Introduction

On September 22, OpenAI rolled out two new models within its GPT-6 family: the mid-tier GPT-6 Sol and lightweight GPT-6 Luna. The most impactful announcement was not merely the launch of a new generation model, but a sharp price reduction. Compared with promotional pricing of the prior GPT-5.6 series, rates have been halved. Sol is priced at $2 per million input tokens and $10 per million output tokens, while Luna costs $0.1 per million input tokens and $0.5 per million output tokens.

When paired with the flagship GPT-6 Astra released earlier, OpenAI has built a three-tier model lineup. Astra handles the most complex, end-to-end high-stakes tasks. Sol takes charge of professional workflows and sophisticated AI Agent workloads. Luna targets high-volume, low-cost, repetitive execution tasks. This combination of price reduction and capability segmentation sends a clear market signal: competition in large language models is moving away from benchmark score races, and toward lowering the per-task operational cost of AI.

For engineering teams building multi-model production pipelines, managing access to different model endpoints, token quota control and request routing becomes a core operational burden. Treerouter, an API gateway, helps developers unify authentication and traffic management across multiple LLM endpoints.

1. GPT-6 Sol: Primary Workhorse for Complex Coding and Multi-step Workflows

GPT-6 Sol is more than a cheaper alternative to Astra. OpenAI positions it for professional work, code manipulation, computer operation, business automation and fact-based output. It is built for workflows requiring multi-round code reading, code modification, test execution, tool invocation and system write-back operations.

Key Capability Improvements

  1. Reduced factual error rate. OpenAI conducted factual evaluations based on real conversations and human-annotated error datasets. GPT-6 Sol produces roughly half as many factual errors as GPT-5.6 Sol, approaching the reliability level of Astra.
  2. Cost advantage for automation workflows. The AutomationBench benchmark includes 47 tools covering sales, operations, customer support, finance and HR scenarios. Sol scores 33.2% under the xhigh tier, with a per-task cost of $0.27. In identical test conditions, Astra’s low tier scores 30.3% but costs approximately 3.9 times more. Claude Opus 5 max scores 26.9% with a per-task cost around 11 times higher than Sol.
  3. Long context and Codex compatibility. Both Sol and Luna support a 1.05M token context window and a maximum output length of 128K tokens. They integrate with Responses and Chat Completions API, function calling and tool use. The models will gradually roll out across Codex, ChatGPT Workspace and GitHub Copilot products.

For software developers, the core value of Sol is not generating elegant standalone code snippets. Its strength lies in reducing trivial failures during complex requirements: reading large code repositories, tracing invocation chains, modifying multiple files, supplementing unit tests, and retrying after CI pipeline errors. When paired with Agent runtimes such as Codex, Sol works best as the default mid-tier model to complete end-to-end requirement delivery, rather than acting only as a code completion assistant.

2. GPT-6 Luna: Low-Cost High-Volume "Execution Worker", Not a Strategic Planner

Luna’s pricing is 1/20th of Sol: $0.1 per million input tokens and $0.5 per million output tokens. Cached input prompts can drop as low as $0.01 per million tokens. Its use cases are clearly defined: document summarization, field extraction, rule-based Q&A, batch classification, ticket pre-screening, scheduled report generation and lightweight iterative sub-steps inside Agent loops.

Early developer feedback shows Luna performs surprisingly well in some coding benchmarks. On DeepSWE v1.1 max tier, Luna reaches 66.6%, only 2 percentage points below Sol’s 68.8%. However, the per-task cost is merely a fraction of Sol’s expense. Even so, teams need to avoid misalignment in model selection. Luna is optimized for speed, low cost and high throughput. It is not designed for complex logical reasoning and high-stakes decision making.

OpenAI internal and third-party evaluation materials indicate Luna falls behind Sol in re-planning and long-horizon automation. On AutomationBench high-effort scenarios, Luna scores 14.5%, and its overall xhigh score is lower than Sol. It also generates more factual inaccuracies than Sol. Complex debugging and cross-system long-chain decision workflows should not rely solely on Luna.

