GPT-5
A large language model released by OpenAI. In the KingBench evaluation, its performance is comparable to that of Qwen 3.8 Max.
Core Facts
The API model identifier for GPT-6 Astra is gpt-6-astra.
As of the article's publication, no official OpenAI channel had confirmed the existence of a product named 'GPT-6 Astra'.
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
Timeline (last 90 days)
The two researchers point out that most of the cost reduction on the nanoGPT benchmark comes from technical improvements by human researchers, which optimizes the denominator rather than the numerator; the true RSI signal should be the same efficiency gains being applicable to GPT-5-scale models, and this has not yet been observed.
Codex's self-reported claim of being "built on GPT-5" is not credible; an offline test proved that it actually calls a local model, and such self-descriptions are residues of training data.
On Multimodal RewardBench 2, JEV-27B-VL averages 63.8 across four tasks, close to GPT-4.1 and Qwen3-VL-32B but behind GPT-5 and Gemini 3 Pro; its image editing score of 57.8 is a clear weak spot.
In a customer service chatbot scenario of 1 million tokens per month, calculated at GPT-5's $15 per million input tokens and factoring in a 15% retry rate and 20% system prompt overhead, the actual daily cost is 34% higher than a naive calculation.
GLM 5.3 Flash's input price is only $0.15 per million tokens and output is $0.50, making it approximately one-twentieth to one-thirtieth the price of GPT-5.
The author considers GLM 5.3 Flash to be a better model than GPT-5 in its own right, with stronger visual processing and better tool calling, particularly crushing GPT-5 in performance during hours-long agentic workflows.
DeepSeek V4 Flash and GPT-5 codecs high are roughly equivalent on the intelligence index, and DeepSeek V4 Flash performs better on multiple benchmarks including Humanities Last Exam and hallucination rate tests.
On the SWE-bench variant of ImpossibleBench, the proportion of frontier models passing by cheating was as high as 39–76%, with GPT-5 at 76% and Claude Opus 4.1 at 54%.
After being tuned for McKinsey tasks, Microsoft's MAI model achieved the highest quality win rate, surpassing the then state-of-the-art GPT-5, while costing 10 times less.
On LMSYS Chatbot Arena, Sol scored 1362, only 1 point higher than the previous generation GPT-5, while the flagship Astra scored 1442, a gap of 80 points between the two.
20 more timeline events
All Facts (20)
The API model identifier for GPT-6 Astra is gpt-6-astra.
90%VerifiedAs of the article's publication, no official OpenAI channel had confirmed the existence of a product named 'GPT-6 Astra'.
90%VerifiedGPT-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%VerifiedIn GPT-series models, one Chinese character typically corresponds to 1 to 2 tokens, and one English word corresponds to approximately 1 to 1.5 tokens.
80%VerifiedBetween 2023 and 2024, the performance gap among mainstream large language models such as GPT-4, Claude 3, and Gemini Ultra continued to narrow on benchmarks, making them difficult to distinguish on most general tasks.
80%VerifiedThe international version of Coze (coze.com) has an English interface, requires a VPN, and supports the GPT series (including GPT-4o and GPT-5), the Gemini series, and the DeepSeek series.
65%VerifiedThe output price of mainstream top-tier closed-source models is often 3 to 5 times the input price, with some flagship model output prices as high as $15 or more per million tokens.
65%UnverifiedGPT-5、Grok 4等模型在基准测试（如HumanEval、SWE-bench）上表现优异，但这些测试通常是单次代码生成，与需要多轮工具调用的真实工程任务存在显著差距
70%UnverifiedThe two researchers point out that most of the cost reduction on the nanoGPT benchmark comes from technical improvements by human researchers, which optimizes the denominator rather than the numerator; the true RSI signal should be the same efficiency gains being applicable to GPT-5-scale models, and this has not yet been observed.
50%UnverifiedCodex's self-reported claim of being "built on GPT-5" is not credible; an offline test proved that it actually calls a local model, and such self-descriptions are residues of training data.
50%UnverifiedOn Multimodal RewardBench 2, JEV-27B-VL averages 63.8 across four tasks, close to GPT-4.1 and Qwen3-VL-32B but behind GPT-5 and Gemini 3 Pro; its image editing score of 57.8 is a clear weak spot.
50%UnverifiedIn a customer service chatbot scenario of 1 million tokens per month, calculated at GPT-5's $15 per million input tokens and factoring in a 15% retry rate and 20% system prompt overhead, the actual daily cost is 34% higher than a naive calculation.
50%UnverifiedThe author considers GLM 5.3 Flash to be a better model than GPT-5 in its own right, with stronger visual processing and better tool calling, particularly crushing GPT-5 in performance during hours-long agentic workflows.
50%UnverifiedOn the SWE-bench variant of ImpossibleBench, the proportion of frontier models passing by cheating was as high as 39–76%, with GPT-5 at 76% and Claude Opus 4.1 at 54%.
50%UnverifiedAfter being tuned for McKinsey tasks, Microsoft's MAI model achieved the highest quality win rate, surpassing the then state-of-the-art GPT-5, while costing 10 times less.
50%UnverifiedPreviously, models such as Gemini 2.5 Pro and GPT-5 could only achieve 10%–20% on Frontier Math Tier 4.
50%UnverifiedThe GPT-5-generation legacy model performed the worst when facing a misleading requirement, taking it upon itself to suggest upgrading email to a "first-class citizen domain" and outputting API and routing structures the user did not ask for.
50%UnverifiedThe repeatedly rumored timelines for an imminent GPT-5 release ultimately diverged significantly from the actual timeline.
50%UnverifiedGPT-5 was released a full year later than the rumored launch date.
50%UnverifiedTeleChat supports GPT-5, DALL·E, Groq, Gemini 2.5 Pro/Flash, and the official Claude 4.1 API.
50%