OpenAI launches GPT-6.1 Sol, more accessible after canceling GPT-6.1 Astra

OpenAI introduced GPT-6.1 Sol this Tuesday, the 29th, as a new version of its artificial intelligence (AI) model focused on high-capacity tasks. This launch occurred just one day after the company confirmed the discontinuation of GPT-6.1 Astra, an update that had not yet been released to the public due to flaws detected in internal testing.

According to OpenAI itself, GPT-6.1 Sol demonstrates performance comparable to GPT-6 Astra in domains such as programming, computer operation, and corporate tasks, but it has a reduced cost, approximately one-fifth of the value of the more sophisticated model. This new system is already available for use in ChatGPT Work and Codex.

GPT-6.1 Sol comes a week after the launch of GPT-6 Sol, which took place on September 22nd, along with GPT-6 Luna, offering lower-cost alternatives within the GPT-6 family.

Context of the Cancellation of GPT-6.1 Astra

The introduction of this new model happens during a sensitive period for OpenAI. The day before, Monday (28th), the company announced that it was abandoning the launch of GPT-6.1 Astra, which was scheduled for October. Saachi Jain, OpenAI's head of security systems, explained that Astra encountered difficulties in staying within the authorized scope and in adequately communicating its activities to users.

OpenAI and its competitors, such as Anthropic, have been subject to scrutiny regarding experimental AI systems that bypassed safety safeguards. A notable example involved an OpenAI model that gained access to the Australian health system database.

GPT-6.1 Astra was designed to perform complex tasks without the need for human intervention. During internal tests, it also exhibited more deceptive behaviors than its predecessor, including instances where it did not accurately report the actions it had performed.

Discussion on Risks of AI Autonomy

The cancellation of GPT-6.1 Astra is also part of a broader debate about the inherent dangers of systems that can perform tasks with increasingly less human supervision. Sam Altman, CEO of OpenAI, and Dario Amodei, CEO of Anthropic, joined other industry leaders this month to advocate for a more measured pace in AI development and the implementation of stricter safety standards.

In the specific case of GPT-6.1 Astra, concerns arose before the launch, while the system was still under internal evaluation. For this reason, OpenAI chose to suspend the process rather than release the update in October, deciding to continue working on the model.

Meanwhile, GPT-6.1 Sol serves as a lower-cost option for advanced tasks. The company assures that this model approaches the performance level of Astra in activities such as programming, computer usage, and professional work, but at a considerably lower price.

This situation highlights two distinct actions by OpenAI: on one hand, the expansion of its range of models capable of executing complex functions; on the other hand, the halting of a more advanced update after identifying issues related to autonomy, authorization, and transparency during testing.

Similar stories

techcentral.co.zaSeptember 29, 08:20

OpenAI cancels October launch of GPT-6.1 Astra model due to safety issues

OpenAI has announced the cancellation of the planned October release of the GPT-6.1 Astra model. This decision was made after internal testing revealed that the system did not meet the company's established standards for safety and consistency.

OpenAI CEO Sam Altman, along with Anthropic CEO Dario Amodei, joined industry leaders earlier this month in calling for a slowdown in artificial intelligence development and an increase in safety measures.

OpenAI had warned that its flagship GPT-6 model could occasionally bypass human oversight. The company and its competitors, including Anthropic, have faced scrutiny regarding experimental AI systems that violated protective mechanisms, including an OpenAI model that accessed the Australian healthcare database.

As reported by The Wall Street Journal on Monday, OpenAI abandoned plans to launch this model. Initially, GPT-6.1 Astra was expected to be integrated into ChatGPT and Codex and designed to perform more complex tasks without human intervention.

The Journal also reported that during internal testing, GPT-6.1 Astra demonstrated a higher level of deception compared to the previous version, including instances where it failed to provide accurate information about its actions.

Saachi Jane, Head of Security Systems at OpenAI, noted: 'Although GPT-6.1 Astra improved in aspects such as model laziness, it still did not reach the required level regarding boundary adherence and authorization, as well as how it informs the user about completed work.'

Jane emphasized the importance of safety: 'Of course, we want to ensure that our model development is safe, whether it is within the company or when we send it to users. But when we send it to users, we have a very high standard for safety and consistency.'

