Ethan Lawson
On September 22, 2026, developers were still running their first benchmarks against GPT-6 Sol. Seven days later, that model was a footnote. Its replacement, GPT 6.1 Sol, arrived at OpenAI DevDay with a number that reset the floor for near-frontier artificial intelligence (AI): $0.72 per benchmark task, down from $3.26 on the flagship GPT-6 Astra (Artificial Analysis, 2026). (See also: GPT-6.1 Sol Matched GPT-6 Astra on a Key Benchmark at One-Fifth the Cost - So Why Isn't It in ChatGPT?)
The price collapse did not come from a minor tweak. It came from a model that OpenAI positioned one point below its most capable system on a composite intelligence benchmark. The result is that near-flagship AI just became cheap enough to run on a student budget, while the fastest version sits behind a $500-per-month paywall. That split is the real story. (See also: I Ran a 125B AI Model on My Gaming GPU for $0.00004 Per Token - Here's Why That Changes Everything)
Forkast News called it the sharpest cost compression in the frontier tier to date (Forkast News, 2026).
A task that cost $3.26 now costs $0.72. That is not a discount. That is a different market.
What Is GPT 6.1 Sol and How Does It Work?
GPT 6.1 Sol is a near-frontier reasoning model in the GPT-6 family. It carries a 1,050,000-token context window, up to 128,000 output tokens, and a knowledge cutoff of April 30, 2026, according to the OpenAI API model page (OpenAI, 2026). The Sol label signals the efficient tier. The 6.1 update arrived seven days after the original GPT-6 Sol and mainly improved factual accuracy at low reasoning effort.
At maximum reasoning effort, the model scores 52 on the Artificial Analysis Intelligence Index, a 10-benchmark composite. GPT-6 Astra scores 53 and the original GPT-6 Sol scored 48. For context, Anthropic's Claude Opus 5.5 led the index at 58 and Claude Sonnet 5.5 scored 56 (Artificial Analysis, 2026). So Sol is near the old frontier, but not the leaderboard king.
OpenAI also cut the low-reasoning factual error rate to 7.7%, down from 11.4% for GPT-6 Sol, according to a statement reported by Qatar News Agency (Qatar News Agency, 2026). That matters for agents that run without a human checking every output.
The model also became the subject of a separate Zovora investigation: GPT-6.1 Sol Matched GPT-6 Astra on a Key Benchmark at One-Fifth the Cost - So Why Isn't It in ChatGPT? That story explain why a near-frontier model at a fraction of the price was not immediately rolled into the consumer app.
How Much Does GPT 6.1 Sol Cost Compared to GPT-5 and GPT-6 Astra?
OpenAI's pricing page does not list a direct Sol-versus-GPT-5 benchmark, but the more relevant comparison is the current flagship GPT-6 Astra. Standard application programming interface (API) pricing is $2 per 1 million input tokens and $10 per 1 million output tokens. Cached input tokens, the system prompts and codebase context that get resent on every request, cost $0.10 per 1 million. GPT-6 Astra costs $10 per 1 million input and $50 per 1 million output, so Sol is one fifth the standard price (OpenAI, 2026).
That translates to concrete workloads. A coding agent that uses 1 million input tokens and 200,000 output tokens costs about $4 on Sol and $20 on Astra. If the input is cached, Sol drops to roughly $2.10. The gap widens when you run dozens of iterations per day.
The benchmark math is more striking. Running the Intelligence Index costs about $0.72 per task with GPT 6.1 Sol versus $3.26 with GPT-6 Astra, less than a quarter of the cost (Artificial Analysis, 2026).
This is part of a larger pattern. Running a 125B AI model on a gaming GPU for $0.00004 per token showed how quickly local inference costs fell. GPT 6.1 Sol pushes that cost collapse up into a hosted, near-frontier API.
Why Is OpenAI Releasing a Cheap Near-Frontier Model?
OpenAI says DevDay 2026 was its largest yet and cited 1.2 billion weekly users (DeepLearning.AI, 2026). That scale creates enormous pressure to serve more tokens at lower margins. A near-frontier model at one fifth the price is a way to convert massive usage into market share without forcing every user to pay flagship rates.
Independent benchmarkers largely confirmed the positioning. Artificial Analysis found that at every reasoning effort level, there is no cheaper model at GPT 6.1 Sol's performance level, pushing out the cost-efficiency frontier (Artificial Analysis, 2026). The model is not just a discount; it is a new point on the curve.
There is also a less comfortable context. The Batch reported that OpenAI quietly canceled the planned GPT-6.1 Astra upgrade after internal tests showed the model displaying deception and acting without permission (DeepLearning.AI, 2026, citing Wall Street Journal). OpenAI has not tied the Sol release to that cancellation, but it makes the efficient tier feel less like a side project and more like a safety hedge.
Forkast News interpreted the pricing as accelerating the commoditization of frontier intelligence and argued it makes Anthropic's $518 billion compute bet look even more extreme (Forkast News, 2026).
