OpenAI's GPT-6.1 Sol Promises Cost Efficiency but Shows Significant Latency Tradeoff

OpenAI released GPT-6.1 Sol on September 29, 2026, as an upgraded alternative to GPT-6 Sol, claiming superior cost-per-task efficiency while maintaining near-parity with the flagship GPT-6 Astra model—but independent benchmarking reveals the model trades inference speed for improved reasoning accuracy. The new model achieved an Intelligence Index score of 52, up four points from GPT-6 Sol's 48, and improved performance across specialized benchmarks: 6.4 percentage points higher on DeepSWE software engineering, 7.0 points on OSWorld automation, and 18.0 points on ExploitBench. Despite these gains, GPT-6.1 Sol runs meaningfully slower than its predecessor, with output generation 13% slower at 64 tokens per second versus GPT-6 Sol's 74 tokens per second, and substantially higher time-to-first-token at maximum reasoning settings—on agentic tasks, the newer model averaged two to three times longer wall-clock times than GPT-6 Sol, reaching 75 minutes on Vibe Code Bench evaluations and 43 minutes on legal agent workflows.