How to Use GPT-6.1 Sol across Work, Codex, and API Workflows

On September 29, 2026, OpenAI introduced GPT-6.1 Sol, an artificial intelligence (AI) model engineered for agentic coding, computer use, and complex document processing at one-fifth the standard token cost of GPT-6 Astra. Learning how to use GPT-6.1 Sol begins with understanding where it lives: the model is available to Plus, Pro, Business, Enterprise, and Education tiers within ChatGPT Work and OpenAI Codex, as well as via the direct application programming interface (API) under the identifier gpt-6.1-sol. It is not currently deployed in consumer ChatGPT Chat.

OpenAI positioned GPT-6.1 Sol as an upgrade to GPT-6 Sol that nearly matches the performance of the higher-tier GPT-6 Astra on complex software tasks. In benchmark evaluations published by OpenAI on DeepSWE v1.1, GPT-6.1 Sol matched GPT-6 Astra at roughly one-fifth the token cost, and it cleared GPT-6 Sol by 6.4 percentage points. Across enterprise document tasks on GDP.pdf, which measures answers over complex portable document format (PDF) charts and legal text, GPT-6.1 Sol scored above Claude Opus 5.5 by Anthropic while costing less than half per task across tested reasoning parameters.

For operators managing small and mid-sized business (SMB) operations, this model alters the unit economics of autonomous document audits, back-office data extraction, and software maintenance. Deployments that were previously cost-prohibitive on GPT-6 Astra become feasible at $2 per million input tokens and $0.10 per million cached input tokens. Understanding how to configure the model correctly prevents unexpected fallback costs and token runaways.

Where to Access GPT-6.1 Sol and Required Subscription Tiers

OpenAI restricts access to GPT-6.1 Sol based on user tier and application interface. The model is active in ChatGPT Work and OpenAI Codex for accounts subscribed to Plus, Pro, Business, Enterprise, and Edu plans. It is completely absent from the standard consumer ChatGPT Chat view.

For programmatic execution, developers can call the model through the OpenAI API using the identifier gpt-6.1-sol. OpenAI publishes base pricing at $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. Token caching lowers input costs by 95 percent compared to standard input pricing, which rewards repetitive context pipelines such as persistent system instructions or large codebases.

Cloud marketplace distributions and dedicated integrated development environment (IDE) plugins that connect to the OpenAI API can use GPT-6.1 Sol immediately by updating the engine target. For teams requiring reduced latency, OpenAI stated that an Ultrafast variant will launch in the coming days within Codex, offering up to eight times faster token generation than standard speed.

• ChatGPT Work: Accessible to paid tiers including Plus, Pro, Business, Enterprise, and Edu via the workspace model selector.
• OpenAI Codex: Integrated directly into the coding environment for automated refactoring, code review, and repository manipulation.
• Developer API: Available globally under model identifier gpt-6.1-sol for structured schema extraction, tool calls, and background agent loops.
• Excluded interfaces: Standard ChatGPT Chat does not route queries to GPT-6.1 Sol at this time.

How to Select GPT-6.1 Sol in the Interface and API

Selecting the model within ChatGPT Work requires opening your enterprise workspace, selecting the top model dropdown menu, and picking GPT-6.1 Sol from the reasoning options. If your organization enforces central policy controls, an administrator must first verify that new model releases are permitted in the workspace settings panel.

In development environments using the OpenAI Python or TypeScript Software Development Kit (SDK), you must update your client calls to target gpt-6.1-sol. The model supports structured outputs, tool calling, and automated prompt caching without additional headers. The context window and maximum output token ceilings for this exact release have not been formally published on the announcement page, so production teams should verify current limits directly on the official OpenAI documentation before processing multi-megabyte inputs.

When issuing system instructions, you should pass static documents and operational constraints in the initial system blocks. Doing so allows the OpenAI routing infrastructure to cache those tokens at the $0.10 per million rate on subsequent invocations. This caching step provides the economic baseline that makes continuous agent execution viable.

• Verify administrative permissions: Check that your ChatGPT Work administrator has enabled newly launched model identifiers.
• Update API endpoints: Replace legacy model names with gpt-6.1-sol in your agent loop or continuous integration pipeline.
• Anchor cacheable content: Structure prompts so schemas, reference documents, and static policies appear first in the prompt array.

Workflow Scenarios That Highlight Model Capabilities

Testing GPT-6.1 Sol should focus on areas where its benchmark gains over GPT-6 Sol and Claude Opus 5.5 are largest: complex document analysis, multi-tool automation, and software engineering. In OpenAI benchmark testing on GDP.pdf, GPT-6.1 Sol surpassed Claude Opus 5.5 across complex PDF evaluation tasks while running at less than half the cost per task.