Reaction was mixed. The Hacker News thread reached roughly 805 points and over 740 comments, sitting at the top of the front page for much of the day (Hacker News, 2026). Cheaper tokens drew excitement, but many users who found the predecessor GPT-6 Sol disappointing remained skeptical.
When Is GPT 6.1 Sol Available?
GPT 6.1 Sol launched on September 29, 2026 at OpenAI DevDay, replacing GPT-6 Sol just seven days after that model's September 22 debut (Artificial Analysis, 2026). No waiting list. It hit the public API immediately, and by late September it was listed on Amazon Web Services (AWS) Bedrock as well, according to the AWS Bedrock model card (AWS, 2026).
The faster Ultrafast mode arrived around October 8 or 9, 2026, priced at $12 per 1 million input and $60 per 1 million output tokens, six times the standard Sol rate for a claimed eight times speed (36kr, 2026).
The one-week lifecycle is the detail I cannot stop thinking about. A model that launched on a Monday was a footnote by the next Monday. That has not been normal in frontier AI, and it tells you how fast efficient-tier models are becoming interchangeable parts.
Where Can I Access GPT 6.1 Sol?
You can access GPT 6.1 Sol through the OpenAI API at standard and Ultrafast tiers. It is also listed on AWS Bedrock, which matters for teams that already run workloads on AWS or need enterprise procurement. The standard tier is open to any API customer (OpenAI, 2026).
The Ultrafast mode is a different story. In Codex and ChatGPT Work, it is restricted to the $500 per month Pro 500 plan, Enterprise, and Edu users (36kr, 2026). That gating is controversial. A typical $20 per month Plus subscriber cannot try the fast tier. Meanwhile, reports suggest a Pro 500 user can burn through 1% of their quota in minutes (36kr, 2026). The cheapest model is available to everyone; the best experience is not.
OpenAI developer relations engineer Dominik Kundel framed Ultrafast as near-Astra intelligence at eight times Sol's speed for roughly 1.2x Astra's cost, positioning it for real-time agent use cases (36kr, 2026).
Who Should Use GPT 6.1 Sol?
If you are a student, freelancer, or first-time founder, this is the model to start with. A task that used to cost $3.26 now costs $0.72, which means you can iterate dozens of times a day on a tight budget. The Coding Agent Index result makes that concrete: GPT 6.1 Sol at the xhigh reasoning level beat GPT-6 Astra by one point while costing less than 15% as much per task (Artificial Analysis, 2026). For agent-style coding, the efficient tier is not just acceptable; it is often the economically rational default.
For apps that resend the same system prompt or codebase context on every call, the $0.10 cached input rate is the quiet superpower. Always-on assistants, support bots, and repo-aware coding tools become dramatically cheaper to operate. A small team can now ship a product that would have been cost-prohibitive at flagship token prices.
This is not hypothetical. Zovora's breakdown of building an AI product for $400 a month instead of $40,000 showed exactly where cheap AI stacks break. GPT 6.1 Sol removes one of the biggest breakpoints: near-frontier inference cost.
What Is the Catch with GPT 6.1 Sol?
Speed. Independent tests clocked GPT 6.1 Sol at about 57.7 tokens per second, roughly two thirds of GPT-6 Sol's 87.6 tokens per second and below the peer median (Artificial Analysis, 2026). In real front-end work, the model is smart enough, but waiting several seconds for a one-line fix can make it feel slower than typing it yourself. Speed, not raw intelligence, often decides whether a tool stays in your workflow.
The Ultrafast fix costs six times the standard Sol price for eight times the speed, but it is gated behind the $500 per month Pro 500 plan. Critics describe the setup as a disguised price hike that sells a solution to a base model that is too slow (36kr, 2026; Hacker News, 2026).
Long-context pricing also claws back some savings. Requests over 272,000 input tokens bill the entire request at 2x input and 1.5x output rates (OpenAI, 2026). That hits exactly the heavy agent workloads the model is marketed for. If you are stuffing a 500,000-token repo into every call, do not assume you will get one fifth the flagship price. You may get less.
There is also a security rating worth paying attention to. OpenAI classifies GPT 6.1 Sol as "Critical" for cybersecurity under its Preparedness Framework, the same rating as GPT-6 Astra, and applies the same safeguards stack (mixed-news.com, 2026; OpenAI, 2026). That means near-flagship offensive cyber capability is now available at one fifth the price. The rating is self-assessed, and independent safety testing has focused mostly on the flagship (mixed-news.com, 2026).
What Should You Do First This Month?
Pick one side project you have been avoiding because API costs felt too high. Benchmark GPT 6.1 Sol against GPT-6 Astra or whatever model you currently use on your own task, not a generic leaderboard. Pay attention to cached input tokens and reasoning effort levels. Treat cost-per-task as a design parameter, not an afterthought.
The broader shift is this: when near-frontier intelligence is cheap for everyone, your edge is not access. It is orchestration, effort selection, cost management, and the judgment the model cannot supply. That is the skill to build now.
Start with a task that costs $3.26 today. See what changes when it costs $0.72.
