GPT-6 Sol Benchmarks: Published Scores and Business Realities

In September 2026, OpenAI introduced GPT-6 Sol as part of an expansion to its GPT-6 model tier alongside GPT-6 Luna. The release provides a mid-tier commercial model positioned directly beneath OpenAI's flagship GPT-6 Astra, aimed at balancing high reasoning capabilities with lower application programming interface (API) operational costs.

Unlike the earlier GPT-5.6 Sol or standard chat endpoints, GPT-6 Sol cuts token pricing by 50 percent while delivering higher accuracy on long-horizon professional evaluations. On AutomationBench, OpenAI reports that GPT-6 Sol at extra-high effort scores 33.2 percent, topping rival Claude Opus 5 at max effort while consuming roughly 9 percent of the cost per completed task.

For operators running software automations in legal intake, financial documentation, or customer operations, this release signals a fundamental shift in running costs for reasoning agents. Understanding what these benchmark numbers measure, where they fail to predict real workflow reliability, and how API fees scale is necessary before moving production pipelines onto new infrastructure.

AutomationBench and Professional Workflow Performance

OpenAI reports that GPT-6 Sol scored 33.2 percent on AutomationBench version 1.0.6 at extra-high effort. This test measures agent performance across 47 individual business tools spanning sales, marketing, operations, support, finance, and human resources (HR). By comparison, Claude Opus 5 at max effort reached 26.9 percent, while Claude Fable 5.1 with Opus 5 fallback scored 31.4 percent. OpenAI also noted that low-effort GPT-6 Astra scored 30.3 percent on the same test.

Cost per completed task on AutomationBench was reported at $0.27 for GPT-6 Sol at extra-high effort. OpenAI stated that Claude Opus 5 cost 11.1 times more per task, while low-effort GPT-6 Astra cost 3.9 times more. Claude Fable 5.1 was reported as costing over 8.9 times more, though OpenAI noted that this comparison omitted Opus 5 fallback costs that occurred during roughly 40 percent of those tasks.

These figures show that smaller, high-effort inference runs can outperform previous-generation larger models on synthetic computer-use routines. However, operators should confirm current benchmarks directly on official vendor test repositories, as evaluation harnesses receive frequent updates.

Agents' Last Exam and Real-World Factuality Evals

On Agents' Last Exam version 1, GPT-6 Sol scored 56.4 percent at maximum reasoning effort. This benchmark covers long-horizon tasks across 55 professional sub-industries to evaluate economic value on knowledge-work problems. OpenAI reported that this score exceeded Claude Opus 5's top mark in the evaluation, while cutting task cost by 60 percent.

Factuality evaluations conducted by OpenAI showed that GPT-6 Sol produced roughly half as many errors as its predecessor, GPT-5.6 Sol. The test dataset utilized de-identified conversations from ChatGPT where users had flagged factual mistakes in prior models. OpenAI reported that answer length sweeps showed minimal correlation between verbosity and error rates on these queries.

OpenAI did not publish raw point scores for its internal factuality evaluation, reporting instead a relative 50 percent error drop compared to GPT-5.6 Sol. Teams building systems requiring strict factual verification must run their own domain checks rather than relying on vendor averages.

API Pricing Reductions and the Cost-Intelligence Curve

OpenAI cut standard API pricing for GPT-6 Sol to $2 per million input tokens and $10 per million output tokens. This represents a 50 percent price reduction compared to GPT-5.6 Sol, which was priced at $4 per million input tokens and $20 per million output tokens during its promotional period. The lightweight GPT-6 Luna was reduced to $0.10 input and $0.50 output per million tokens.

Lower inference costs allow engineering teams to increase reasoning effort parameters without breaking unit economics. For example, running an agent at extra-high effort requires more intermediate thinking tokens, which previously made multi-step tasks cost-prohibitive on flagship models.

In our client engagements across document automation and administrative workflows, the base token fee is rarely the only operational cost. Schema formatting errors, repeated API tool retries, and context window drift frequently double actual invoice totals if guardrails are not strictly maintained.

Why Public Benchmarks Overstate Practical Business Performance

Public evaluations like AutomationBench measure models inside clean, static software environments with clear API specifications. Real corporate tools, such as customized customer relationship management (CRM) software or enterprise resource planning (ERP) platforms, routinely return ambiguous error codes, unexpected rate limits, and partial database payloads. An agent scoring 33 percent on an isolated benchmark often struggles when handling messy data inputs across three non-standardized systems.

Benchmark tasks are also evaluated on single-shot success criteria without long-term accountability. If a model misinterprets a field during an intake flow, the evaluation marks it as a simple fail, whereas an operational pipeline creates downstream accounting or legal reconciliation work.

Teams should look at published benchmarks as an index of raw reasoning capacity rather than an operational readiness guarantee. True verification requires running identical workflow scenarios with prompt edge cases, bad inputs, and intermittent service disconnects.

Audience Fit and Implementation Realities

GPT-6 Sol fits mid-sized organizations building agentic workflows that require deeper reasoning than small models provide, without the budget burden of GPT-6 Astra. Teams handling structured document verification, lead enrichment across external databases, and internal ticket triage gain the most from its pricing structure.

This model is not suitable for organizations operating under strict local-residency mandates that prohibit external cloud API inference. Teams facing zero-tolerance compliance requirements for data export should instead deploy self-hosted open-weight architectures inside private clouds.

Our assessment would shift if OpenAI removes tiered reasoning effort controls or if enterprise data privacy addendums require higher subscription tiers. Monitor OpenAI's official developer terms and API status notices regularly to confirm current SLA parameters.