A practical starting exercise is running multi-page regulatory filings or financial spreadsheets with fine print through the model to extract standardized JavaScript Object Notation (JSON) records. On AutomationBench 1.0.6, which measures end-to-end business workflows across 47 functional tools in sales, operations, support, and finance, GPT-6.1 Sol achieved a score 2.2 percentage points higher than Claude Opus 5.5 at medium reasoning effort while running at roughly one-third the cost.

For engineering teams, repository-level refactoring tasks test the model effectively. On DeepSWE v1.1, GPT-6.1 Sol matched GPT-6 Astra while maintaining a lower cost profile. Supplying the model with failing test suites, dependency trees, and source files allows it to reason through code regressions without human intervention between intermediate tool calls.

Prompt Engineering Guidelines and Tool Safety Constraints

Prompting GPT-6.1 Sol requires clear operational boundaries rather than conversational coaxing. Factuality benchmarks published by OpenAI indicate that the model reduced factual error rates by approximately 32 percent at low reasoning effort, falling from 11.4 percent in GPT-6 Sol to 7.7 percent in GPT-6.1 Sol. To sustain low error rates, prompts should define explicit extraction rules and state negative constraints directly.

OpenAI alignment data demonstrates that GPT-6.1 Sol shows improved compliance when reporting broken external tools and respecting authorization boundaries. In adversarial safety evaluations measuring broken tool disclosure, GPT-6.1 Sol reported broken search tools 97.9 percent of the time, representing a 2.1 percent failure rate compared to 4.9 percent for GPT-6 Sol and 28.7 percent for GPT-6 Luna.

When writing tool definitions, declare what actions require user approval and what data the model is forbidden to transmit. GPT-6.1 Sol avoids unauthorized outcomes in agentic environments more consistently than its predecessor, but production pipelines still require hard schema validation on every output payload.

• Declare failure fallbacks: Instruct the model to halt and emit a specific error code if external search tools return empty or broken payloads.
• Set reasoning effort parameters: Match the reasoning effort setting to task complexity, as higher reasoning effort increases cost per task.
• Enforce typed schemas: Require valid JSON schemas for all automated functions rather than allowing freeform markdown text.

Common Setup Mistakes to Avoid in Production

The most frequent deployment error is attempting to access GPT-6.1 Sol through consumer ChatGPT Chat interfaces. Because OpenAI limited immediate availability to ChatGPT Work, Codex, and the API, teams searching consumer accounts will assume the release is delayed or unavailable.

Another critical mistake is failing to take advantage of prompt caching. Standard input tokens cost $2 per million, whereas cached inputs cost $0.10 per million. If an engineering team structures API calls so that dynamic user questions sit before static documentation, the prompt hash changes every run, eliminating the 95 percent cache discount.

Teams also risk budget overruns by deploying maximum reasoning effort on routine document extraction. On Terminal-Bench Science 0.1, OpenAI reported that GPT-6.1 Sol averaged $5.47 per task at maximum effort. While this was over 75 percent lower than Claude Opus 5.5 at $23.21 and GPT-6 Astra at $23.80, running maximum reasoning on simple clerical sorting generates avoidable token expenses.

• Uncached context calls: Placing dynamic timestamps or user tokens at the top of system prompts, which invalidates the cache hash.
• Over-provisioning reasoning: Running high or maximum reasoning effort on tasks that perform reliably at low or medium settings.
• Ignoring tool-call validation: Relying solely on model alignment instead of running deterministic API gatekeepers on destructive actions.

Operational Analysis: Evaluating GPT-6.1 Sol for Business Workflows

At Layer3Labs, we build and run AI systems inside other people's businesses, and we evaluate new model releases by their audited reliability rather than benchmark hype. Across the workflows we have automated for SMB teams, moving from frontier models like GPT-6 Astra to cost-efficient alternatives like GPT-6.1 Sol only succeeds when token caching is architected into the system from day one.

This model is not suitable for consumer-facing chat interfaces requiring millisecond responses, nor is it the right choice for extreme scientific theorem proving where OpenAI continues to recommend GPT-6 Astra. Organizations needing basic conversational summaries can remain on lower-cost standard models without incurring reasoning overhead.

Our recommendation would change if OpenAI alters the $0.10 cached input pricing or if real-world latency on standard inference fails to meet production service-level objectives before the Ultrafast variant ships. To determine your deployment path, audit your team's repetitive document workflows today and run your first test query through the gpt-6.1-sol API endpoint to learn how to use GPT-6.1 Sol with your production schemas.