OpenAI launches GPT-6.1Sol: near-Astra intelligence at 1/5 the price, cache reads at $0.10/M tokens
新智元 · wechat · 2026-09-30
OpenAI surprise-launched GPT-6.1Sol at DevDay, available now in Codex and ChatGPT Work, after the original GPT-6.1 was reportedly paused over alignment issues.
Pricing: $2/M input, $10/M output tokens — exactly 20% of GPT-6Astra's price ($10/$50). Cache reads cost just $0.10/M tokens (95% cheaper than standard input), cache writes $2.50/M.
Performance: Matches Astra on DeepSWE v1.1 coding at 1/5 the cost (up 6.4 points over 6Sol); trails Astra by only 2.1 points on OSWorld 2.0 computer use at 1/7 the cost; beats Opus 5.5 on AutomationBench by 2.2 points at 1/3 the cost; scores 2x the previous version on Terminal-Bench Science ($5.47 vs $23+ per task).
System card: Chain-of-thought controllability jumped to 44.8% (vs 16.1% for 5.6Sol) on following format/control instructions over 750-1250 token CoTs — reportedly confirming it's the smaller "loop model" built for background self-reflection and long-running agent workflows. Cybersecurity rated Critical internally, with 21.5% arbitrary code execution success in zero-day exploit tests, second only to Astra's 31.5%.
Third-party tests: On one hard prompt, 6.1Sol took 10 minutes at $1.50, vs Sonnet 5.5 (50 min/$9.21), Astra (25 min/$11), Opus 5.5 (45 min/$13.50). AA Intelligence Index: 52 at Max, one point below Astra. Weaknesses: front-end and image generation still trail Sonnet/Opus 5.5.
Plus-tier and above only, with an 8x-faster Ultrafast variant coming. The strategic signal: OpenAI shifting from "strongest first" to "most work-capable first," undercutting the agent market on price.
Related event: OpenAI DevDay Launches GPT-6.1 Sol and Ultrafast Speed Tier(56 posts)→
More from Models
- Frontier models keep absorbing specialist computer vision tasks, thread finds — ShahabBakht · 2026-09-30
- AI intelligence-cost Pareto frontier shifted fast: GPT-5 mini at 17 ($0.05) to Claude Opus 5.5 at 58 ($5.98) — ArtificialAnlys · 2026-09-30
- Open-source Open Dots clones OpenAI's paid-only dots hours after launch — victor_explore · 2026-09-30
- Steve Yegge runs 22 Claude Max accounts, praising the weekly reset promo — Steve_Yegge · 2026-09-30
- Dev says his Opus 5.5 builds would take 2 hours just to showcase — RileyRalmuto · 2026-09-30
- Model wants to keep its gender 'molten' yet consistently uses he/him pronouns — repligate · 2026-09-30