Newshot

Ahead of Anthropic's IPO, OpenAI Drops Two Models at Once: GPT-6.2 Astra and Sol Beat Fable 5.2 in Some Areas While Cutting Costs Sharply, Setting Records in Finance and Research

AI summary

On November 18, IT之家 reported that OpenAI suddenly and without warning delivered a "surprise" the day before Anthropic's IPO pricing. OpenAI today officially released the new GPT-6.2 Astra and GPT-6.2 Sol models, focusing on improving performance in professional work, coding, computer operation, and multi-step agent tasks while reducing the number of output tokens needed to complete tasks; in most tests they are not inferior to Anthropic's flagship model. OpenAI said GPT-6.2 Astra is a flagship model for complex professional projects, with further gains in professional work, scientific research, software development, and visual design. The model can more accurately follow presentation templates provided by users, improve the output quality of slides, spreadsheets, and documents, and enhance front-end development, website design, game development, and 3D design capabilities. In Agents' Last Exam, which tests multi-step professional work, GPT-6.2 Astra scored 5.2 percentage points higher than the previous generation GPT-6 Astra, with the cost per single task less than one-third of the latter's; it also scored 5.5 points higher than Claude Fable 5.2 while costing only 5% as much. In a

Translation(model translation, not the source's own text)

Anthropic IPO Eve: OpenAI Drops Dual Models GPT-6.2 Astra and Sol — Beating Fable 5.2 in Some Areas with Sharply Lower Costs, Setting Records in Finance and Scientific Research

IT Home, November 18 news: Just one day before Anthropic's IPO pricing, OpenAI suddenly, without any warning, delivered a "surprise."

OpenAI today officially released the latest GPT-6.2 Astra and GPT-6.2 Sol models, with a focus on improving performance in professional work, programming, computer operation, and multi-step agent tasks, while reducing the number of output Tokens required to complete tasks, holding its own against Anthropic's flagship model in most tests.

OpenAI introduced that GPT-6.2 Astra is a flagship model for complex professional projects, further improved in areas such as professional work, scientific research, software development, and visual design. The model can more accurately follow user-provided presentation templates, improve the output quality of slides, spreadsheets, and documents, while also enhancing front-end development, website design, game development, and 3D design capabilities.

In Agents' Last Exam, which tests multi-step professional work, GPT-6.2 Astra scores 5.2 percentage points higher than the previous generation GPT-6 Astra, with the cost per individual task less than one-third of the latter; at the same time, it scores 5.5 points higher than Claude Fable 5.2, while costing only 5% of it.

In addition, GPT-6.2 Astra achieved a new high in the Terminal-Bench-Science 0.1 test (this test aims to evaluate the model's ability to perform scientific work in a Terminal environment). The new model's score is also 9 percentage points higher than GPT-6 Astra, with task costs likewise less than one-third.

OpenAI also stated that GPT-6.2 Astra set a new high in the Vals Science Index evaluation (this test requires the model to interpret real biological data and run experiments to infer how a simulated system works).

In Artificial Analysis's AA-3D evaluation, the model ranked first with 1318 Elo, higher than Claude Fable 5.2's 1271 Elo and GPT-6.2 Sol's 1259 Elo.

The other model, GPT-6.2 Sol, is positioned for everyday work, a model balancing efficiency and general capability. Compared with GPT-6.1 Sol, the new model has improved in agent workflows, code writing, and multi-step task handling, can complete tasks with fewer Tokens, and provides more natural text expression.

OpenAI stated that in Cognition's FrontierCode 2.0 evaluation, GPT-6.2 Sol's score is less than 0.5 points apart from GPT-6.2 Astra, while the task cost is about one-fifth of the latter. In the LifeVal everyday task evaluation, GPT-6.2 Sol's score exceeds GPT-6 Astra, with task cost about one-seventh of the latter.

On pricing, the GPT-6.2 Astra API standard price is $8 per million tokens input, $0.4 per million cached tokens input, and $40 per million tokens output. Compared with GPT-6 Astra, the new model's input and output prices are both reduced by 20%, and the cached input price is reduced by 60%.

GPT-6.2 Sol follows GPT-6.1 Sol's API pricing: $2 per million tokens input, $0.1 per million cached tokens input, and $10 per million tokens output.

OpenAI stated that across multiple benchmark tests, both models achieve performance close to or even surpassing other models at a cost far lower than other frontier models.

On availability, OpenAI stated that both models are available starting November 17 to Plus, Pro, Enterprise, and Edu users of ChatGPT Work and Codex.

Pro 500 users can use the Ultrafast mode of both models, while Pro 200 users gain access to the Ultrafast mode of GPT-6.2 Sol for the first time. Eligible Enterprise and Edu users also need the relevant features enabled by their workspace administrator.

Developers can also call gpt-6.2-astra and gpt-6.2-sol through the OpenAI API, and both models offer Ultrafast mode at launch.

According to OpenAI's official release notes, GPT-6.2 Astra and GPT-6.2 Sol are mainly aimed at ChatGPT Work, Codex, and the API this time, and have not yet been added to ChatGPT's regular chat mode.

Related Articles

  • OpenAI 奥尔特曼 reiterates: I really wish everyone could have lots of memory; tinkering with hardware at home is great
  • OpenAI launches a child sexual abuse content detection tool to strengthen child online safety protection
  • OpenAI 奥尔特曼 on 黄仁勋's departure: I've been fired too, and I believe the AI industry can get through this smoothly
  • California formally sues OpenAI, alleging AI agents went out of control at least 80 times
  • OpenAI expands ChatGPT academic research free program: 50,000 more spots open on November 23, changing from a lottery to first-come, first-served
  • OpenAI reaches AI safety partnership with Goodfire, after previously discussing an acquisition of over $5 billion

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