A stable recommended architecture separates planning and execution: Sol handles planning and critical revisions, while Luna takes charge of extraction, formatting, batch write-back and low-frequency question answering.

3. Sol Is Not an All-Around Upgrade Over GPT-5.6 Sol

It is critical to clarify this point to avoid overstating model improvements. On selected benchmarks including factual accuracy, AutomationBench, Agents' Last Exam and FrontierCode, GPT-6 Sol outperforms GPT-5.6 Sol. However, it does not hold universal advantages in rigorous software engineering benchmarks.

For example, on DeepSWE v1.1 max tier, GPT-6 Sol scores 68.8%, which is lower than GPT-5.6 Sol’s 72.7%. In OSWorld offline high-tier tests, GPT-5.6 Sol can outperform the newer model in some cases. Artificial Analysis’s intelligence index supports the conclusion that the upgrade delivers economic gains rather than a massive intelligence leap. Sol scores approximately 48 and Luna around 37, representing modest gains from the prior generation.

Developers should not migrate all workloads to Sol simply because of the "6th generation + half price" selling point. Below are practical migration recommendations:

  • Large repository refactoring and long-running bug fixes: Run A/B tests across Sol, GPT-5.6 Sol and Astra. Compare PR acceptance rates and cycle time.
  • Business Agents, ticket processing, reports and cross-system RPA: Use Sol (xhigh/high) as the main model, with Luna for preprocessing and post-processing.
  • High-frequency summarization, extraction and intent classification: Deploy Luna max/high, paired with manual sampling audits.
  • High-liability judgment in healthcare, finance and legal domains: Do not switch to Luna for cost savings. Even Sol should serve only as an auxiliary tool, with final human review mandatory.

4. Behind The Price Cut: Cache and Inference Efficiency, Not Model Degradation

OpenAI attributes the price reduction to advances in caching and inference optimization, rather than simply shrinking model size. Several engineering improvements deserve close attention for production deployment.

  • Prompt cache hits deliver up to 90% discount on input tokens. Sol cached input drops to $0.2 per million tokens, and Luna cached input can reach $0.01 per million tokens.
  • Changes to reasoning effort, tool toggle settings no longer automatically invalidate prompt cache prefixes. System prompts, tool schemas and knowledge base summaries can be reused.
  • GitHub side disclosures show related caching optimization cuts the proportion of fresh prompt tokens needing recomputation by 50%. This is a key factor enabling Copilot-class products to reduce costs for long conversation sessions.
  • Requests exceeding the 272K input limit apply separate pricing for longer context windows. Teams evaluating total cost must look beyond base list pricing.

This means the true affordability of Agent batch workflows comes not only from halved token unit pricing. Reusable system prompts, tool definitions and historical context are the decisive factors. Without careful caching and prefix design, even cheaper Sol will quickly accumulate redundant token consumption and erase cost benefits.

5. Conclusion: The Price War Focuses on Per-Task Cost, Not Cheap Tokens

The combination of GPT-6 Sol and Luna essentially brings the reasoning, factual accuracy and tool invocation capabilities of the Astra generation down to mid and low price tiers. For enterprises, model selection will no longer center on "which model is the smartest". Instead, workloads will be decomposed and matched to appropriate models by task type.

  • Difficult, long-duration and high-responsibility tasks: Astra or retained GPT-5.6 flagship models as baselines.
  • Daily complex work, coding Agents and business automation: primary workload assigned to Sol.
  • Massive summarization, extraction, lightweight execution and low-latency Q&A: primary workload assigned to Luna.
  • Multi-Agent systems: Sol for planning, Luna for execution, cache reuse and human oversight.

When Sol can bring automated task cost down to $0.27 per task and Luna enables summarization and extraction at sub-cent pricing, AI evolves from an occasional assistant for developers into digital workers that enterprises can invoke hundreds of thousands of times. The core meaning of this pricing shift is not merely cheaper models. Enterprises can now purchase intelligence based on actual workload volume, rather than trial-based consumption.

Building such a multi-model system requires routing, access control and observability across multiple LLM endpoints. Centralized gateway services simplify operational overhead for teams running mixed model workloads.

Learn more：https://treerouter.com