Two new models, two different jobs
OpenAI has released GPT-6 Sol and GPT-6 Luna in its API. The OpenAI announcement places the release in the 22 September section and names the two model identifiers, gpt-6-sol and gpt-6-luna. Both are reasoning models that accept text and image input and generate text through the Responses and Chat Completions APIs. The practical choice is not simply which model has a higher benchmark score, but which one matches the complexity and frequency of an application’s work.
OpenAI describes Sol as suited to complex coding and agentic workflows, while Luna is its most efficient option for focused, high-volume tasks. A product might use Sol to investigate a difficult software defect and Luna to classify a large stream of incoming requests. Treating every step as the same model call can waste money or leave a hard decision underpowered.
This is a direct API launch. Availability in ChatGPT Work and Codex, and availability through cloud partners, are related but separate deployment surfaces. Developers should check the identifier, endpoint and account access that apply to their application rather than assuming all three surfaces share the same controls.
What the published prices mean
For standard short-context usage, OpenAI lists Sol at US$2 per million input tokens, US$0.20 per million cached input tokens and US$10 per million output tokens. Luna is listed at US$0.10 input, US$0.01 cached input and US$0.50 output. Those are large differences, but a model that requires more retries or produces an unusable answer can still be the more expensive choice overall.
The published short-context tier covers prompts up to 272,000 input tokens. OpenAI’s model page says requests above that threshold are billed at twice the input and cache rates and one-and-a-half times the output rate for the full request. A large context window therefore should not be read as a flat-price invitation to send an entire document store on every call.
Prompt caching is another conditional saving. It rewards repeated stable material such as instructions or schemas, but only if the application actually achieves cache hits. Developers should record cache reads, new input, output, latency and retries for a complete task. Australian teams should also check currency conversion, contracted pricing and any regional-processing premium before comparing local costs.
Capabilities and endpoint boundaries
The Sol model page lists a 1,050,000-token context window and 128,000 maximum output tokens. These are ceilings, not guidance that the best result will come from filling every available token. Retrieval and careful context selection remain useful when the source material is noisy or changes frequently. The quality test is whether the model identifies the right evidence and follows constraints.
Sol supports built-in tools and function calling through the Responses API. OpenAI’s documentation says Chat Completions function calling for Sol requires reasoning_effort set to none. That detail can affect a migration from an older chat-completions integration: code that assumes tool calls work identically at every reasoning setting may need an endpoint change or a different request configuration.
The model documentation also lists image input but not audio or video input for Sol. Teams building a voice or video application should not infer direct support from the broad GPT-6 label. They may need a separate transcription or media-processing step, with its own latency, privacy and cost implications.
How to evaluate the pair
A fair trial should include common short requests, long documents, coding tasks and cases that need tool use. Give each model the same task and score the final result rather than only the first response. Track factual errors, format failures, human correction time and whether any required tool step was skipped. For routing, test a simple Luna-first process with escalation to Sol when uncertainty or task complexity warrants it.
Safety and permissions remain application responsibilities. A stronger model may request more ambitious tool actions, but it should receive only the authority needed for the user’s task. Keep consequential actions behind explicit approval and record enough evidence to review what the system did. Price savings are not useful if they come from loosening oversight.
The release gives developers a wide economic range within one model family. Sol appears designed to make substantial recurring work affordable, while Luna makes narrow work cheap enough to run often. The right architecture will emerge from task-level measurements: quality, throughput, total spend and how reliably each model behaves within the application’s guardrails.
Migration planning should include a small rollback mechanism. Keep the previous model identifier available, preserve test prompts and compare changes under the same service tier. If a downstream parser expects a particular structure, test those outputs explicitly. If a model is used for advice, inspect where it expresses uncertainty and whether its citations can be checked. Pricing and model specifications can change, so pin the versions and assumptions used in the initial business case. This gives a team a defensible basis for expanding the rollout and a way to diagnose regressions without guessing whether the model, prompt or surrounding application changed.