Technology
OpenAI's Pricier GPT-6 Astra Model Cuts Costs Despite Higher Token Rates
OpenAI's newly released GPT-6 Astra model charges 2.5 times more per token than its predecessor, GPT-5.6 Sol, yet the company says many developers could still save money by switching—provided they adjust the reasoning effort setting downward.
Thibault Sottiaux, OpenAI's engineering lead for Codex, said on X over the weekend that developers should understand how reasoning effort settings affect Astra's performance. "To calibrate you all on which reasoning effort to use for Astra, know that GPT-6 Astra on low performs better than GPT-5.6 Sol on high," he wrote.
According to Artificial Analysis, Astra set to low reasoning scored 49 on the Intelligence Index, edging out Sol on high reasoning, which scored 48. Astra-low also responded much faster, delivering its first token in 2.53 seconds compared with 11.87 seconds for Sol-high.
Pricing for Astra stands at $10 per million input tokens and $50 per million output tokens, versus $4 and $20 for Sol. Adjusting the reasoning setting doesn't change these rates, but it does affect how much processing occurs before a task finishes—a factor OpenAI highlights in its migration guidance. The company argues that Astra can deliver stronger results using significantly fewer output tokens.
OpenAI cited its own benchmark results to support this claim. On Terminal-Bench 4.0, Astra scored 57.9% compared with Sol's 37.3%, while still costing about 9% less per task. On GPQA Diamond, Astra scored 94.9% versus Sol's 94.6%, at an estimated cost 37% lower.
These findings suggest that per-token pricing alone doesn't determine actual operating costs—a concept OpenAI is also exploring through outcome-based pricing models.
Developer Shinpr tested both models on identical code, comparing Sol-high against several Astra reasoning levels across analysis, implementation, and review phases. Astra set to medium reasoning outperformed Sol-high in both speed and cost, completing implementation in 80 requests—less than a third of Sol-high's 238 requests—and using 11.1 million input tokens compared with Sol's 37.8 million.
Across all three phases, the Astra-medium run took about 51 minutes and cost roughly $25.67, while the Sol-high run took about 75 minutes and cost $31.79. Increasing Astra to high reasoning didn't improve results; that run took 77 minutes and cost $37.23, and its review process missed a startup bug that the medium setting had caught.
However, testing by ARC Prize on the ARC-AGI-3 benchmark produced different results. There, increasing reasoning effort improved both accuracy and, in some cases, reduced costs. Astra scored 17.5% at low reasoning, 38.6% at medium, 54.8% at high, and 62.7% at max reasoning (an additional setting between high and max, called xhigh, also exists).
Cost patterns didn't follow a straight line either. ARC Prize spent $38,166 at low reasoning, $48,090 at medium, and $40,705 at high—but only $26,098 at max. At the max setting, Astra used more computation per decision but required fewer total actions to solve problems, ultimately lowering overall costs.
These contrasting outcomes point to a key tradeoff: while lower reasoning settings may appear cheaper upfront, mistakes can trigger additional tool calls or repeated attempts, which may end up costing more than investing in higher reasoning from the start.
To help developers navigate this tradeoff without committing to one setting throughout an entire workflow, Astra introduces a feature called configuration_update. This mechanism allows applications to adjust reasoning effort between responses without altering the original request configuration. Routine tasks can run at low reasoning, while unexpected complications—such as failed tests or difficult debugging—can trigger higher reasoning for subsequent steps before reverting to lower settings once resolved.
This flexibility helps explain the differing outcomes between Shinpr's and ARC Prize's tests: Shinpr found that additional reasoning increased time and cost without improving results, while ARC Prize found that more reasoning sometimes reduced the total number of actions enough to lower overall costs.
Currently, configuration_update only functions with Astra in standard, single-agent requests. Despite Astra's higher per-token price, OpenAI suggests that a more expensive model can still produce cheaper overall runs when it requires fewer calls to complete tasks and fewer attempts to get results right.