Skip to content
Theme
OnDevice Tools

AIAI Models & Research

OpenAI launches GPT-6.2 Astra and Sol with lower API prices

OpenAI has introduced GPT-6.2 Astra and GPT-6.2 Sol, two models aimed at professional and everyday work. The Decoder reports lower API costs for Astra and mixed benchmark results, with gains in science and design but flat coding performance.

Illustration of OpenAI GPT-6.2 Astra and Sol models with lower API prices and benchmark charts

Key takeaways

  • OpenAI released GPT-6.2 Astra for professional work and GPT-6.2 Sol for everyday tasks.
  • Astra's API cost is 20 percent below the previous model, while Sol keeps the GPT-6.1 Sol rate card.
  • Artificial Analysis ranks Astra second behind Claude Fable 5.2, but reports flat coding performance and high hallucination rates on hard tests.
  • OpenAI says both models reach the Critical cybersecurity level and is adding stricter safeguards.
In this article
  1. Two new OpenAI models
  2. API prices and access
  3. Benchmark gains
  4. Independent rankings and coding
  5. Science, design and security

Two new OpenAI models

Illustration: AI, OpenAI logo, code, Claude logo and the labels Artificial Analysis, OpenAI, AI Models & Research

According to The Decoder, in a report published on 17 November 2026, OpenAI has introduced GPT-6.2 Astra and GPT-6.2 Sol. Astra is aimed at demanding professional work in science, law, finance and accounting. Sol is intended for quick, everyday tasks. The Decoder describes both systems as competitive with Anthropic's leading models while costing less.

The models are open to Plus, Pro, Enterprise and Edu customers. Developers can reach both systems via the OpenAI API. OpenAI says Astra is designed to follow professional conventions more reliably, including in presentations. It says slide layouts, colours and typography can follow templates, with exact font sizes. OpenAI promotes Astra as a leading model for scientists.

The Decoder also notes that Astra is the first released model to hit the High level for AI self-improvement on OpenAI's scale. OpenAI treats that as a risk.

API prices and access

OpenAI says the price for Astra is 20 percent below its predecessor. Astra costs $8 for each million input tokens, $0.40 for cached input tokens, and $40 for each million output tokens.

Sol keeps the earlier GPT-6.1 Sol rate card. It charges $2 for each million input tokens, $0.10 for cached input tokens, and $10 for each million output tokens. This keeps Sol at a lower price point for routine tasks, while Astra remains priced for heavier professional workloads.

Benchmark gains

OpenAI says Astra improves on GPT-6 Astra in several areas. On Agents' Last Exam, which measures complex professional tasks, OpenAI says Astra scores 5.2 points above GPT-6 Astra while costing under one third as much. The company also says Astra outscores Claude Fable 5.2 on that test by 5.5 points while spending only 5 percent as much. Google's Gemini 4 Carbon is the only rival OpenAI says stays close on that measure.

On Harvey LAB-AA, a legal test that includes contract clause review, OpenAI says Astra's pass rate is more than twice Claude Fable 5.2's, at roughly one sixth of the expense. GDPval-v2, an internal OpenAI test covering consulting, finance and design, shows a 6.2 percentage point gain over the earlier model, with 22 percent fewer output tokens.

Independent rankings and coding

Artificial Analysis ranks Astra second in its Intelligence Index, behind Claude Fable 5.2. The index combines nine evaluations spanning knowledge work, coding, science and general knowledge. Artificial Analysis says Astra rises five points above GPT-6 Astra. It passes Gemini 4 Carbon, which scores 60, and Claude Opus 5.5, which scores 59.

Sol gains two points to reach 57. That matches GPT-6 Astra's score while costing about 15 percent as much per task. Artificial Analysis also says OpenAI models dominate the cost-performance frontier. This is the set of models where none is both cheaper and better. Anthropic and Meta also have models on that frontier. OpenAI also holds most leading places for efficiency, using output tokens per task as the measure. Claude Fable 5.2 is the only non-OpenAI model listed there.

Coding results are less impressive. Artificial Analysis says coding progress has stalled. In the Coding Agent Index, using the Codex harness at maximum compute, Astra remains at GPT-6 Astra's level. Sol falls three points below GPT-6.1 Sol. The drop is linked to a 14-point fall on SWE-Atlas-QnA-AA. Sol leaves out key facts more often than the earlier model.

Artificial Analysis also reports high hallucination rates. At maximum effort, Astra hallucinates in 59 percent of cases, while Sol does so in 55 percent. The measure is AA-Omniscience Hallucination Rate. It tracks how often a model guesses instead of admitting uncertainty on factual questions. The questions are unusually difficult, so the figures may not represent everyday hallucination rates.

Science, design and security

OpenAI says Astra is stronger in science and computer use. It reports gains on Terminal-Bench-Science, which checks scientific work in a terminal environment. Astra also reaches a record result on the Vals Science Index, a measure of biological data interpretation and experiment handling.

In demonstrations, Astra created a three-dimensional car model, produced a Python-based physics solver, and animated the output in Blender. It also built a remote-controlled car component by component and made an animated assembly guide. On AA-3D, a three-dimensional design test where people compare two anonymous interactive experiences, Astra ranks first with 1,318 Elo points. Claude Fable 5.2 follows at 1,271, and Sol at 1,259. Artificial Analysis also places Astra first in AA-3D and AA-Robotics v0.1. Its robotics score is 79, compared with 78 for Sol and 75 for Claude Fable 5.2.

OpenAI says both models are stronger in cybersecurity. They can locate and exploit security weaknesses more effectively, produce safer code, and provide improved bug fixes. Several tests also show better token efficiency. Both reach the Critical cybersecurity level in OpenAI's Preparedness Framework, matching GPT-6 Astra.

OpenAI is tightening safeguards around potentially dangerous cybersecurity requests. The company says this should reduce misuse, but it may also restrict lawful vulnerability studies and security testing. The cybersecurity benchmark figures OpenAI published come from internal tests where safeguards and classifiers were disabled. OpenAI describes its method in an accompanying safety report, which provides the relevant details. OpenAI also plans to make the models available through OpenAI Daybreak. That programme already provides vetted security researchers with access to models that have fewer restrictions.

Sources

  1. OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gainsThe Decoder

Tags

Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.