GPT 5.6 Sol, Terra, Luna Cheat Sheet: Features, Pricing at a Glance, and Comparison with Claude – Which is Better?

OpenAI has released the GPT-5.6 series of models. This article will walk you through the Sol, Terra, and Luna model family’s highlights, billing, and API costs. However, in response to the Trump administration’s security review requirements, the preview is currently limited to a limited number of specific partner organizations.

GPT-5.6 Goes Live in a Flash—Access Restricted by Trump Administration Requirements

On June 26, OpenAI unexpectedly unveiled its new generation of GPT-5.6 model series, including the flagship model Sol, the balanced model Terra, and the high-value model Luna.

However, due to the Trump administration’s requirements for security review of cutting-edge AI models, OpenAI currently offers limited preview versions only to a small number of trusted partner organizations, and they are not yet available in ChatGPT; access will be expanded gradually in the future. Even so, the surprise release of GPT-5.6, OpenAI’s preview offerings, Sol, Terra, and Luna has still become a popular search keyword in the tech industry.

The Three Main Differences in GPT-5.6 Models: Sol, Terra, Luna

The GPT-5.6 series has made major adjustments to its naming system, abandoning commonly used names from the past such as nano or mini. OpenAI states that the number in the new naming system represents the model generation, while Sol, Terra, and Luna represent the corresponding independent and persistent capability tiers, with the goal of helping enterprises and developers make clearer choices among intelligence performance, speed, and cost.

Although, for people in the crypto community, Sol is likely to bring to mind the Layer 1 blockchain Solana ($SOL), and Terra and Luna may remind them of the namesake project Terra ($LUNA) that has gone defunct, these three models have nothing to do with blockchains or cryptocurrencies. In Latin, Sol means the moon, Terra refers to the Earth, and Luna is the moon.

These three models are designed for different enterprise needs:

• Sol is the top-tier flagship option, built for the most challenging tasks such as complex reasoning, extended coding, advanced agent-driven workflow orchestration, and cybersecurity defense
• Terra strikes a balance between strong performance and efficiency, making it suitable for production environments that need to handle large volumes of work (such as customer support, internal tools, and document analysis) while aiming to control costs
• Luna is the lightest and most cost-effective option in the family, optimized for speed and everyday routine automation tasks.

GPT-5.6 Sol, Terra, Luna Billing: Differences in API Costs

In terms of pricing, GPT-5.6 is billed according to three model sizes, at a rate per 1 million tokens. The detailed pricing and positioning are shown in the table below (mobile users can swipe the table left or right):

| Model | Positioning | Cost per 1M Token Input | Cost per 1M Token Output | | --- | --- | --- | --- | | GPT-5.6 Sol | Flagship model, suitable for advanced reasoning, agents, and security research | 5 USD | 30 USD | | GPT-5.6 Terra | Balanced enterprise model | 2.5 USD | 15 USD | | GPT-5.6 Luna | High-efficiency, low-cost model | 1 USD | 6 USD |

GPT-5.6 API Price Comparison with Claude, Gemini, Grok, and More

In addition to publishing GPT-5.6 series pricing, VentureBeat has also compiled API pricing for current mainstream large language models (LLMs).

Looking at the overall market, GPT-5.6 Luna is positioned in the mid-to-low price tier: the input price is 1 USD per 1 million tokens, and the output price is 6 USD per 1 million tokens, for a total cost of about 7 USD—placing it between GLM-5.2 and Grok 4.3 (Low Context).

If you need higher reasoning capability, GPT-5.6 Terra’s total cost is about 17.5 USD, the same as GPT-5.4. The flagship model GPT-5.6 Sol keeps the same pricing as GPT-5.5—5 USD for every 1 million tokens of input and 30 USD for every 1 million tokens of output—bringing its total cost to about 35 USD, which is lower than Anthropic’s latest Claude Fable 5 / Claude Mythos 5 at 60 USD.

Mainstream AI Model API Price Comparison (per 1 million tokens), compiled by foreign media VentureBeat

| Model | Input | Output | Total Cost | | --- | --- | --- | --- | | MiMo-V2.5 Flash | 0.10 USD | 0.30 USD | 0.40 USD | | DeepSeek V4 Flash | 0.14 USD | 0.28 USD | 0.42 USD | | DeepSeek V4 Pro | 0.435 USD | 0.87 USD | 1.305 USD | | MiniMax M3 | 0.30 USD | 1.20 USD | 1.50 USD | | Gemini 3.1 Flash Lite | 0.25 USD | 1.50 USD | 1.75 USD | | Qwen3.7 Plus | 0.40 USD | 1.60 USD | 2.00 USD | | MiMo-V2.5 | 0.40 USD | 2.00 USD | 2.40 USD | | Grok 4.3 (Low Context) | 1.25 USD | 2.50 USD | 3.75 USD | | MiMo-V2.5 Pro (≤256K) | 1.00 USD | 3.00 USD | 4.00 USD | | Kimi-K2.6 | 0.95 USD | 4.00 USD | 4.95 USD | | GLM-5.2 | 1.40 USD | 4.40 USD | 5.80 USD | | GPT-5.6 Luna | 1.00 USD | 6.00 USD | 7.00 USD | | Grok 4.3 (High Context) | 2.50 USD | 5.00 USD | 7.50 USD | | MiMo-V2.5 Pro (>256K) | 2.00 USD | 6.00 USD | 8.00 USD | | Qwen3.7 Max | 2.50 USD | 7.50 USD | 10.00 USD | | Gemini 3.5 Flash | 1.50 USD | 9.00 USD | 10.50 USD | | Gemini 3.1 Pro Preview (≤200K) | 2.00 USD | 12.00 USD | 14.00 USD | | GPT-5.6 Terra | 2.50 USD | 15.00 USD | 17.50 USD | | GPT-5.4 | 2.50 USD | 15.00 USD | 17.50 USD | | Gemini 3.1 Pro Preview (>200K) | 4.00 USD | 18.00 USD | 22.00 USD | | Claude Opus 4.8 | 5.00 USD | 25.00 USD | 30.00 USD | | GPT-5.5 | 5.00 USD | 30.00 USD | 35.00 USD | | GPT-5.5 Instant | 5.00 USD | 30.00 USD | 35.00 USD | | Sakana Fugu Ultra | 5.00 USD | 30.00 USD | 35.00 USD | | GPT-5.6 Sol | 5.00 USD | 30.00 USD | 35.00 USD | | Claude Fable 5 / Claude Mythos 5 | 10.00 USD | 50.00 USD | 60.00 USD |

