If you are adding function calling to gpt-6-sol or gpt-6-luna, choosing between Chat Completions and the Responses API is more than an SDK preference. According to a September 23 write-up of the release published by CometAPI, it also decides whether the model can reason while it chooses and sequences tools. Below, what the source reports is kept apart from what follows from it, and the gaps it leaves are listed at the end.
What the source reports
The article summarizes OpenAI’s changelog, model comparison and launch post. Those OpenAI pages are not quoted directly here, so treat each point as the article’s account of them.
- OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, and both support the Responses and Chat Completions APIs.
- The constraint is stated plainly: “Use Responses for reasoning with built-in tools and function calling. Chat Completions supports ordinary requests, but function calling for Sol and Luna requires reasoning_effort=“none”.”
- Sol’s reasoning effort is listed “from none through max”. No explicit range is given for Luna, though one Luna result is reported “at max effort”.
- Sol targets work that “requires sustained reasoning, tool decisions, and multiple steps”, such as repository analysis, debugging and agent-driven software changes. Luna targets extraction, classification, support triage and first-pass summaries, “provided outputs are checked against the application’s requirements.”
Why this is an architecture boundary
The inference: on these models, a Chat Completions tool loop runs every function-calling turn at reasoning effort none. If your loop earns its value by deciding which tool to call next, recovering from a bad intermediate result, or carrying a plan across steps, it is running without the capability Sol is positioned for.
The published numbers do not close that gap. Every agentic score in the article is at elevated effort: Sol at 33.2% on AutomationBench at xhigh, Sol at 68.8% and Luna at 66.6% on DeepSWE v1.1 at max, and Sol at 60.5% on OSWorld 2.0 offline at xhigh. No result is reported at effort none, and the article does not say which API surface or tool harness produced the figures. It also cautions that “the best model can change with reasoning effort, tool use, and task design.” Those benchmarks therefore say little about a Chat Completions tool loop at effort none.
The article’s routing pattern points the same way. It sends routine work to Luna and demanding cases to Sol, and lists “tool requirements” among the escalation signals. A request escalated to Sol because it needs tools, but left on Chat Completions with function calling, still reaches Sol at effort none: the escalation changes the model without turning reasoning on.
A decision rule
These are recommendations, not measured outcomes:
- Keep Chat Completions for shallow, checkable tool calls: a single lookup, a classification that selects one function, or a schema-bound extraction validated downstream. That matches the Luna-style work described above. There is no published quality evidence at effort none, so confirm it on your own prompts instead of borrowing the benchmark figures.
- Move to Responses for loops where the model must choose among several tools, complete multi-step execution, or debug its own intermediate output. That is the work the article assigns to Sol, and Responses is the surface it names for reasoning with function calling.
- Make the API surface part of the routing decision. When a router escalates on “tool requirements”, switch the surface together with the model.
What to measure before switching
The article names the failure modes to evaluate: “incorrect tool choice, incomplete multi-step execution, schema failures, factual errors, latency, and cost per accepted result.” Run the same representative prompts on both surfaces and compare exactly those. In the article’s words, “A small representative evaluation set is more actionable than a broad leaderboard score.”
For CometAPI’s OpenAI-compatible route, the article’s advice is conditional: if the dashboard documents Responses and chat support for the selected route, “test it separately before rollout”. It does not confirm Responses support for either model there. It also recommends logging “provider, route, model ID, reasoning setting, latency, tokens, validation result, and billed cost.” For this decision the reasoning-setting field matters most, because it separates a Chat Completions run at effort none from a Responses run with reasoning enabled. The general point holds for any gateway: which API surfaces a route supports for a given model is a per-route fact, not something inherited from OpenAI’s direct release.
What the source does not answer
- What happens when a Chat Completions request combines function calling with effort above none: an error or a silent override.
- Which built-in tools Responses offers and what they cost; tool fees appear only as an unquantified cost input.
- Whether the restriction also applies to GPT-6 Astra.
- Differences in conversation state, streaming, parallel tool calls or structured-output handling between the two APIs.
- Any measured tool-call quality difference between the surfaces.
Check OpenAI’s own changelog and model documentation for these before committing a production tool loop to either surface.