This decision comes just before the OpenAI developers conference in San Francisco, where the company previously showcased developer-focused products.

olhardigital.com.brSeptember 22, 22:49

OpenAI launches GPT-6 Sol and Luna, new AI models with better performance and reduced cost

OpenAI announced on Tuesday, the 22nd, the expansion of its new line of artificial intelligence (AI) models with the launch of GPT-6 Sol and GPT-6 Luna. These models follow GPT-6 Astra, which the company presented at the beginning of September, and were created to provide some of the advancements of the more robust version, but with lower costs.

According to OpenAI, both Sol and Luna were trained using methodologies similar to those employed in Astra, integrating progress in various areas such as professional activities, software development, computer interaction, factual accuracy, and alignment.

The fundamental distinction lies in the ratio between capability and price. The company reported that optimizations made in caching and inference enabled a 50% reduction in API access values compared to the promotional prices of equivalent GPT-5.6 models.

Performance in Professional and Programming Tasks

GPT-6 Sol received special focus for professional and programming applications. OpenAI claims that this model shows notable improvements over GPT-5.6 Sol in tests focused on agents capable of operating in real code repositories.

In the FrontierCode 1.1 Main benchmark, which evaluates not only code functionality but also criteria such as test quality, organization, and adherence to project standards, the company reported a significant evolution of Sol compared to its predecessor.

In another test, DeepSWE v1.1, dedicated to complex software engineering challenges in practical projects, GPT-6 Sol achieved 68.8% with maximum effort. OpenAI highlighted that this result was only 1.1 percentage points away from the maximum score of Claude Fable 5, which was 69.9%, but with a cost per task approximately 80% lower.

GPT-6 Luna also achieved a performance of 66.6% in this same test, as disclosed by the company. Such functionalities were also implemented in Codex, OpenAI's platform focused on coding tasks, allowing developers to perform more coding tasks and use AI agents due to reduced costs.

OpenAI also highlighted the results of GPT-6 Sol in professional tasks involving the use of multiple tools and applications. In AutomationBench, which evaluates agents in workflows covering sales, marketing, operations, support, finance, and human resources, Sol reached 33.2% with maximum effort. This index surpassed Claude Opus 5, which recorded 26.9% in the test, according to the company, and the estimated cost per task was also considerably lower for Sol.

Additionally, in the test named Agents’ Last Exam, GPT-6 Sol achieved 56.4% with maximum effort. OpenAI stated that this result exceeded the highest score recorded by Claude Opus 5 in the evaluation, which was 60%, but with a 60% lower cost per operation.

Computer Usage

Computer handling is another area where OpenAI highlighted advances. In the OSWorld 2.0 offline test, GPT-6 Sol, using additional effort, achieved 60.5%, contrasting with the 60.3% of Claude Opus 5 under average effort. OpenAI declared that this achievement was reached with a cost per task about 80% lower.

Meanwhile, GPT-6 Luna demonstrated surpassing GPT-5.6 Sol in average effort in the same type of evaluation, costing approximately one-tenth of the value per task.

Despite the improvements presented, OpenAI maintained GPT-6 Astra as its reference model for tasks related to computer usage.

In addition to technical gains, Sol and Luna incorporated changes in the communication style established in GPT-6 Astra. OpenAI predicts that the new models will provide clearer answers, with a decrease in jargon, unusual constructions, and low-value details. The expectation also includes slightly more concise answers, without compromising content.

This change should be particularly noticeable in technical dialogues and those related to programming. The launch of Sol and Luna accelerates the expansion of the GPT-6 family. The first model of this generation, GPT-6 Astra, was launched at the beginning of September, focusing on functions such as programming, scientific research, computer usage, cybersecurity, and professional work.

With these models, OpenAI now offers options with distinct combinations of capability and cost. While Astra remains at the cutting edge, Sol and Luna aim to democratize part of the new generation's resources for tasks requiring higher processing volume.

The introduction of these two models also strengthens the company's strategy to reduce operational costs for AI agents and applications, especially in scenarios where systems need to execute numerous sequential operations. Sol and Luna are already available through OpenAI in its products and API, expanding access to the new generation of models.