Claire Donovan
A task that cost $3.26 with GPT-6 Astra now costs $0.72 with GPT-6.1 Sol. The intelligence gap: one point. That is the whole story.
OpenAI just reset the cost of near-frontier AI. According to Artificial Analysis (2026), GPT-6.1 Sol sits one point below GPT-6 Astra on the Intelligence Index while costing less than a quarter per task. That gap is small enough to change how you build.
What Is GPT-6.1 Sol?
OpenAI unveiled GPT-6.1 Sol on September 29, 2026 at its DevDay conference in San Francisco (TechCrunch, 2026). It is a lower-cost model positioned just below GPT-6 Astra, the company's flagship, and it replaced GPT-6 Sol only seven days after that model launched.
The model is available now in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu tiers, but not in the regular ChatGPT chat interface. It supports a 1,050,000-token context window, a 922,000-token maximum input, and a 128,000-token maximum output (OpenAI, 2026).
How Much Does GPT-6.1 Sol Cost?
The GPT-6.1 Sol API pricing is $2 per million input tokens and $10 per million output tokens. GPT-6 Astra runs $10 and $50. That is exactly one-fifth the price across both token types (TechCrunch, 2026).
Cached input costs $0.10 per million tokens. OpenAI says that is 95 percent cheaper than standard input and half the cached-input price of GPT-6 Sol. Cache writes are $2.50 per million tokens (Windows Report, 2026). For agent workloads that read the same context repeatedly, this number matters more than the headline input price.
Why Is GPT-6.1 Sol So Cheap?
OpenAI has not explained the exact cost engineering, but the pricing strategy is clear. This is the frontier AI price war reaching the API. A model that nearly matches the flagship is priced as a utility, not a premium product.
The cached input rate is the clearest signal. At $0.10 per million tokens, developers can keep long system prompts, codebases, and session history in context without multiplying their bill. For a given level of intelligence, Artificial Analysis found no cheaper model (Artificial Analysis, 2026). This is a cheap GPT-6.1 alternative to running Astra for every task.
When Did GPT-6.1 Sol Come Out?
The GPT-6.1 release date was September 29, 2026, at OpenAI DevDay in San Francisco (TechCrunch, 2026). It replaced GPT-6 Sol seven days after that model launched. If you were evaluating GPT-6 Sol, the upgrade cycle just lapped you.
The model is currently available in ChatGPT Work and Codex for paid tiers, but not in the general ChatGPT chat interface. For more on why it is not in ChatGPT, see GPT-6.1 Sol Matched GPT-6 Astra on a Key Benchmark at One-Fifth the Cost - So Why Isn't It in ChatGPT?
How Does GPT-6.1 Sol Compare to GPT-5?
OpenAI has not released a direct GPT-5 benchmark table, so any GPT-6.1 vs GPT-5 answer is more about context and error rate than a single score. The 1,050,000-token context window is far larger than GPT-5-class models, and the low-reasoning factual error rate on difficult prompts dropped to 7.7 percent for GPT-6.1 Sol, compared with 11.4 percent for GPT-6 Sol (TechCrunch, 2026).
The model has a 922,000-token maximum input, a 128,000-token maximum output, and five API reasoning effort levels from low through max (OpenAI, 2026). That is built for agents and long documents, not quick chat replies.
How Good Are GPT-6.1 Sol Benchmarks?
OpenAI's numbers show the gap between cheap and flagship is not a cliff. At low reasoning effort, the share of responses containing a factual error on difficult prompts fell from 11.4 percent for GPT-6 Sol to 7.7 percent for GPT-6.1 Sol. Across all reasoning settings, OpenAI says its error rate is within 1.9 percentage points of GPT-6 Astra (TechCrunch, 2026).
On the DeepSWE v1.1 software-engineering benchmark, OpenAI says GPT-6.1 Sol matches GPT-6 Astra at roughly one-fifth the cost. On OSWorld 2.0 computer-use tasks, it lands within 2.1 points of Astra at about one-seventh the cost per task (WION, 2026).
Independent testing from Artificial Analysis puts GPT-6.1 Sol 2 points below Astra on the coding agent index, but the cost per task is so much lower that the tradeoff is hard to ignore (Artificial Analysis, 2026).
What This Means for Your Build
Near-frontier AI now costs $0.72 per task instead of $3.26 (Artificial Analysis, 2026). For a 22-year-old indie developer, an overnight batch of coding agent tasks that once cost a dinner out now costs less than a coffee.
Cached context at $0.10 per million tokens makes long-running agents affordable. A coding tutor, job-application assistant, or research agent can reuse a long system prompt and project context all semester without burning your budget.
This is the same shift described in GPT 6.1 Sol Cut Near-Frontier AI Costs by 80% and Changed Who Gets to Build. The cost floor dropped, and that changes what you can afford to try.
The Other Side of Cheap GPT-6.1 Sol
Cheap near-frontier capability is available to everyone. Your side project can move faster, but so can every competitor with the same API key. Entry-level coding and knowledge-work tasks face faster automation because the cost floor just dropped.
The model is also a reminder to verify. OpenAI replaced GPT-6 Sol within a week, and GPT-6.1 Sol is not in the regular ChatGPT interface. Treat autonomous agents as tools to direct and check, not as finished judgment.
AI agents are already common in companies, but 79% of Companies Are Using AI Agents Right Now - But Only a Fraction Actually Have Them Working. Cheap models like GPT-6.1 Sol will push that number higher, and the gap between adopting agents and actually making them work will matter even more.
Verdict: GPT-6.1 Sol Removes the Cost Barrier
Should you care? Yes, if you build or experiment. GPT-6.1 Sol removes the cost barrier to near-frontier AI. The real skill now is directing agents, verifying output, and applying judgment, because the raw intelligence is cheap enough to treat as a utility.
Start with a small agent task, measure the output, and check the bill. The price floor moved, and that changes what you can afford to try.
