Introduction
On September 23, 2026, OpenAI officially rolled out two new members of the GPT-6 family: GPT-6 Sol and GPT-6 Luna. These two models inherit the leading capabilities of GPT-6 Astra in professional work, factual reasoning, coding, computer use and safety alignment, while delivering these features at a much lower price point. The API pricing for both models has been reduced by 50% compared with promotional pricing of GPT-5.6. Sol is priced at $2 per million input tokens and $10 per million output tokens. Luna carries pricing of $0.10 per million input tokens and $0.50 per million output tokens. This article analyzes their product positioning, division of labor, benchmark performance, cache optimization strategies and user selection guidance.
GPT-6 Sol and Luna represent scaled industrial-grade models built after the Astra flagship, trained using methodologies similar to Astra. Where Astra is engineered to deliver peak intelligence performance, Sol and Luna function as industrial amplifiers deployed for mass adoption across vertical industries. The two models follow a clear work split: Astra serves scenarios requiring maximum performance, Sol acts as the primary high value-for-money workhorse, and Luna targets extreme throughput tasks. According to OpenAI’s official statement published on September 23, 2026, both models are live on the OpenRouter platform, with API identifiers gpt-6-sol and gpt-6-luna respectively.
The related ecosystem includes native OpenAI services, ChatGPT Work, Codex and OpenRouter. Comparative benchmark suites for evaluation include AutomationBench, FrontierCode and DeepSWE.
Pricing: Dual-Tier Strategy Behind the 50% Price Cut
The most immediate change in this release is pricing, with token costs halved from the previous GPT-5.6 generation.
| Model | Input per million tokens | Output per million tokens | Price reduction |
|---|---|---|---|
| GPT-5.6 Sol → GPT-6 Sol | $4 → $2 | $20 → $10 | 50% |
| GPT-5.6 Luna → GPT-6 Luna | $0.20 → $0.10 | $1.20 → $0.50 | 50% |
| GPT-6 Astra (Reference) | $10 | $50 | Flagship tier |
Data source: OpenAI official blog, September 2026.
Sol and Luna sit one magnitude apart in capability and pricing. Sol undertakes complex iterative workloads, while Luna handles high-volume, lightweight tasks. Data from OpenRouter in the same period shows OpenAI’s monthly consumption volume surpassed Anthropic for the first time in 30 months, and this low-price high-volume model lineup is a core driver of this shift.
The release also introduces major cache optimization. Agent workflows that reuse context can access a 90% discount on cached reads. GitHub reports show the proportion of freshly processed prompts has dropped by more than 50% over several months, which directly speeds up Copilot-style response generation.
Developers can call these models through standard compatible API interfaces. 4sapi, an API gateway service, has already onboard these two new models. It maintains compatibility with mainstream model API specifications, supports direct domestic access, and is suitable for combined workload architectures using Sol as the primary worker and Luna for batch processing tasks.
Professional Work: Sol Delivers Superior Performance at Only 9% of Opus 5 Cost
On AutomationBench business workflow benchmarks, GPT-6 Sol achieves a peak score of 33.2%, with a cost of only $0.27 per task.
- It exceeds Claude Opus 5’s maximum score of 26.9%, at just 9% of Opus 5’s task cost.
- It surpasses Claude Fable 5.1’s 31.4% benchmark result, with a cost less than one-ninth of Fable 5.1.
- It even outperforms the low-tier Astra score of 30.3%, consuming merely 1/3.9 of Astra’s cost.
On Agents Last Exam, Sol reaches 56.4%. This result approaches Opus 5’s highest score, but task cost falls by 60%.
Luna also brings meaningful upgrades. Its high-end benchmark score rises by 5.4 percentage points over the prior generation, while per-task cost drops by 58%. In terms of factual error rate, Sol produces approximately half as many misstatements as the prior generation, approaching Astra-level reliability. Luna’s high-end matching capability reaches the standard of GPT-5.6 Sol at roughly 1/100 of the cost.
Coding Performance: DeepSWE Benchmark Near Fable, Cost Cut by 80%
Coding represents the primary application domain for Sol and Luna. Internal OpenAI data shows median token consumption for coding workflows has exceeded $600, and the 90th percentile reaches $700. Low-cost iteration has become a critical requirement for development teams.
- FrontierCode: Sol delivers substantial gains against GPT-5.6 Sol, with peak performance matching Fable 5.1 while carrying far lower operational costs.
