OpenAI Releases GPT 5.6, Is It Any Good?

OpenAI just released GPT 5.6 on July 9th, as a successor to GPT 5.5, which was released last April. However, GPT 5.6 was rolled out to “trusted partners,” such as government employees and large companies, starting on June 26th, to comply with Executive Order 14409.

The Executive Order, titled "Promoting Advanced Artificial Intelligence Innovation and Security," focuses on making frontier AI labs give their models to the government before they are released, so the government can test them for safety purposes.

Anyhow, GPT 5.6 is split into 3 major tiers, Luna, Terra, and Sol, where Luna is the smallest (think moon), Terra is the medium (think earth), and Sol is the largest (think sun). All three models support a 1 million token context window and up to 128,000 output tokens in one response. They have a knowledge cutoff of February 16, 2026.

Let’s start with GPT 5.6 Luna. It’s OpenAI’s equivalent to Claude’s Haiku lineup, however it is significantly more powerful, scoring 51 points on the Artificial Analysis Intelligence Index, while Claude Haiku 4.5 scores only 23.7. A more fair comparison, considering that Claude Haiku was released in October 2025, is with Grok 4.5, which was released a day before, but surprisingly is priced very similarly (2$/m input tokens and 6$/m output tokens for Grok, and 1$/m input tokens and 6$/m for ChatGPT) scores a slightly higher 54.

Nevertheless, different AI models consume different amounts of token for the same tasks, so the best way to measure the cost effectiveness of a model is by using the Artificial Analysis Cost per Intelligence Index. According to that, Luna costs around $0.21 per task, and Grok costs around $0.31 per task, so Luna is actually cheaper when you use it. Overall, the winner at this price point is GPT 5.6 Luna, simply because it is a little cheaper.

Here’s a handy chart comparing GPT 5.6 Luna, Grok 4.5, and Claude Haiku 4.5:

GPT 5.6 Terra is OpenAI’s medium model, costing $2.50 for every million input tokens, and $15 for every million output tokens. It competes with Claude 5 Sonnet, which is $2/M Input, and $10/M Output until August 31, when it becomes $3/M Input, and $15/M Output. Other than Anthropic, no other major AI labs have models priced similarly. Terra scores 55.0 on Artificial Analysis’s Intelligence Index. However, Claude 5 Sonnet is a little worse, scoring 53.4. In terms of efficiency, Terra annihilates Sonnet, costing $0.55 per task, while Anthropics model costs $1.53 per task. Overall, GPT 5.6 Terra is the obvious winner from the two, as it is cheaper and more intelligent.

The real question on most people’s minds is simple: is GPT 5.6 Sol, the top tier AI model in the GPT 5.6 family, better than Claude Fable 5? Sol costs $5/M Input and $30/M Output, while Claude Fable 5 costs nearly double that, at $10/M Input and $50/M Output, so a comparison between these models isn’t exactly the most fair, however, if GPT 5.6 Sol can score close enough to Claude Fable 5 while being significantly cheaper, I’ll chalk it up as a win for OpenAI. GPT 5.6 Sol scores 58.9 on the Intelligence Index and costs $1.04 per task, while Fable 5 scores 59.9 on the Intelligence Index, while costing $2.75 per task. From personal experience, Fable 5 is a slightly better model, however the incredible price and the fact that the model isn’t included in the Pro or Max plans, meaning you have to pay for every million tokens individually, really makes using it ridiculously expensive. So, the winner is GPT 5.6 Sol, simply because Fable 5 has similar intelligence, but is too expensive to justify using it.

Overall, GPT 5.6 is a massive step change for OpenAI, and Artificial Intelligence in general. Their models are incredibly powerful, and, more importantly, more cost-effective when compared to competitors.