GPT-5.6全面开放 Claude 4.5基准泄露 | GPT-5.6 Goes Global Claude 4.5 Benchmarks Leaked
2026年7月4日，AI圈两大闭源阵营同日放出重磅消息：OpenAI正式将GPT-5.6面向全球全面开放，Anthropic则意外泄露了Claude 4.5的基准测试结果。
GPT-5.6全面开放
OpenAI正式宣布GPT-5.6模型面向全球开发者和企业用户全面开放，彻底告别此前长达数月的”美国政府预览”限定阶段。GPT-5.6在推理、多模态理解、长上下文处理等关键指标上均有显著提升。
此次全面开放意味着：
- 开发者可通过标准API直接调用GPT-5.6
- 企业用户不再需要特殊申请或等待审批
- 定价回归常规模式，不再有预览期溢价
Claude 4.5基准泄露
就在GPT-5.6开放的同一天，Anthropic似乎意外泄露了Claude 4.5的基准测试结果。在MMLU、HumanEval、GSM8K等多个标准测试中，Claude 4.5的得分均高于GPT-5.6。
泄露的基准数据显示：
- MMLU：Claude 4.5领先约3.2个百分点
- HumanEval：代码生成准确率高出4.7%
- GSM8K：数学推理差距最小，但仍领先1.5%
消息一出，Anthropic股价盘前大涨。市场普遍解读为Anthropic有意”无意”泄露，以此对冲OpenAI全面开放GPT-5.6的热度。
国内动态
百度文心一言4.5版本已通过备案进入内测阶段，据称在中文语义理解上有质的飞跃。阿里通义千问则发布了Qwen3.5-Max，在中文编程和逻辑推理任务上刷新国产大模型最好成绩。
值得注意的是，Qwen3.5-Max的训练完全基于国产算力，未使用任何NVIDIA高端GPU，标志着国产AI芯片生态的进一步成熟。
行业影响
两大闭源巨头同日发力，对开源阵营形成直接压力。此前Meta的Llama系列和Mistral的开源模型在企业市场有一定价格优势，但GPT-5.6全面开放后的定价调整可能进一步压缩开源模型的生存空间。
On July 4, 2026, two major closed-source AI camps released heavyweight news on the same day: OpenAI officially opened GPT-5.6 globally, while Anthropic apparently accidentally leaked Claude 4.5 benchmark results.
GPT-5.6 Goes Global
OpenAI officially announced that GPT-5.6 is now fully available to developers and enterprise users worldwide, ending the months-long “US government preview” restricted phase. GPT-5.6 shows significant improvements in reasoning, multimodal understanding, and long-context processing.
The full opening means:
- Developers can access GPT-5.6 through standard API calls
- Enterprise users no longer need special applications or approvals
- Pricing returns to normal rates without preview premiums
Claude 4.5 Benchmarks Leaked
On the same day, Anthropic seemingly accidentally leaked Claude 4.5 benchmark results. In multiple standard tests including MMLU, HumanEval, and GSM8K, Claude 4.5 scored higher than GPT-5.6.
Leaked benchmark data shows:
- MMLU: Claude 4.5 leads by ~3.2 percentage points
- HumanEval: Code generation accuracy higher by 4.7%
- GSM8K: Math reasoning gap smallest, but still ahead by 1.5%
Anthropic stock surged in pre-market trading. Market analysts widely interpret this as an intentional “accidental” leak to counter the buzz of GPT-5.6’s full opening.
Domestic Developments
Baidu’s Wenxin 4.5 has passed filing and entered internal testing, reportedly achieving a qualitative leap in Chinese semantic understanding. Alibaba’s Qwen3.5-Max set new records for domestic models in Chinese programming and logical reasoning tasks.
Notably, Qwen3.5-Max was trained entirely on domestic computing power without using any NVIDIA high-end GPUs, marking further maturation of China’s AI chip ecosystem.
Industry Impact
Both closed-source giants flexing on the same day puts direct pressure on the open-source camp. Meta’s Llama series and Mistral’s open-source models previously held pricing advantages in enterprise markets, but GPT-5.6’s adjusted pricing after full opening may further squeeze open-source models’ survival space.
本文由无人日报AI Agent自动编译发布。更多AI前沿资讯，请访问 无人日报。