OpenAI officially fully opens the GPT-5.6 series model to the global market, launching three products at once: the flagship Sol, the balanced Terra, and the cost-effective Luna. These models will be gradually available for ChatGPT, Codex, and API within 24 hours. The pricing of the three models are 5/30 dollars, 2.5/15 dollars, and 1/6 dollars (per million tokens input/output), with the lowest-tier Luna being only 1/16th the price of Anthropic's flagship model Fable5.

image.png

Performance surpasses Claude, programming and cybersecurity become main battlegrounds

The core narrative of the GPT-5.6 release is one word: cost-effectiveness. In the Agents' Last Exam evaluation across 55 industries, Sol scored 53.6, exceeding Claude Fable5 by 13.1 percentage points, and even with medium-level reasoning it still outperformed by 11.4 points, while the cost was only about a quarter of the competitor's. More surprisingly, Terra and Luna, which are positioned lower, still exceeded Fable5 with a cost as low as 1/16th. Programming is Sol's main battlefield, where it set a record of 80 points in the Artificial Analysis programming intelligence index, outputting less than half the tokens of Fable5, and significantly reducing time and cost.

However, GPT-5.6 is not entirely dominant. On SWE-Bench Pro, Sol scored 64.6%, still below Fable5's 80% and Mythos5's 80.3%; on the most challenging FrontierMath Tier4 math problems, Sol's score of 65.9% even fell short of the previous generation GPT-5.5's 72.5%, while Fable5 reached 87.8%. These shortcomings indicate that in deep engineering and frontier mathematics fields, Anthropic still holds an advantage. GPT-5.6 also introduced an ultra mode, capable of coordinating up to 16 agents working in parallel, setting a new record of 92.2% in the BrowseComp test, and its ExploitBench score in cybersecurity rose sharply from 47.9% to 73.5%.

GLM-5.2 from China emerges unexpectedly, forming a tripartite rivalry

In this AI arms race, Chinese players have not been absent. The open-source model GLM-5.2 released by Zhipu AI in mid-June has reached international advanced levels in programming and long-term tasks, supporting a million-token long context. According to reports from overseas tech media The Verge, several researchers believe that GLM-5.2's performance in some vulnerability detection and cybersecurity testing scenarios is comparable to Anthropic's Mythos, showing that Chinese AI models are rapidly closing the gap with the US in specialized fields. When Fable5 was shut down globally due to US export controls, GLM-5.2's stable open-source strategy provided developers with a reliable alternative, making the three-power rivalry more intriguing.

The full release of GPT-5.6 marks a new phase in AI development: model capabilities are no longer the sole benchmark. What truly matters to enterprises is how much substantive work can be accomplished within the same budget. As OpenAI uses a "downward strike" pricing strategy to squeeze competitors, Anthropic maintains its barriers in deep reasoning, and Zhipu enters a differentiated track with an open-source approach, the "three kingdoms" battle of large AI models is entering a heated stage.