GPT-6 Sol vs Luna: Which Model Should You Actually Ship?

OpenAI shipped GPT-6 Sol and Luna at half the API price of the 5.6 series, landing roughly 90 minutes after Anthropic's Opus 5.5.

The pair completes the GPT-6 matrix under the Astra flagship, and the pitch is explicitly economic rather than a new capability ceiling. OpenAI frames them as two models that "bring frontier intelligence to everyday work with different balances of capability and cost."

What actually changed

The headline is pricing. Per TechCrunch, "the 6 series models will be available at half the cost of the 5.6 series of Sol and Luna, a price drop that OpenAI attributes to improvements in caching and inference." That attribution matters: the savings are architectural, not a promotional rate.

The second claim is accuracy. OpenAI's announcement states that on an internal factuality evaluation "based on de-identified real-world conversations where users flagged mistakes by our models, GPT-6 Sol makes about half as many mistakes as its predecessor, reaching Astra-level reliability at much lower cost." OpenAI also claims Sol and Luna handle tasks substantially better than Anthropic's Fable and Opus — a vendor comparison, published by the vendor, with no third-party run behind it yet.

A separate OpenAI post covers the caching layer doing the work: GPT-6 improves prompt caching with higher cache hit rates, new diagnostics, explicit breakpoints, and controls that reduce latency and cost.

Sol vs Luna vs Astra

Model	Built for	Where it fits
Astra	Most demanding reasoning, long-context analysis, high-risk professional tasks	Top-tier flagship; fallback for high-stakes work
Sol	Complex tasks like coding, multi-file refactoring, agent task chaining	Default primary model for agent pipelines
Luna	"High-volume tasks with a clear goal, like summarizing documents, extracting information, or answering quick questions"	Throughput layer for repetitive batch jobs

One developer guide reports both models keep million-token context windows and native tool calling, function invocation and structured JSON output, and that the pair directly phases out the older Terra line.

Where the cost claim breaks down

The same guide claims per-task cost can fall "to less than one-tenth of flagship competing models" — then immediately qualifies it. That figure "only holds under specific favorable runtime conditions," requiring well-matched effort levels and high cache hit rates for repeated context. In long sequential agent workflows, output token expansion narrows the advantage, and complex multi-step tasks can generate far more output tokens than benchmark testing estimates.

Read that as the practical rule: list price is not your bill. If your agent reuses a large static context — one repository, one contract, one reference corpus — the caching discount is real. If every call carries fresh context, you are buying a 50% token cut and nothing more.

Which one to ship

Route coding and multi-turn agent work to Sol, and push classification, extraction, summarization and support triage to Luna. The failure mode worth avoiding is the one both releases are priced against: sending every production request to a flagship because routing was never built. Before switching, measure your actual cache hit rate — that single number decides whether the price cut reaches your invoice.

Availability and what to watch

The new Sol and Luna are live in ChatGPT Work and Codex for most paid accounts and in the API, with Luna also reaching the desktop app and Free and Go users; the ChatGPT app and website roll out gradually.

Two things to watch. First, whether independent evaluations reproduce the halved-error claim, which so far rests entirely on OpenAI's internal factuality test. Second, whether the 90-minute gap between Opus 5.5 and this launch turns into a sustained price war — the competitive question here is total cost of ownership across millions of task runs, not another benchmark table.