You finally picked an OpenAI model last week. Bad news: it’s already old. GPT-6.1 Sol landed on September 29 and replaced GPT-6 Sol after just seven days, at the same price with better scores. Here’s what actually changed and which model your money should go to.
What happened in those 7 days
OpenAI shipped GPT-6 Sol on September 22. It was the flagship for complex coding and agentic work, priced 50% below the older GPT-5.6 Sol. Then, one week later, the company quietly posted an addendum introducing GPT-6.1 Sol, a new model in the same family that scores higher across the board.
The official docs now point anyone opening the GPT-6 Sol page toward GPT-6.1 Sol as “the newer Sol model.” That’s about as clear a “move along” as a documentation page can give.
Seven days is an absurdly short model lifecycle. For comparison, most OpenAI flagship generations used to last six months or more. This is the new normal: incremental point releases arriving fast enough that your “default model” decision expires monthly.
But before the eye-roll, look at what a week bought.
The numbers that matter
Independent benchmark lab Artificial Analysis ran GPT-6.1 Sol through its Intelligence Index suite the day it shipped. The gains are real:
- +4 Intelligence Index points vs GPT-6 Sol, landing just 1 point below GPT-6 Astra, OpenAI’s most expensive model
- +12 points on Terminal-Bench 4.0 (agentic coding in terminals)
- +5 on Humanity’s Last Exam, +6 on GDP.pdf
- Hallucination rate down from 60% to 54% on their omniscience test
- +3 points on the Coding Agent Index, 2 points below Astra
That last gap deserves a highlight. For months, Astra has been the “if you know, you know” model for serious agent work. GPT-6.1 Sol closes most of that gap at a quarter of the operating cost.
What it costs, in tokens and in real tasks
The sticker price didn’t move: $2 per million input tokens, $10 per million output. Two things under the hood did.
First, the cache discount grew. Reading cached input now costs 95% off, which is $0.10 per million tokens instead of the old 90% discount. If you run agents that re-send long system prompts and reference docs all day, that’s an instant cut on your biggest recurring cost.
Second, and this is the part the press release won’t emphasize: GPT-6.1 Sol writes more. Artificial Analysis measured roughly 10-30% more output tokens than GPT-6 Sol across effort levels. Output tokens are the expensive ones. So your raw bill can creep up even while the per-task numbers improve.
How do we square that? Cost per task, the metric that actually matters for agent work. Here’s the full picture:
| Model | Cost per Intelligence Index task | Intelligence vs Astra |
|---|---|---|
| GPT-5.6 Sol | $1.99 | 6 points below |
| GPT-6 Sol | $1.05 | 5 points below |
| GPT-6.1 Sol | $0.72 | 1 point below |
| GPT-6 Astra | $3.26 | baseline |
Read that middle row again. GPT-6.1 Sol delivers near-Astra intelligence for less than a quarter of Astra’s cost per task. Artificial Analysis puts it plainly: at every effort level, there is no cheaper model for a given level of intelligence.
Which OpenAI model should you actually use
Here’s my take after staring at these numbers.
Most people should switch to GPT-6.1 Sol. It’s the default choice for coding, agent workflows, document work, and anything where you’re paying by the token. Same price as the model it replaced, better results, cheaper per task. There’s no real trade-off unless verbosity genuinely bothers you.
Keep Astra for the small set of problems where maximum capability earns its price: gnarly multi-hour agent runs, high-stakes analysis, whatever your “break glass” tasks are. A 4x cost premium needs a reason.
One technical note if you write API calls yourself: GPT-6.1 Sol doesn’t support the “none” reasoning effort setting. Effort starts at low and goes up to max. Chat Completions works without tool calling, and the Responses API handles everything else, including computer use, web search, file search, and MCP.
What about ChatGPT users, not developers?
Maybe you don’t touch the API at all. You just open ChatGPT and use whatever shows up. Fair enough, but the switch still matters to you.
OpenAI listed GPT-6.1 Sol as a product release on September 29, the same day as its Dots agent launch. Model rollouts inside ChatGPT happen progressively: paid tiers typically see a new flagship first, free users get it later, and everyone keeps whatever the model selector shows until the rotation completes.
Practical advice for non-developers: don’t chase model names in the picker. Watch how your actual tasks come out. If your weekly newsletter draft, your resume rewrite, or your code debugging session suddenly gets sharper with no prompt changes, that’s the new model reaching your account. Our Opus 5.5 vs GPT-6 Sol comparison showed exactly how much real-world output differs between generations, sometimes in ways benchmarks miss.
Also worth knowing: GPT-6.1 Sol’s knowledge cutoff is April 30, 2026. Anything after that date needs the browsing tool. Older models won’t know about it, which is why asking about recent events sometimes feels like the AI “forgot” things it never knew.
Switching without breaking anything
If you call the API directly, change your model ID from gpt-6-sol to gpt-6.1-sol and run your normal test prompts. Both IDs still work today, but the old page already redirects you to the new model, so treat the swap as inevitable.
Running agents through a router or an automation platform? Your platform will likely flip the default for you at some point. Don’t wait for that. Our model routing guide covers how to pin models per task so a silent provider switch never surprises your workflow.
And if your bill makes you nervous, batch your non-urgent work. Batch processing prices are 50% lower on this model, and our Batch API guide walks through the setup step by step. Overnight jobs at half price is about as painless as AI savings get.
The takeaway
OpenAI just made its best-value model better for free, seven days after you got used to the last one. Switch your daily driver to GPT-6.1 Sol, keep Astra for the heavy stuff, and check your output token usage after the first week. The model’s verbosity is the one bill that can quietly grow.
Full pricing lives on the official OpenAI model page, and the complete benchmark breakdown is on Artificial Analysis. Test with your own tasks before you commit, because your workload isn’t their benchmark suite.
