From Startup Fortune · · 1m 10s
GPT-6 Astra Beats Claude Sonnet on Science Tasks
GPT-6 Astra uses fewer tokens to outperform Claude Sonnet 5.5 on science benchmarks.
In brief
GPT-6 Astra uses fewer tokens to outperform Claude Sonnet 5.5 on science benchmarks. GPT-6 Astra scores 63% on Terminal-Bench-Science vs Claude Sonnet 5.5's 53% using 27,000 tokens versus 193,000 tokens. Originally reported by Startup Fortune.
GPT-6 Astra scores higher on science tasks
GPT-6 Astra scores 63% against Claude Sonnet 5.5's 53% on Terminal-Bench-Science.
“GPT-6 Astra scores 63% against Claude Sonnet 5.5's 53% on Terminal-Bench-Science”
Astra uses fewer tokens per task
Astra needs 27,000 tokens versus Sonnet 5.5's 193,000 tokens for a typical task.
“Astra needs about 27,000. Sonnet 5.5, run at its highest reasoning setting, generates roughly 193,000 output tokens”
Cost per task differs significantly
Astra costs $3.26 per task versus Sonnet's $7.60, despite lower token prices.
“Astra ends up costing $3.26 per task against $7.60 for Sonnet”
Token efficiency flips economics
Astra's 7x token gap makes it cheaper despite higher per-token pricing.
“That's close to a 7x gap, not the 22x some chatter around the models has claimed, but still enough to flip the economics entirely”
OpenAI's pricing strategy
OpenAI charges $10 per million input tokens and $50 per million output tokens for Astra.
“OpenAI charges $10 per million input tokens and $50 per million output tokens for it”
GPT-6.1 Sol pricing advantage
GPT-6.1 Sol costs $5.47 per task versus Astra's $23.80.
“Sol completes a task for about $5.47, compared with $23.80 for Astra and $23.21 for Claude Opus 5.5”
Astra classified as critical for security
Astra is the first model OpenAI classified 'Critical' for cybersecurity under its Preparedness Framework.
“Astra is the first model OpenAI has classified 'Critical' for cybersecurity under its own Preparedness Framework”
Cost per task matters more than leaderboard rank
For R&D workloads, token efficiency outweighs benchmark scores in cost calculations.
“A lab that needs thousands of research-style completions a day will feel a 7x token gap in its cloud bill long before it feels a 10-point difference on an abstract benchmark”
The essence
GPT-6 Astra scores 63% on Terminal-Bench-Science vs Claude Sonnet 5.5's 53% using 27,000 tokens versus 193,000 tokens.
