This Is a New Tier, Not an Astra Replacement
OpenAI has released GPT-6 Sol and GPT-6 Luna as two faster and more affordable models below GPT-6 Astra. They are not aimed at exactly the same work: Astra remains the choice for the most demanding and important projects, while Sol and Luna bring professional work, coding, factuality, and computer-use capabilities to more frequent tasks. OpenAI says the two models were trained with methods similar to Astra, but are intended to distribute frontier capability across lower-cost tiers.
The important change is not simply that the model family has gained two more names. The constraint is shifting from whether a model can complete a task to whether it can complete that task repeatedly at a sustainable cost. An agent may need several attempts, tool calls, checks, and corrections before it produces a usable result. In that setting, the highest single-run score does not determine whether the system is deployable. A cheaper model that supports more iterations can change the default routing policy rather than merely serving as a budget compromise.
The Price Cut Reflects Serving Economics, Not Just a Discount
OpenAI attributes the efficiency gains to serving-side improvements in caching and inference. Cached input reads can receive discounts of up to 90%, which matters when an agent repeatedly processes the same project context or long instruction set. The practical question is not only how much one response costs, but whether an application can turn stable context, tool definitions, and project state into reusable input.
The pricing evidence is straightforward:
| Model | Input price | Output price | |---|---:|---:| | GPT-5.6 Sol | $4 per million tokens | $20 per million tokens | | GPT-6 Sol | $2 per million tokens | $10 per million tokens | | GPT-5.6 Luna | $0.20 per million tokens | $1.20 per million tokens | | GPT-6 Luna | $0.10 per million tokens | $0.50 per million tokens |
These comparisons are against GPT-5.6 promotional pricing. The announced 50% reduction should not be read as a permanent halving against every previous price baseline. Engineering teams still need to model cache hit rates, input-output ratios, tool calls, and retries together. Otherwise, a lower token price can be consumed by longer agent trajectories.
The Agent Benchmarks Point to Cheaper Useful Attempts
On AutomationBench, GPT-6 Sol scores 33.2% at xhigh effort. The reported task cost is about one thirty-ninth of GPT-6 Astra at low effort and one eleventh of Claude Opus 5 at maximum effort. It also exceeds Claude Fable 5.1 at 31.4%, but that comparison has an important gap: Opus 5 fallbacks occurred on roughly 40% of tasks, and the published cost does not include them. The result supports Sol’s cost efficiency, but it does not establish that every real deployment will produce the same total bill.
On Agents’ Last Exam, Sol at maximum effort scores 56.4%, below Claude Opus 5’s highest reported score of 60%, while costing 60% less per task. That is a more operationally relevant signal than asking only which model has the highest score. If a workflow can retry failures, or if a system must process a large volume of moderately complex work, lower cost per attempt may be more valuable than a few extra percentage points at the ceiling. A website redesign task described in the material reflects the same preference: the model used native page transitions instead of introducing React, then checked desktop, narrow mobile, and browser-back behavior. The goal was a shippable result, not a larger stack.
Luna’s Value Appears When Volume Becomes the Constraint
Luna is positioned more like a high-frequency, low-cost foundation layer. The material reports a 5.4 percentage-point improvement over its predecessor at high effort while reducing cost per task by 58%. On the factuality evaluation, Luna can also match GPT-5.6 Sol at roughly one hundredth of its cost. That does not make Luna suitable for every complex task. It suggests that classification, triage, batch processing, and lightweight agent work that was previously too expensive to run continuously may now have a different economic boundary.
Sol is more likely to become the default routing tier. Its results on professional and complex agent tasks can cover work that previously required a more expensive model, while higher effort levels provide a way to trade cost for quality. An application could use Luna for low-risk, high-throughput work, route more demanding requests to Sol, and reserve Astra for high-value or ambiguous cases. This is not a promised automatic-routing feature from OpenAI. It is a deployment judgment based on the published capabilities and prices.
The Factuality Gains Still Have a Narrow Evidence Base
OpenAI reports that GPT-6 Sol makes about half as many mistakes as its predecessor on an internal factuality evaluation and approaches Astra-level reliability. GPT-6 Luna also improves substantially, reaching the level of GPT-5.6 Sol at higher effort. This matters for agent systems because one factual mistake can trigger an incorrect tool call, an incorrect business update, or a chain of later steps built on unreliable information.
The evidence has a defined boundary. The evaluation uses de-identified conversations in which users had previously flagged factual errors, so the sample is not representative of ordinary usage. Scores are also not controlled for response length. “About half as many mistakes” should therefore be read as progress on a targeted error-inducing sample, not as a claim that factual errors have been halved across all conversations. Teams changing their routing policy should still measure factual errors, tool misuse, fallback rates, and human review cost on their own domain data.