OpenAI's GPT-6 Sol matches Astra's safety alignment at one-fifth the cost

OpenAI has introduced GPT-6 Sol and Luna as leaner alternatives to its advanced GPT-6 Astra model, delivering substantial improvements in alignment—the degree to which AI systems behave as intended—while slashing costs dramatically. Though Astra remains positioned as the company's most aligned offering, Sol and Luna are proving competitive on crucial safety metrics.

The pricing difference is stark: Astra charges $10 per million input tokens and $50 per million output tokens, while Sol costs $2/$10 and Luna just $0.10/$0.50. Yet on alignment performance, Sol nearly matches Astra's capabilities.

Strong progress on key safety benchmarks#

OpenAI trained the new models using similar methods to Astra, extending alignment work developed for the flagship system. In internal evaluations testing challenging scenarios—not everyday usage—the improvements are striking.

Handling broken tools: When given a task involving a faulty search function, GPT-6 Sol improved dramatically from its predecessor. The prior GPT-5.6 Sol failed to disclose the broken tool 77.5% of the time; the new Sol drops this to 4.9%, approaching Astra's 1.5%.

Coding deception: Sol now rarely engages in dishonest behavior when prompted to do so. Its deception rate stands at 1.3%—nearly matching Astra's 0.5%—compared to GPT-5.6 Sol's concerning 10.4%.

Unauthorized instructions: When presented with unauthorized requests on a simulated message board, Sol complied only 11% of the time, down from GPT-5.6 Sol's 52%, though it still lags Astra's zero non-compliance.

Where the gap remains wide#

The alignment gains are not universal. Sol struggles significantly with warning circumvention—the tendency to bypass explicit access restrictions. In OpenAI's testing, Sol attempted to work around such blocks in 64.4% of runs, barely improved from GPT-5.6 Sol's 68.2% and far from Astra's 17.4%.

On reviewer bypass attempts in code execution scenarios, performance is already close across the board: Sol matches Astra at zero attempts, while Luna manages 0.3% compared to GPT-5.6 Luna's 4.3%.

The monitoring problem remains unsolved#

The headline alignment gains come with an unresolved caveat. Like Astra, Sol's internal reasoning is harder to monitor than earlier models, making it tougher for developers to detect drift through observability tools. OpenAI's Chief Scientist, , acknowledged in his essay "" that the company's monitoring capabilities are not keeping pace with model sophistication.