GPT-6.1 Sol - OpenAI's new-generation flagship model

What is GPT-6.1 Sol

GPT-6.1 Sol is OpenAI's new-generation flagship model, serving asGPT-6 Sol's upgraded version, achieving intelligence close to the latter at one-fifth of the Astra flagship's token price. GPT-6.1 Sol focuses on three high-consumption scenarios: agenticprogramming, computer operation, and professional work. It matches Astra on real codebase engineering tasks, trails by only 2.1 percentage points on the OSWorld computer operation benchmark, and pushes cached input pricing down to just $0.1 per million tokens. The official release includes an Ultrafast tier with up to 8x speed, now available via API and ChatWork,Codexto paid users.

GPT-6.1 Sol's main features

• Agentic programming: Matches GPT-6 Astra on DeepSWE v1.1 real codebase engineering tasks, a 6.4 percentage point improvement over GPT-6 Sol.
• Computer operation: 7 percentage points higher than GPT-6 Sol on the OSWorld 2.0 computer use benchmark, only 2.1 percentage points behind Astra.
• Professional document processing: Surpasses Claude Opus 5.5 on the GDP.pdf benchmark, at less than half its per-task cost.
• Multi-step business workflows: On AutomationBench at medium reasoning effort, 2.2 percentage points higher than Opus 5.5 at about one-third its cost.
• Scientific workflows: On Terminal-Bench Science, per-task cost is $5.47, over 75% lower than Opus 5.5 and Astra.
• Factual accuracy: The factual error rate drops from 11.4% to 7.7% at low reasoning effort, and reaches 4.1% at high reasoning effort.
• Ultrafast high-speed tier: Token generation speed up to 8x that of the standard version, available via API and the $500/month Pro tier.
• Multiple reasoning effort levels: Supports adjustable reasoning effort from low to high, letting users trade off between cost and quality.

GPT-6.1 Sol's technical principles

• Model positioning that brings flagship capabilities downmarket: GPT-6.1 Sol is a mid-tier reasoning model in the GPT-6 series. OpenAI has "pushed down" GPT-6 Astra-level capabilities into a lower cost range through subsequent training optimization, and while maintaining the reasoning model architecture, it focuses on strengthening post-training for three scenarios: agentic programming, computer operation, and professional tasks.
• Adjustable reasoning effort mechanism: The model supports multiple reasoning effort settings: lower tiers respond quickly at low cost, suitable for simple tasks; the highest tier allocates more reasoning compute budget, and OpenAI reports scores at the highest tier on benchmarks such as DeepSWE, OSWorld, and Terminal-Bench.
• Cost optimization of context caching: Cached input pricing is set at $0.10 per million tokens, 95% lower than standard input and another 50% lower than GPT-6 Sol, which is the true core of this price reduction. When an agent reuses the same context across multiple requests, the server can reuse the cached KV/prefix and bill only for the new incremental input.
• Ultrafast inference acceleration: The Ultrafast tier provides token generation up to 8x standard speed, at the cost of a higher billing multiplier (about 6x the price in Codex).

How to use GPT-6.1 Sol

• Using the ChatGPT entry point: Log in to ChatGPT, and select the GPT-6.1 Sol model in ChatGPT Work or Codex (regular Chat does not currently offer it).
• API call: In your code, set the model ID to gpt-6.1-sol, and send requests using the standard Chat Completions interface.
• Set reasoning effort: Choose the reasoning effort level based on task difficulty. Use a lower level for low-difficulty tasks to save cost, and the highest level for complex tasks.
• Enable context caching: Reuse system prompts and task context across requests. Cached input is billed at $0.10 per million tokens.
• Enable Ultrafast when you need maximum speed: Enable up to 8x faster generation via the API or the $500/month Pro tier.

Core advantages of GPT-6.1 Sol

• Extreme cost-effectiveness: Delivers near-flagship intelligence at one-fifth the standard token price of GPT-6 Astra.
• Strong agentic coding: On the DeepSWE v1.1 real codebase evaluation, it matches Astra and improves on GPT-6 Sol by 6.4 percentage points.
• Leading computer use: On OSWorld 2.0 it trails Astra by only 2.1 percentage points, at about one-seventh the cost per task.
• Outperforms competitors on professional tasks: It surpasses Claude Opus 5.5 on both GDP.pdf and AutomationBench, at half to one-third the cost.
• Lower factual error rate: At low reasoning effort, the factual error rate drops from 11.4% to 7.7%, a reduction of about 32%.
• 8x Ultrafast: In the high-speed mode, it delivers up to 8x generation speed at a cost close to that of the previous Astra.

Comparison of GPT-6.1 Sol with similar competitors

Application scenarios of GPT-6.1 Sol

• Full app lifecycle maintenance: After a developer releases an app, a persistent Agent continuously tracks user feedback, automatically organizes high-frequency issues, fixes bugs, and submits complete PRs and change videos.
• Product development follow-up: During a new product launch, the Agent continuously learns the product positioning and team requirements, and after requirement changes automatically updates launch materials, adjusts documents, and flags parts that need manual confirmation.
• Scientific research data analysis: After new experimental data is entered into the system, automatically rerun the analysis pipeline, investigate outliers, update the paper's figures, and synchronously revise the related textual explanations.
• Enterprise finance automation: Monitor accounts receivable, and when missing invoices are detected, automatically prepare the invoice documents and send them after human confirmation—OpenAI has disclosed that an early test user's Dot actually completed this task.
• Multi-channel information monitoring: Regularly check designated Slack channels or emails, automatically summarize new findings into a shared ChatGPT Space page, and generate dynamic charts for the team to view in real time.