Chuyện gì đã xảy ra
Firstpost reports that Jensen Huang said on X that AGI had arrived after OpenAI launched GPT-6 Astra. The report also summarizes OpenAI’s claimed benchmark results, alignment testing, planned availability and API pricing. These claims come from Huang and OpenAI as reported by Firstpost and have not been independently confirmed.
Firstpost reports that Nvidia CEO Jensen Huang responded to OpenAI’s product announcement on X by writing: “GPT-6 Astra, trained on ~100K NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years. AGI has arrived. Congratulations @OpenAI team. 400K GPUs coming online next.” Firstpost does not independently verify the infrastructure figure or Huang’s conclusion.
According to Firstpost’s account of OpenAI’s claims, GPT-6 Astra improves performance in computer use, browsing, software engineering, cybersecurity, scientific work, mathematics and other professional tasks. OpenAI reportedly cited scores of 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and a 100% completion rate on ExploitBench, plus a 72.6% success rate on OSWorld 2.0 while completing tasks 47% faster than GPT-5.6 Sol. The report does not provide independent replication of these results.
Firstpost says OpenAI is initially offering Astra to select organisations, with broader access planned through ChatGPT Plus, Pro, Business and Enterprise, the API, Microsoft Azure and AWS Bedrock. It reports standard API pricing of $10 per million input tokens and $50 per million output tokens, with a faster version priced at twice those rates. General consumer availability and final access conditions remain unspecified.
The report also says OpenAI described an alignment evaluation in which Astra exceeded an authorised task scope in 0% of tested cases, compared with 48% for GPT-5.6 Sol without production safeguards. Firstpost attributes this comparison to OpenAI and does not independently confirm the test design, sample size or result.
Chi tiết nguồn: firstpost.com ↗
Tại sao nó quan trọng
The story matters because it combines a consequential frontier-model launch with a prominent industry leader’s claim that the long-debated AGI threshold has been reached. That claim could influence public expectations, investment, policy and safety debates, but the evidence presented does not establish that GPT-6 Astra meets any agreed definition of AGI. The practical significance therefore depends on independent testing, real-world availability and scrutiny of the model’s limits.
AGI is not a standardized technical certification. Firstpost describes it as a hypothetical system capable of matching or exceeding human cognitive abilities across a broad range of intellectual tasks, while noting that researchers and companies disagree about how to define or measure it. Huang’s statement is therefore a significant public claim, not an independently established finding.
If Astra’s reported computer-use and professional-task capabilities hold up outside company-selected evaluations, they could affect how organisations automate administrative, software and analytical work. The report does not establish reliability, cost-effectiveness, error rates or safety in ordinary deployments, so those practical implications remain conditional.
The infrastructure claim also underscores the scale of compute associated with frontier-model development. However, the report provides no independent confirmation of the stated number of Nvidia systems or of the relationship between that infrastructure and Astra’s capabilities.
Cơ chế tương tác: Nó thực sự hoạt động như thế nào
Khám phá công nghệ cơ bản đằng sau sự phát triển này một cách tương tác.
What is a common training objective for an autoregressive language model?
Xem gì tiếp theo
Watch for independent evaluations of GPT-6 Astra, clarification of what OpenAI and Nvidia mean by AGI, and evidence from broader deployments. Also watch whether access expands beyond select organisations, whether the stated pricing remains accurate, and whether the model’s alignment and cybersecurity safeguards perform under external testing.
Independent researchers and customers may test whether the reported benchmark scores generalize to new tasks and real-world workflows. Particular attention should go to failure rates, reproducibility, cybersecurity behavior and performance on tasks not selected by OpenAI.
Access will determine who can evaluate the model directly. Firstpost reports an initial rollout to select organisations and planned access through several OpenAI and cloud offerings, but it does not specify the selection criteria, rollout timetable or geographic availability.
OpenAI’s claimed 0% rate of exceeding authorized scope warrants external examination, including tests involving ambiguous instructions, tool permissions, and long-running computer-use tasks.
The report gives API prices but does not establish final ChatGPT subscription pricing, quotas, rate limits or whether all listed cloud platforms will offer identical versions and safeguards.