context is a useful place to inspect costs
Our suggested trial is a repository investigation with a stable instruction prefix and several narrow follow-up questions. Keep the task specification fixed, record cached and uncached input separately and save any corrections. Compare the bill with a single long request as well as a repeated-request workflow. The cheaper configuration must still preserve the evidence needed for the answer; token savings alone are not a useful definition of success.
Provider evaluations need their boundaries
When reading a coding or computer-use chart, note the tools, effort setting and scoring procedure before comparing it with your environment. Our recommendation is to write a small acceptance packet: one representative task, one awkward task and one task the agent should decline or escalate because information is missing. Keep the outputs and verification receipts. We have not run an independent model evaluation and do not infer reliability from the launch adjectives.
Budget the orchestration around the model
A practical agent budget also includes search, browser runtime, storage, retries and monitoring. Set a task-level spending boundary before granting execution tools. For teams, separate read-only investigation from externally visible actions and require a reviewable artifact before approval. These are our deployment recommendations, not a description of every product default. Read the Sonnet effort and cache article for another task-cost worksheet; comparing prices is only one part of selecting an execution workflow. An agent that cannot retrieve a necessary file should report that limitation rather than fill the gap with a plausible implementation story.
Frequently asked questions
Can I infer Chat availability from Codex access?
No. Check the announcement’s product-specific availability; it explicitly distinguishes those surfaces.
Can cached input prices apply to every token?
Only treat input as cached when the provider usage and billing show it. Keep an uncached baseline.
What else can an agent cost?
Tool calls, search, browser execution, persistence, retries and human review can all matter beyond the model’s token bill.
Sources & further reading
Primary sources checked Oct 11, 2026. Vendor statements are attributed; editorial advice is our own.
- 1
- 2
Help us keep this useful. Check contact availability →
