GPT-5.6 Luna
A lightweight, high-speed model in OpenAI's GPT-5.6 series, balancing inference speed and cost efficiency, priced at $1 per million input tokens and $6 per million output tokens.
Timeline (last 90 days)
After the V7 platform integrated GPT-5.6 Luna, the hallucination rate dropped by 38%.
After the V7 platform integrated GPT-5.6 Luna, costs were cut by 78%.
Table 1 of the paper shows that ARPO (a pure-text 8B agent without latent-space mode) scores 38.8 on GAIA and 32.0 on WebWalker, both slightly higher than GPT-5.6 Luna.
According to Table 1 of the paper, HybridAgent 8B scores 43.7 on the GAIA benchmark, surpassing GPT-5.6 Luna's 37.9.
Official model card scores for Qwen3.8-27B: GPQA Diamond 89.2, LiveCodeBench v6 90.3, SWE-bench Pro 61.7, Terminal Bench 2.1 73.0; Artificial Analysis Intelligence Index score of 52, on par with GPT-5.6 Luna, and just one point below GLM-5.2 (753B) and DeepSeek V4 Pro (1.7T).
GPT-6.1 Sol and GPT-5.6 Luna are the only models that scored 0.9375 on both tasks, that is, 15 out of 16 problems, and did so in two independent samplings.
In Copilot quality scoring, Microsoft-Decision-1 is on par with GPT-5.6 Luna and about 100 times faster (Microsoft's statement).
The Copilot team used Decision-1 to measure the quality of chat and agent responses, claiming its performance is competitive with GPT-5.6 Luna at approximately 100 times the speed.
With two to three attempts, GLM-5.3 Flash scored 82%, which is within the margin of error of GPT-5.6 Luna's 81%.
On the server side, Jev can output 12 to 14 decisions in the time it takes Luna to generate one decision, and Jev processes all decisions in parallel.
12 more timeline events
All Facts (15)
The tool names GPT-5.6 Luna and DeepSeek-V4-Flash may be in community discussion or forward-looking speculation stages and may not correspond to official releases
60%UnverifiedFlagship or experimental models like GPT-5.6 Luna represent a vendor's capability ceiling and excel at complex reasoning, long-context understanding, and multimodal processing
60%UnverifiedAfter the V7 platform integrated GPT-5.6 Luna, costs were cut by 78%.
50%UnverifiedAfter the V7 platform integrated GPT-5.6 Luna, the hallucination rate dropped by 38%.
50%UnverifiedTable 1 of the paper shows that ARPO (a pure-text 8B agent without latent-space mode) scores 38.8 on GAIA and 32.0 on WebWalker, both slightly higher than GPT-5.6 Luna.
50%UnverifiedAccording to Table 1 of the paper, HybridAgent 8B scores 43.7 on the GAIA benchmark, surpassing GPT-5.6 Luna's 37.9.
50%UnverifiedOfficial model card scores for Qwen3.8-27B: GPQA Diamond 89.2, LiveCodeBench v6 90.3, SWE-bench Pro 61.7, Terminal Bench 2.1 73.0; Artificial Analysis Intelligence Index score of 52, on par with GPT-5.6 Luna, and just one point below GLM-5.2 (753B) and DeepSeek V4 Pro (1.7T).
50%UnverifiedGPT-6.1 Sol and GPT-5.6 Luna are the only models that scored 0.9375 on both tasks, that is, 15 out of 16 problems, and did so in two independent samplings.
50%UnverifiedIn Copilot quality scoring, Microsoft-Decision-1 is on par with GPT-5.6 Luna and about 100 times faster (Microsoft's statement).
50%UnverifiedThe Copilot team used Decision-1 to measure the quality of chat and agent responses, claiming its performance is competitive with GPT-5.6 Luna at approximately 100 times the speed.
50%UnverifiedIn benchmark comparisons, except for Banking 77 (bank intent classification), Jev can generally match or surpass Luna, with a particularly notable lead on guardrail-related tasks (prompt injection and toxic detection).
50%UnverifiedOn the server side, Jev can output 12 to 14 decisions in the time it takes Luna to generate one decision, and Jev processes all decisions in parallel.
50%UnverifiedIn terms of input token counts, both Astra and Luna used 16 tokens, while Sol and Terra used 26 tokens.
50%UnverifiedThe similarity in token usage patterns between Astra and Luna may hint at shared commonalities in their tokenizers or underlying architectures.
50%UnverifiedSimon Willison uses GPT-5.6 Luna to automatically generate a summary of key points for each change, as he does not trust Claude to summarize its own system prompts due to a conflict of interest.
50%