This time, OpenAI has adopted the Sol, Terra, and Luna naming scheme, making it easier for enterprises to select the right model based on capability, speed, and cost, rather than using model size as the primary differentiator.

GPT-5.6 New Technology: Deeper Reasoning and Sub-agent Collaboration

Beyond improvements to model capabilities, GPT-5.6 also introduces a new reasoning mode. According to OpenAI’s announcement, Sol adds a Max Reasoning mode that allows the model to spend more time on reasoning, improving the quality of complex task completion.

At the same time, OpenAI has introduced an Ultra mode, which improves efficiency for large-scale projects and long workflows by having multiple sub-agents (Subagent) collaborate to handle large tasks rather than having a single agent complete all work.

This design is mainly aimed at enterprise agent workflows. By splitting complex tasks and having multiple sub-agents work on them together, it can improve efficiency in long-duration reasoning, software development, and large project execution. This is also one of the key differences between GPT-5.6 and GPT-5.5.

Benchmarks Improve Across the Board—TerminalBench Hits New Highs

Based on the test results published by OpenAI, GPT-5.6 Sol achieved the best performance to date on the TerminalBench 2.1 command-line workflow test, surpassing GPT-5.5 and outperforming competing models such as Claude Mythos 5.

In addition, in the GeneBench biological research test, GPT-5.6 Sol achieved better results than GPT-5.5 while using fewer tokens.

Multiple tests show that Sol performs better than GPT-5.5 on benchmarks such as Agent’s Last Exam, TerminalBench, and ExploitBench. Terra also surpasses the previous generation flagship model in many workflow tests. Meanwhile, Luna maintains low costs while delivering capability performance close to GPT-5.5.

Prompt Cache Mechanism Updated—Cerebras to Provide Up to 750 Tokens/Second

OpenAI is also updating the Prompt Cache mechanism by adding an explicit Cache Breakpoint and providing at least 30 minutes of cache retention time. The official statement says that cache writes will be billed at 1.25 times the price of uncached input, while cache reads will retain a 90% discount—making it easier for enterprises to control token costs in agent workflows.

On the other hand, OpenAI also announced that GPT-5.6 Sol will be available on the Cerebras platform starting this July, reaching up to 750 tokens per second in inference speed. It mainly targets large enterprise applications that need low latency and high performance.

For high-volume repetitive workflows, the combination of the new caching mechanism and fast inference is expected to reduce overall compute costs.

Stronger for Enterprise Use, While Security Protection Is Also Upgraded

OpenAI says that the GPT-5.6 series adopts the most complete multi-layer security defense architecture available to date. This includes model-level refusal mechanisms, real-time biological and cybersecurity classifiers, account-level risk analysis, and ongoing red team testing, aiming to balance security defense needs while reducing malicious use.

The official statement indicates that GPT-5.6 Sol is better at helping identify vulnerabilities and patch weaknesses, but it has not yet reached the Cyber Critical level, where it can autonomously complete a full attack chain.

OpenAI invested about 700,000 A100 GPU hours to conduct automated red team testing for GPT-5.6, and also added mechanisms such as real-time risk detection, an Activation Classifier, and security reviews during reasoning time. However, because security research itself can be dual-use, OpenAI also admitted that even some legitimate work may still be affected by security checks.

GPT-5.6 Rolls Out in Phases, Reflecting a New Direction in AI Governance

This limited preview of GPT-5.6 also serves as an important case study for AI governance in recent times.

According to OpenAI’s explanation, before the official release, the company had already briefed the U.S. government on the model’s capabilities and release plans, and—per government requirements—it first provides the model to a small number of trusted partner organizations for testing before gradually expanding availability.

OpenAI says it hopes this government-first review approach is only a transitional measure, and it still expects to make the latest models more widely available to developers, enterprises, and general users in the future.

According to a report by Decrypt, the limited release this time is also related to the U.S. government’s recent establishment of a new evaluation framework for next-generation AI models. After Anthropic’s Claude Mythos 5 and Claude Fable 5 were restricted, GPT-5.6 became the second frontier AI model recently affected by the U.S. government’s release process.

Based on the recent situations of these two major players, when large AI models are launched in the future, government security reviews, corporate governance, and model capability evaluations may gradually become part of the release process.

• **Related Report:**Claude Fable 5 Banned and Removed! White House Adviser Reveals Inside Story of the Ban, Yann LeCun Criticizes It as Self-Inflicted Harm