GPT-6 Sol and Luna API Prices: Version Names, Cache Hits and Cost Boundaries

Identify the exact service request

OpenAI documents GPT-6 Sol under the literal API name gpt-6-sol, while a newer page and the live tariff also mention gpt-6.1-sol. This distinction matters for billing, reproducibility and historical reporting. Never rewrite a stored gpt-6-sol call into gpt-6.1-sol simply because a new model has appeared. Likewise, gpt-6-luna is a separate SKU rather than a discount setting on Sol.

For this 25 October 2026 snapshot, capture the literal model identifier returned by the API or specified in the request, together with any explicit deployment or processing tier. Release and retirement dates need vendor-issued lifecycle notices; a page that says a model is newer does not establish its launch date.

Documented starting point for Sol

OpenAI’s GPT-6 Sol model page quotes Standard short-context rates of $2 per million uncached input tokens, $0.20 cached input, $2.50 cache writes and $10 output. A cached-input rate is not the same thing as a cache-write fee: repeated prompts may initially cost more because the prefix must be written. Output can dominate an agent task when generated explanations are long. For context beyond 272,000 input tokens the model reference notes whole-request rate changes.

Pricing tables found at different URLs may expose gpt-6-sol or gpt-6.1-sol. Keep each as a distinct version; do not interpolate prices between them or claim that their cached-read multipliers are identical.

Luna illustrates why token volumes matter

The pricing page lists GPT-6 Luna Standard short-context rates of $0.10 for input, $0.01 for cached input, $0.125 for a standard cache write and $0.50 for output, all per million tokens. That rate card is relevant to high-volume lightweight work, not evidence that Luna can replace a more capable model for every prompt. Selection belongs to technical evaluation, outside this team’s remit.

For a purely hypothetical 2 million ordinary input tokens and 100,000 output tokens, short-context Luna rates produce $0.20 + $0.05 = $0.25, excluding all tools, cache writes and taxes. Do not multiply a monthly aggregate by a short-context rate if any constituent request crossed the threshold.

Build a version-safe ledger

Use columns for the requested model ID, resolved model ID if disclosed, response timestamp, input, cached input, cache-write tokens, output, processing mode and region. Group a daily report by these fields before attaching a price identifier. A dashboard showing only “GPT-6” is incapable of explaining the bill or responding to a model retirement. Version aliases, as opposed to pinned snapshots, can change over time; store the literal API metadata.

Where release evidence is incomplete, mark lifecycle dates null. The current model page is evidence about the named product’s posted rates, not a complete immutable history of every week in 2026.

Separate subscriptions and tool charges

ChatGPT consumer entitlements do not automatically fund the API. Search tools, hosted execution, storage and long-context handling may be separately billed. If a team is switching between Sol and Luna as a routing strategy, compute total spend from actual requests and include failures, retries and human review, then evaluate output quality with another team’s benchmark protocol.

Region-specific residency can also carry a published premium and eligibility constraints; avoid applying one uplift automatically to every model or every data centre. Reconcile with invoices and preserve the original USD list-price currency for this snapshot.

What can still change

The public pricing page and product pages are mutable. The exact initial effective date of the displayed tariff is not stated, and entitlement to specific services may differ by organisation. This makes the record useful as a date-checked commercial snapshot, not as a guaranteed 2020–2025 historical schedule. For long-running workflows, capture a new snapshot whenever a model identifier, deployment region or tier changes.