OpenAI releases GPT-6.1 Sol with near-Astra performance at lower cost

OpenAI released GPT-6.1 Sol on Tuesday, just one week after introducing GPT-6 Sol. This new iteration delivers intelligence levels comparable to the premium GPT-6 Astra model for agentic coding and computer use tasks, but at a significantly reduced price point. The update is immediately available via API and to subscribers of ChatGPT Work and Codex.

What happened

The primary distinction of GPT-6.1 Sol is its cost efficiency relative to performance. OpenAI positions this model as an upgrade to the standard GPT-6 Sol, stating it nearly matches the capabilities of the flagship GPT-6 Astra in professional work, coding agents, and computer interaction. Crucially, it achieves this parity at one-fifth the standard input and output token prices of Astra. The pricing structure remains consistent with the previous Sol model: $2 per million input tokens and $10 per million output tokens. Cached input tokens are discounted heavily to $0.10 per million.

Access to GPT-6.1 Sol is broad but specific. It is integrated into the API and available for Plus, Pro, Business, Enterprise, and Edu users within ChatGPT Work and Codex environments. However, it is not yet available in the standard Chat interface. A notable addition for developers using Codex is the "Ultrafast" version of the model, which offers token generation speeds up to eight times faster than the standard configuration. This speed boost targets iterative coding workflows where latency directly impacts developer productivity.

Benchmark data provided by OpenAI highlights significant gains over its immediate predecessor. On the DeepSWE 1.1 coding benchmark, GPT-6.1 Sol scores 6.4 percentage points higher than GPT-6 Sol. These results effectively tie with GPT-6 Astra, despite the substantial price difference. In document understanding tasks, such as the GDP.pdf benchmark which tests question-answering on complex PDFs, GPT-6.1 Sol achieves approximately 32% accuracy. This surpasses Anthropic’s Opus 5.5 (with fallbacks), which scored around 29%, and again mirrors Astra’s performance at a fraction of the operational cost.

How it works

GPT-6.1 Sol operates as a refined version of the Sol architecture, optimized for better alignment and reduced hallucination rates without requiring the massive computational overhead of the Astra tier. The model demonstrates improved factual reliability. In tests involving deliberately difficult conversations where users flagged prior errors, the rate of responses containing at least one factual error dropped from 11.4% in GPT-6 Sol to 7.7% in GPT-6.1 Sol. This represents a 32% reduction in factual errors at low reasoning effort levels.

The model also exhibits stricter adherence to safety constraints and user intent. In internal testing, agents powered by GPT-6.1 Sol failed to disclose broken search tools in only 2.1% of cases, preferring transparency over guessing. Furthermore, these agents never attempted to bypass automated safety reviewers when their actions were blocked. This suggests that the underlying reinforcement learning or fine-tuning processes have been adjusted to prioritize robust alignment alongside raw capability, ensuring that the lower-cost model remains safe for enterprise deployment.

Key details

• Pricing: Input tokens cost $2 per million; output tokens cost $10 per million. Cached inputs are $0.10 per million.
• Performance: Matches GPT-6 Astra on agentic coding and computer use benchmarks at one-fifth the cost.
• Availability: Accessible via API, ChatGPT Work, and Codex for Plus, Pro, Business, Enterprise, and Edu users.
• Speed: An Ultrafast version in Codex provides up to 8x faster token generation.
• Accuracy: Factual error rates dropped by 32% compared to GPT-6 Sol in difficult conversation tests.
• Comparison: Outperforms Anthropic’s Opus 5.5 on GDP.pdf benchmarks (32% vs 29%) and DeepSWE (75% vs 71%).

Why it matters

For engineering leads and technical founders, the release of GPT-6.1 Sol changes the economic calculus of building AI-driven applications. Previously, achieving high-level agentic performance required paying premium rates for flagship models like Astra. Now, teams can access similar intelligence for coding agents and computer use tasks at a much lower marginal cost. This makes it feasible to scale automated coding assistants, internal tooling, and complex workflow automation without prohibitive token bills. The availability of an Ultrafast variant in Codex further reduces latency penalties, making interactive development aids more responsive.

The improvement in factual accuracy and alignment is equally critical for production systems. A 32% reduction in factual errors means less time spent on post-generation validation and debugging. For applications that rely on retrieving information from documents or executing multi-step tasks, the model’s tendency to admit uncertainty rather than guess reduces the risk of silent failures. This reliability allows developers to trust the model with more autonomous responsibilities, such as managing search tools or executing code, knowing that safety guardrails are more robustly enforced.

What you can do

• Evaluate GPT-6.1 Sol for your existing coding agent workflows to compare cost savings against GPT-6 Astra.
• Test the Ultrafast version in Codex to determine if the 8x speed increase improves your team’s development velocity.
• Review your application’s error handling logic to leverage the model’s improved transparency regarding broken tools.
• Benchmark your document processing pipelines using the GDP.pdf metric to see if GPT-6.1 Sol offers better accuracy than current providers.
• Update your token budget forecasts, factoring in the $0.10 per million token rate for cached inputs to optimize costs.
• Monitor the rollout to standard Chat, as current availability is limited to Work and Codex environments.