If you blinked, you missed GPT-6 Sol. Seven days after it launched, GPT-6.1 Sol has already taken its seat — and according to Artificial Analysis, the new model sits just 1 point below the flagship GPT-6 Astra on the Intelligence Index while costing less than a quarter as much per task. If you pay for AI out of your own pocket, that one sentence is the whole story.
What happened
Artificial Analysis published the full breakdown on September 29, 2026. The short version: GPT-6.1 Sol replaces GPT-6 Sol after just 7 days, lands 1 point below GPT-6 Astra in the Intelligence Index, and does it for less than one quarter of the Cost per Task.
Two terms, in plain English. The Intelligence Index is a scoreboard — a fixed batch of hard tasks, boiled down to one number. Cost per Task is what you'd pay, on average, to get one of those tasks finished. GPT-6.1 Sol moved up the scoreboard and the bill went down at the same time.
The score jumps
Against GPT-6 Sol, the new model gains 4 points on the Intelligence Index. Against GPT-5.6 Sol, it's 5. The detailed wins:
• Agentic knowledge work: +4 points on AA-Briefcase v1.1 and +5 on GDPval-AA v2.1
• Terminal-Bench 4.0: +12 points
• Humanity's Last Exam: +5 points
• GDP.pdf: +6 points
• AA-Omniscience Accuracy: +8 points, with the hallucination rate dropping from 60% to 54%
"Agentic" is the industry's word for an AI that carries out a whole job rather than answering one question — it opens files, runs the steps, checks its own work. Think of the difference between asking a friend for directions and hiring a courier to actually deliver the parcel.
The pricing story
The list price hasn't budged: $2 per million input tokens and $10 per million output tokens, same as GPT-6 Sol. Tokens are the chunks of text the model reads and writes; you're billed by the chunk, like paying for groceries by weight.
One thing did improve. The cache read discount climbs from 90% to 95%. "Cache" is text the model has already seen — a contract you keep pasting into the same chat, for example — so re-reading it now costs half of what it used to.
That nudges GPT-6.1 Sol's blended price for agentic workloads slightly below GPT-6 Sol's. And it stacks on top of an earlier cut: GPT-6 Sol had already knocked 50% off GPT-5.6 Sol.
At max effort, a single task costs $0.72, versus $3.26 for GPT-6 Astra. That's 31% less than GPT-6 Sol ($1.05) and 64% less than GPT-5.6 Sol ($1.99).
What it means for you
Benchmarks are abstract. Here's what they look like in an ordinary week.
At home
Long documents are the quiet win. GDP.pdf jumped 6 points, and that's the benchmark for reading and reasoning over PDFs. That means the pile of lease agreements, insurance policies and appliance manuals you've been avoiding — you can ask "what's the cancellation clause?" and actually get a straight answer.
At work
Agentic knowledge work is where most office jobs live: pulling figures out of one tool, writing them up, sanity-checking the result before it goes anywhere. The +4 on AA-Briefcase v1.1 and +5 on GDPval-AA v2.1 are exactly that kind of task. If you already hand a weekly report to an AI, this is the release that makes it less likely to hand back nonsense.
Running a business
This is where the 95% cache discount earns its keep. If your assistant works from the same product catalogue, refund policy or style guide over and over, those repeated reads now cost half of what they cost on GPT-6 Sol — and a tenth of the original rate. For a two-person team, that's the difference between reusing your own documents and retyping the context every single time. If you'd rather test prompts before you commit to an API bill, the free tools at MyKreaTool are an easy place to start without a credit card.
Studying
Humanity's Last Exam gained 5 points. It's a set of brutally hard expert-level questions — the kind a professor would have to pause over. A 5-point move there means better step-by-step explanations of genuinely difficult material, which is exactly what you want when you're stuck on a problem set at 11pm.
Creative work
AA-Briefcase improved by roughly 80 Elo, driven by a better rubric score and Analytical Quality Elo — while Presentation Elo slipped slightly. Translation: sharper thinking, marginally less polish. Use it for structure, research and honest critique, then do the final formatting pass yourself.
Earning money with it
Freelancers and small agencies bill for revisions. When a task costs $0.72 instead of $3.26, you can afford to run the same job several times, keep the best result and still come out ahead. At this point the margin is the product.
How to try it right now
You don't need to be an engineer for any of this.
1. Start free — read the scoreboard. Open the GPT-6.1 Sol model page on Artificial Analysis. It costs nothing and needs no account. You'll see the Intelligence Index, the Cost per Task and the effort settings sitting side by side with GPT-6 Astra.
2. Find the model you already use on the same chart. If yours is behind on both score and price, that's your answer.
3. Pick your effort level deliberately. Low and medium effort are the most token-efficient settings here. And in the Coding Agent Index, xhigh actually beat max by 3 points — so "crank it to maximum" isn't automatically the smart call.
4. Do the price math if you're paying. $2 per million input tokens, $10 per million output tokens, 95% off cached reads. Multiply that by your real monthly volume before you migrate anything.
5. Run one actual task. Not a demo — your report, your PDF, your spreadsheet. One real job tells you more than any leaderboard.
Upsides and what changes
Cheaper per task at every setting
"Pareto frontier" sounds technical. It just means: no better deal exists anywhere on that line. All effort levels of GPT-6.1 Sol push that frontier outward, so for a given level of intelligence, nothing is cheaper. At max effort you're paying $0.72 a task where Astra costs $3.26.
Coding gets noticeably sharper
At max effort, the Coding Agent Index gains 3 points over GPT-6 Sol and sits 2 points below GPT-6 Astra. The interesting wrinkle: xhigh scores 1 point above GPT-6 Astra for under 15% of the Cost per Task — a 6-point gain over GPT-6 Sol at max.
The boring win: fewer invented facts
Accuracy up 8 points, hallucination rate down 6. That's the improvement you feel after a month of use, not the one that shows up in a launch post.
Limitations
Be honest about the fine print. GPT-6.1 Sol uses roughly 10–30% more output tokens than GPT-6 Sol at the same effort levels, so a cheaper rate per token doesn't automatically mean a cheaper bill — watch volume, not just prices. It's still 1 point behind GPT-6 Astra, and the hallucination rate on AA-Omniscience only came down to 54%, which means it still gets things wrong there more often than not. Presentation Elo slipped slightly, so anything client-facing deserves a human read before it ships. And the context matters most of all: a model replaced its predecessor in 7 days. Build your workflow so that swapping the model underneath takes five minutes, not a rebuild.
Conclusion
The best value in frontier AI isn't the flagship anymore — it's the model one rung down that costs a quarter as much and trails by a single point. GPT-6.1 Sol is that model today, and today is all anyone can promise in this market.
Your one action: open the free comparison page at Artificial Analysis, find whatever AI you currently pay for, and see where it lands next to GPT-6.1 Sol. If it's behind on both score and price, run a single real task through a Sol-tier model this week. One task, one line on your own invoice — that's the only benchmark that actually pays you.
Comments 0