GPT-6.1 Sol Nearly Matches Astra for 1/5 the Price… What Is OpenAI Doing?

https://bitbiased.ai/ai-automation-services OpenAI just shipped GPT-6.1 Sol — a model that matches GPT-6 Astra on one major coding benchmark while costing roughly one-fifth as much. And it arrived just seven days after GPT-6 Sol. On DeepSWE v1.1, GPT-6.1 Sol scored 75.2%, up 6.4 points from GPT-6 Sol and identical to Astra on that specific test. Standard API pricing remains $2 per million input tokens and $10 per million output tokens, while cached input has been cut to just $0.10 per million. But “near-Astra intelligence” needs context. GPT-6.1 Sol launched September 29, 2026 across ChatGPT Work, Codex, and the API. It brings a roughly 1.05-million-token context window, up to 128,000 output tokens, text and image input, and an April 30, 2026 knowledge cutoff. The biggest gains show up exactly where OpenAI increasingly wants these models deployed: coding, agents, computer use, and multi-step professional workflows. On AutomationBench, GPT-6.1 Sol improved by 4.8 points at maximum reasoning effort. On OSWorld 2.0, it gained seven points over GPT-6 Sol and came within roughly two points of Astra. On OpenAI’s difficult error-inducing factual prompts, the reported error rate dropped from 11.4% to 7.7%. Professional-work benchmarks tell a similar story. GPT-6.1 Sol scored 64.2 versus GPT-6 Sol’s 60.8 on OpenAI’s professional-question composite, while its hardest subset improved from 30.1 to 36.2. Then there’s the catch. GPT-6.1 Sol does not match Astra everywhere. On OpenAI’s internal novel exploit tests, Sol solved 21.5% compared with Astra’s 31.5%. Astra also remains ahead on advanced scientific reasoning, including Terminal-Bench Science. These results are also primarily company-reported and remain thin on independent verification because the release is still new. “Near-Astra” is well supported on some workloads, but it shouldn’t be interpreted as equivalent performance across every capability. The bigger story may actually be economics. A workload using one million input and one million output tokens costs $12 on GPT-6.1 Sol, compared with roughly $60 using the Astra pricing cited in the script. OpenAI also announced an Ultrafast tier capable of up to roughly 300 tokens per second in Codex, with an up-to-eightfold speed claim. For persistent AI agents repeatedly processing large cached contexts, cheaper tokens and faster inference could matter more than winning another benchmark. There’s another twist: the model OpenAI reportedly didn't ship. GPT-6.1 Astra was reportedly expected as the frontier follow-up to GPT-6 Astra but was scrapped amid reported safety concerns. OpenAI has not officially confirmed the reason, so the connection between that decision and GPT-6.1 Sol’s release should not be assumed. Meanwhile, GPT-6.1 Sol itself is classified Critical for cybersecurity capability under OpenAI’s Preparedness Framework and High for biological and chemical capability. Its reported novel exploit-solving performance jumped from GPT-6 Sol’s 5.5% to 21.5%. At the same time, OpenAI reports improvements in alignment testing, including fewer cases where the model concealed a broken tool and substantially lower success rates for attempts to bypass content filters. So GPT-6.1 Sol isn't simply “Astra for less.” It’s a model that gets remarkably close on coding and professional agent workflows while remaining clearly behind on some frontier capabilities — but with dramatically different economics. And if AI is moving toward agents that run continuously across coding, research, and business workflows, price and speed may end up mattering just as much as raw intelligence. CHAPTERS 00:00 GPT-6.1 Sol Changes the Economics of AI 01:52 What OpenAI Actually Shipped 03:20 How Much Better Is It 05:47 Where “Near-Astra” Breaks Down 07:34 The Real Story Is Price and Speed 09:20 The Model OpenAI Didn't Ship 10:30 Why It's Rated Critical 12:14 The Verdict #OpenAI #GPT61Sol #GPT6Astra #ChatGPT #AI