GPT-5.6
A next-generation large language model series released by OpenAI, comprising three capability tiers named Sol, Terra, and Luna. The series was made available in preview form via API on June 26, 2024.
Core Facts
GPT-5.5是OpenAI于2025年中期发布的模型，相比GPT-4o在推理深度和上下文理解方面有显著提升
Under the standard framework of ARC Prize, GPT-6 Astra's ARC-AGI-3 score is only 62.7.
V4.1-Flash scores 1 point higher than a GPT-5.6-level model in comparison.
Timeline (last 90 days)
Meta's "8161" model surpasses Opus and GPT-5.6 with a score of 77.0 in the evaluation data published by the model itself.
These PHP vulnerabilities were discovered by a model at the GPT 5.6 level, while the Wiz team independently discovered the same vulnerabilities.
The performance degradation caused by lost formatting cues is reproducible across six answer models, covering five models across the Claude Opus, Sonnet, and Haiku tiers plus OpenAI's GPT-5.6.
The competitors in the official comparison table are GPT-5.6 and OPUS-5, while these two companies released OPUS-5.5 and GPT-6.1 on September 22 and September 29 respectively, neither of which entered the comparison table.
According to third-party estimates, the input price of DeepSeek V4.1 Flash during weekday daytime is about one-thirteenth to one-seventeenth that of GPT-5.6 and Claude Opus 5, and the output price is one-seventeenth to one-twenty-first.
The gateway operated by the original author resells the same models at 10–25% below vendor official prices (GPT-5.6 at $1.70/$10.20 versus the official $2/$12), but this gateway has no batch endpoint and does not support embeddings.
ChatGPT (GPT-5.6) was relatively cautious, but still guessed from the product name that Mayank makes "no-code tools".
In William Liu's experiment, across all 31 test samples, the token counts reported by the seven GPT models were completely identical, all at 44,794.
William Liu compared the tokenization results of multiple GPT models using a set of fixtures, with the experiment covering seven GPT models, including GPT-5.5, three variants of GPT-5.6 (Sol/Terra/Luna), and three variants of GPT-6 (Astra/Sol/Luna).
In GPT-5.6 and later, OpenAI charges 1.25× for cache writes and 0.1× the uncached input rate for cache reads, with a minimum cacheable prefix of 1,024 tokens.
40 more timeline events
All Facts (20)
GPT-5.5是OpenAI于2025年中期发布的模型，相比GPT-4o在推理深度和上下文理解方面有显著提升
90%VerifiedUnder the standard framework of ARC Prize, GPT-6 Astra's ARC-AGI-3 score is only 62.7.
65%VerifiedV4.1-Flash scores 1 point higher than a GPT-5.6-level model in comparison.
65%VerifiedOpenAI previewed an Ultra Fast speed service mode in the API, enabling GPT-5.6 to achieve up to a 14x speed improvement with an output throughput of 750 tokens per second.
75%UnverifiedThe new model simultaneously achieves higher quality and lower cost, breaking the traditional performance-cost trade-off.
70%Partially VerifiedThese PHP vulnerabilities were discovered by a model at the GPT 5.6 level, while the Wiz team independently discovered the same vulnerabilities.
65%UnverifiedLlama 3.1 405B's score on MMLU is already close to the level of GPT-4o, and Qwen 2.5 72B has surpassed GPT-4 Turbo on certain metrics in math and code generation benchmarks.
60%UnverifiedMeta's "8161" model surpasses Opus and GPT-5.6 with a score of 77.0 in the evaluation data published by the model itself.
50%UnverifiedThe performance degradation caused by lost formatting cues is reproducible across six answer models, covering five models across the Claude Opus, Sonnet, and Haiku tiers plus OpenAI's GPT-5.6.
50%UnverifiedThe competitors in the official comparison table are GPT-5.6 and OPUS-5, while these two companies released OPUS-5.5 and GPT-6.1 on September 22 and September 29 respectively, neither of which entered the comparison table.
50%UnverifiedAccording to third-party estimates, the input price of DeepSeek V4.1 Flash during weekday daytime is about one-thirteenth to one-seventeenth that of GPT-5.6 and Claude Opus 5, and the output price is one-seventeenth to one-twenty-first.
50%UnverifiedThe gateway operated by the original author resells the same models at 10–25% below vendor official prices (GPT-5.6 at $1.70/$10.20 versus the official $2/$12), but this gateway has no batch endpoint and does not support embeddings.
50%UnverifiedIn William Liu's experiment, across all 31 test samples, the token counts reported by the seven GPT models were completely identical, all at 44,794.
50%UnverifiedIn GPT-5.6 and later, OpenAI charges 1.25× for cache writes and 0.1× the uncached input rate for cache reads, with a minimum cacheable prefix of 1,024 tokens.
50%UnverifiedThe author previously used GPT-5.6 to create a Minecraft-style game, spending more than 5 hours repeatedly fine-tuning to optimize the frame rate to 60 FPS, with API costs exceeding $350.
50%UnverifiedGPT-6's error response rate has dropped to 4.1% and its false claim rate to 0.9%, an improvement over GPT-5.6's 6.3% and 1.8% respectively.
50%UnverifiedPollo AI integrates three models—GPT-5.6, GPT-6 Astra, and GPT-Image-2.5—handling different stages of creative concepting, image generation, and final video production respectively.
50%UnverifiedEvoX Agent supports automatic model switching by difficulty level, pairing economy tier with cheaper models such as DeepSeek, balance tier with GPT-5.6 or Kimi K3, and also supports connecting to local models and LM Studio and Ollama.
50%UnverifiedFor Amazon Bedrock, GPT-6 and GPT-5.6 are now routed through inference profiles, and the India regions ap-south-1 and ap-south-2 will correctly resolve profiles with the in. prefix.
50%UnverifiedCompared to GPT-5.6, the price of OpenAI's new model has dropped by approximately 50%.
50%