- DeepSWE v1.1: Sol’s maximum score hits 68.8%. It trails Fable’s top score of 69.9% by only 1.1 percentage points, at roughly 80% cost reduction.
- DeepSWE: Luna reaches a maximum score of 66.6%. This value sits within the mid-range results of Opus 5 and Fable, with cost reduced by 93% to 96%.
- OSWorld 2.0: Sol peaks at 60.5%, matching Opus 5’s mid-tier score of 60.3% at 80% lower cost. Luna surpasses mid-tier GPT-5.6 Sol performance with cost reduced to one-tenth.
Python API Invocation Example
Model names and API keys shown in the sample are placeholders. Production integration must refer to official OpenAI documentation.
Communication Style and Alignment: More Human Output, Lower Hallucination Rates
Sol and Luna inherit Astra’s communication characteristics. Their outputs are clearer, with fewer redundant phrases, unexpected logical jumps and finer granularity in details. In side-by-side comparisons based on Bento official website revision tasks, GPT-6 models first confirm whether React introduction is truly required and explain the reasoning. By contrast, GPT-5.6 models proceed directly to code implementation without validation.
In alignment testing, both models outperform their GPT-5.6 counterparts on coding deception, malicious search evasion and review bypass assessments. OpenAI notes these evaluations are deliberate adversarial stress tests. Failure rates for normal daily usage are far lower, and full details are available in the model system card.
Suitable Application Scenarios
- Sol for primary coding tasks: FrontierCode and DeepSWE benchmark results approach flagship levels. It fits long iterative development tasks in Codex-style workflows.
- Luna for batch processing: The $0.1 input token price makes it ideal for classification, information extraction and large-scale dataset collection.
- Business workflow Agents: AutomationBench delivers outstanding cost-performance. It works well for sales assistant and HR workflow orchestration.
Unsuitable Scenarios
For ultimate flagship-grade tasks, Astra remains the recommended option. The free ChatGPT desktop tier only supports Luna; Sol access requires Plus or higher subscription tiers.
The models are live within ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu plans. Free and Go tier desktop clients can run Luna. ChatGPT web access for these models is still pending full rollout, with 24/7 availability once released.
Frequently Asked Questions
Q: What is the biggest difference between GPT-6 Sol and Luna?
Sol targets medium-to-high complexity workloads, priced at $2 input and $10 output per million tokens. Luna is built for maximum throughput, priced at $0.10 input and $0.50 output per million tokens. Both cost 50% less than GPT-5.6, and their capability tiers sit at mid-high and entry levels respectively.
Q: Can Sol outperform Opus 5?
On AutomationBench, Sol achieves superior results at only 9% of Opus 5’s cost. Its DeepSWE result trails Fable by merely 1.1 percentage points. It fits the majority of daily workloads. For extreme high-stakes tasks, Astra remains preferred. Teams should run validation tests based on their own task datasets before full migration.
Q: How to integrate these models and what are their identifiers?
The API identifiers are gpt-6-sol and gpt-6-luna. Both are available on OpenRouter, ChatGPT Work and Codex with daily refresh cycles. Sol requires Plus or higher subscription tiers. Luna also provides an additional free desktop client quota.
Q: What are the rules for the 90% cache discount?
Enabling prompt cache requires diagnostic tool activation. Prompt tuning and tool toggle operations will not invalidate cached content. Explicit breakpoint controls are available at the front end, delivering the highest reuse gains for multi-turn Agent dialogues.
Q: How can developers conduct low-cost comparative testing?
Multi-model comparison can be completed through compatible API gateway platforms. Teams can run side-by-side tests for Sol, Luna and competing models, recording quality and per-task cost data before locking the primary production model.
Conclusion
GPT-6 Sol and Luna represent a major industrial scaling milestone. They bring Astra’s training methodology to market at Sol’s price and Luna’s throughput scale. Based on OpenAI’s September 2026 official data, their cost-intelligence curve leads the industry, and the consumption growth on OpenRouter reflects strong market adoption. This article uses data published in September 2026. Pricing and availability are subject to change; production decisions must refer to official OpenAI pricing documents.
Teams building multi-model agent pipelines often use unified API gateways to simplify routing, authentication and observability. 4sapi centralizes model endpoint management and supports switching between multiple LLMs within a single application stack.
International access: https://4sapi.com
Domestic access: https://4sapi.cn
