OpenAI’s October 2 GPT-6 model family guide sets out three choices for developers: GPT-6 Astra for the most demanding reasoning, GPT-6.1 Sol for workloads that balance capability and cost, and GPT-6 Luna for focused, repeated work at scale. It also separates model selection from reasoning effort and response speed, giving teams several ways to balance results, latency and cost. OpenAI’s model guide.
The central practical question is which combination fits a workload. OpenAI’s answer is to match the model and effort to the task, then measure success and latency. Its pricing and capability descriptions provide a starting point, but the accompanying safety evidence also shows why a cheaper model with similar headline capabilities should not be treated as interchangeable across every evaluation.
Astra, Sol and Luna serve different workloads
OpenAI positions Astra as its most intelligent model. OpenAI positions GPT-6.1 Sol as a lower-cost option for capable reasoning workloads, while Luna is intended for fast, efficient everyday work at scale. Those are vendor descriptions, rather than independent measurements of equivalent performance. OpenAI’s model guide.
| Model | Positioning in the supplied sources | Documented task examples |
|---|---|---|
| GPT-6 Astra | Highest-capability option; hardest reasoning work | Tasks requiring maximum reasoning capability |
| GPT-6.1 Sol | Near-Astra intelligence at a lower listed price | Complex coding, research and computer use |
| GPT-6 Luna | Focused, repeated work at scale | Invoice-field extraction and request classification |
The Sol task examples and description of Astra’s hardest reasoning role come from daily.dev’s breakdown; OpenAI’s guide supplies Luna’s examples. The distinction is useful because it organizes selection around the work being done. Repeated extraction with a clear goal and difficult debugging present different requirements, even if both sit inside the same application.
Version names also matter. OpenAI’s September 29 system-card addendum introduces GPT-6.1 Sol, while a separate Developer Community announcement discusses GPT-6 Sol and Luna. The guide’s Sol entry is specifically GPT-6.1 Sol. Availability statements about the earlier name should therefore retain that name rather than silently becoming promises about the newer version. GPT-6.1 Sol addendum, Developer Community announcement.
Reasoning effort is a separate decision
The guide describes reasoning effort as a choice of how much effort the model spends on a task in the API. Its recommendations progress from routine changes to demanding analysis:
| Reasoning level | OpenAI’s recommended use |
|---|---|
| Low | Extracting facts or making small edits |
| Medium | Planning a feature or comparing options |
| High | Difficult debugging, deeper analysis or careful review |
| Extra high / Max | Testing where supported when High falls short |
The qualification “where supported” matters: the supplied guide excerpt does not establish that every model exposes every level. It also supplies no API configuration payload for these settings. OpenAI’s model guide.
Response speed adds another variable. daily.dev covers Fast and Ultrafast speed modes. It says Ultrafast accelerates token generation independently of reasoning effort, is available for GPT-6 Astra in Codex and the API, and carries a premium over standard processing. The evidence does not specify that premium or quantify the acceleration. daily.dev’s breakdown.
For developers, the implication is that choosing a faster mode and choosing deeper reasoning address different parts of the workload. Their value still depends on the resulting task success, latency and cost.
What the guide’s prices establish
OpenAI lists the following prices in the supplied guide excerpt:
| Model | Input | Output | Cached input |
|---|---|---|---|
| GPT-6 Astra | $10 | $50 | $1 |
| GPT-6.1 Sol | $2 | $10 | $0.10 |
| GPT-6 Luna | $0.10 | $0.50 | $0.01 |
These are the guide’s listed figures. The supplied excerpt does not state their billing unit, so they cannot establish a cost per request or a complete production budget. Sol’s listed input and output figures are each one fifth of Astra’s; its cached-input figure is one tenth. That distinction qualifies the guide’s broad “a fifth of the price” description. OpenAI’s model guide.
Caching is another substantial cost variable. daily.dev reports that cached input tokens can cost up to 95% less than uncached tokens, depending on the model. Its breakdown covers prompt caching as a production practice. The guide also recommends caching and compaction to manage context and cost. daily.dev’s breakdown, OpenAI’s model guide.
The separate Community announcement reports 50% lower API prices for GPT-6 Sol and Luna relative to GPT-5.6 promotional pricing. That comparison uses a different baseline from the guide’s Astra-versus-GPT-6.1-Sol positioning. The announcement also describes Sol and Luna availability in the API and rollout in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, with Luna trials for Free and Go users in the desktop app. Its undated “today” wording does not establish a calendar rollout date. Developer Community announcement.
Agent workflows bring evaluation questions
According to daily.dev, GPT-6.1 Sol supports beta multi-agent workflows that delegate independent subtasks to subagents. Its breakdown also describes guidance on mid-turn steering, asynchronous tool calls, prompts, skills and AGENTS.md instructions. The supplied evidence establishes those topics, but provides no executable integration example or measured benefit from delegation. daily.dev’s breakdown.
OpenAI’s safety addendum supplies a more qualified picture of Sol’s capabilities. Under its Preparedness Framework, the company treats GPT-6.1 Sol as Critical in cybersecurity and High for Biological and Chemical capability, and says it uses Astra’s safeguards stack. These are OpenAI’s framework classifications. GPT-6.1 Sol addendum.
The supplied evaluation statements require care. One says GPT-6.1 Sol matches or exceeds GPT-6 Sol and Astra in “all categories,” without identifying that statement’s scope. Other statements report improvement over GPT-6 Sol in five of eight Production Benchmark categories and five of six under-18 categories, with a gore-evaluation regression described as statistically insignificant. HealthBench Consensus is also reported at 96.0. Those narrower results prevent interpreting the broad statement as superiority across every evaluation. GPT-6.1 Sol addendum.
A further trade-off concerns unwanted persistence: OpenAI reports it in 23.5% of GPT-6.1 Sol evaluation rollouts, compared with 17.4% for Astra. That is an evaluation result, not a production incident rate. Together, these findings make workload-specific testing central to the guide’s value: model positioning and listed prices can narrow the choices, while task success, latency and observed agent behavior determine whether the choice holds up. GPT-6.1 Sol addendum.
