CVCOVANewsComputing, Opportunities, Value & Analysis
Artificial Intelligence4 min read

OpenAI cuts GPT-6.2 Astra API pricing by 20%: what changes for those paying in dollars

OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol. The flagship model's usage-based pricing drops, but the cheaper model's stays the same, and the fine print will determine how much users in Latin America actually save.

By COVA News newsroom
Turned-on MacBook Pro displaying programming code

Photo: Arnold Francisca on Unsplash

For a startup in Bogotá, Lima or Mexico City that pays in dollars for every query its product sends to OpenAI's artificial intelligence, a price cut changes the math. That is what the November 17, 2026 announcement (U.S. time) brought: OpenAI introduced GPT-6.2 Astra, its most powerful model, with a 20% lower rate on its API, the connection developers use to integrate AI into their own applications. The discount, however, does not apply to everyone.

The company published the launch on its developer forum and on X, and Spanish-language media picked it up on November 18, which is why both dates are circulating. Two models arrived. Astra is, according to OpenAI, a "frontier" model, meaning it belongs to the most advanced tier and is designed for the hardest projects. Sol is the efficient version for everyday work.

Who pays less and who pays the same

These models charge by tokens, the chunks of text into which both what is sent (input) and what the model replies (output) are divided. According to the announcement, Astra goes from $10 to $8 per million input tokens and from $50 to $40 per million output tokens. Sol stays at $2 and $10, the same as its predecessor, GPT-6.1 Sol. One outlet described the cut as a general discount on tokens, but OpenAI and several specialized outlets agree that it applies only to Astra.

An example helps illustrate the change. A company that processes 100 million input tokens and 20 million output tokens with Astra in a month would go from paying $2,000 to $1,600, before taxes. According to TestingCatalog, which cites OpenAI's documentation, Astra's cached input (repeated text that the system reuses without processing it again) also drops, from $1 to $0.40 per million tokens.

The fine print matters. A single outlet reports that when a request exceeds 272,000 input tokens, the input rate doubles and the output rate is multiplied by 1.5, which would bring Astra to $16 and $60. That same outlet says the Ultrafast mode, which delivers faster responses, costs three times the normal rate and would rise to six times on March 1, 2027. OpenAI has not confirmed these details in the material reviewed.

OpenAI also claims that Astra completes each task at less than a third of the cost of GPT-6 Astra, and that on the FrontierCode 2.0 programming benchmark, Sol nearly matches Astra at about a fifth of the cost. These are the company's own figures, without independent verification.

What external measurements say

Independent evaluators are more cautious. According to the firm Artificial Analysis, cited by the outlet Casrai, Astra scores 62 points on its intelligence index, behind Anthropic's Claude Fable 5.2 (65), while Sol scores 57. The Decoder reports that Sol matches GPT-6 Astra at about 15% of the cost per task, and that Astra leads a model comparison ranking that the outlet does not identify.

The same outlet cites hallucination rates (made-up answers presented as true) of 59% for Astra and 55% for Sol at maximum reasoning effort, though it does not specify on which test. Another Artificial Analysis summary indicates that Astra does not improve in coding and that Sol drops 3 points compared with its previous version. In response to the initial criticism, OpenAI's Thibault Sottiaux said on X that there were significant changes in the three or four days before launch that some evaluators did not get to see, as reported by Android Authority.

There is also a security development. OpenAI presents Astra as its most capable model in cybersecurity, scoring 22 points more than GPT-6 Astra on the ExploitBench Internal Port test, and both models ship with stricter restrictions on that kind of assistance. The Decoder and SC World report that Astra is the first released OpenAI model to reach the "High" level of self-improvement, an internal scale the company uses to measure a model's ability to contribute to improving AI itself, and that a monitoring system called Control Room is being deployed as a result.

As for availability, the models are listed in the API as gpt-6.2-astra and gpt-6.2-sol, and are coming to ChatGPT Work and Codex for the Plus, Pro, Enterprise and Edu plans. Ultrafast debuts on the Pro 200 plan, starting with Sol, while Pro 500 gets it with both models.

For Latin America, questions remain open. OpenAI has not specified in what it has published whether payment methods, billing or data storage in the region will change, and the final cost will depend on the exchange rate and each country's taxes. Those currently working with Sol will not see direct savings. Those using Astra will, as long as they do not rely on very long requests or on Ultrafast. Before migrating, it is advisable to wait for official confirmation of those rates and for more independent testing.

Related news

Artificial Intelligence

Mathematician says he has proved the Riemann hypothesis with help from Claude; it still needs to be verified

Levent Alpöge has published a purported proof of one of the most famous problems in mathematics, produced together with Anthropic's AI. Early reviews are encouraging, but the community has not yet validated it.

Artificial Intelligence

Nvidia reports earnings without Jensen Huang: what's at stake for AI chips

Five days after its founder's resignation, the company that dominates artificial intelligence chips faces Wall Street. For Latin America, open questions outweigh measurable effects for now.

Artificial Intelligence

Alibaba launches Qwen4-Max: a giant, cheaper model, still without independent testing to back it up

The Chinese company unveiled a frontier artificial intelligence model with lower prices and compatibility with OpenAI and Anthropic tools. For developers in Latin America it is another option, though questions remain about real-world performance, privacy and access from the region.