OpenAI's DevDay on 29 September was its largest yet, with more than 20 announcements. The headline was GPT-6.1 Sol, a mid-tier model that OpenAI says nearly matches its flagship GPT-6 Astra at one-fifth of the price. Around it came always-on personal agents, a much faster "Ultrafast" mode, a $500-a-month plan and a wave of Codex and ChatGPT updates. And in the background, the planned GPT-6.1 Astra upgrade was cancelled after internal safety tests. Here is what matters.
GPT-6.1 Sol in one chart
The interesting number is not the token price but the cost to complete real work. Artificial Analysis ran its full benchmark suite on both models at their maximum reasoning settings:
GPT-6 AstraGPT-6.1 Sol
Input and output in dollars per million tokens. Cost per task and Intelligence Index from Artificial Analysis, October 2026.
In other words: you give up about one point of measured intelligence and save around 78% on every task. On Artificial Analysis's coding-agent index inside Codex, GPT-6.1 Sol at its "xhigh" setting even beat Astra by a point, at under 15% of Astra's cost per task.
How it compares to the rest of the GPT-6 family
GPT-6 Astra
Flagship- Price
- $10 / $50
- Index
- 53
- Released
- 3 September
- OpenAI's most capable model
- Highest Terminal-Bench Science score of the three
- Five times the price of Sol
For the hardest problems only.
GPT-6.1 Sol
Best value- Price
- $2 / $10
- Index
- 52
- Released
- 29 September
- Near-Astra results on coding, computer use and professional work
- Cached input only $0.10 per million tokens
- Reasoning cannot be switched off
The new default for most API work.
GPT-6 Luna
Smallest- Price
- $0.10 / $0.50
- Index
- 37
- Released
- 22 September
- Cheapest model on the leaderboard
- Powers the new Decisions API
- Far behind on complex reasoning
Classification, routing and simple extraction.
A few practical details: GPT-6.1 Sol accepts text and images, has five reasoning settings from low to max (medium is the default), and its knowledge cut-off is 30 April 2026. In the API it is called gpt-6.1-sol.
Everything else announced at DevDay
- Dots
- Always-on personal agents running on GPT-6 Astra. Each has its own cloud computer and browser, connects to more than 4,000 apps, learns from feedback and asks permission for sensitive actions.
- Ultrafast
- A premium speed tier: up to 8 times faster in Codex (about 300 tokens per second) and up to 6 times faster in the API, at a higher price.
- Pro 500
- A $500-a-month ChatGPT plan with 25 times the Plus usage allowance and Ultrafast access.
- Decisions API
- Uses Luna to answer questions with a fixed set of answers, for classification, routing or choosing an agent's next step. In limited preview.
- Agents API
- Adds computer use, Codex-style multi-agent work, tool search and context compaction to the API.
- ChatGPT Space and Pages
- A shared workspace and a new document type for people and their agents to work on together.
- Codex Code Review
- An automatic first-pass review that posts comments on GitHub pull requests and GitLab merge requests, on all plans.
- OpenAI Marketplace
- Lets enterprises spend part of their OpenAI commitment on approved partner software, starting with 32 partners.
Trying it from Python
If you already use the OpenAI Python library, switching is a one-line change. This example asks the model to write a SQL query from a plain-English question and a schema, which is one of the most common data uses:
from openai import OpenAI
client = OpenAI() # reads OPENAI_API_KEY from your environment
schema = """
orders(order_id, customer_id, order_date, amount_gbp)
customers(customer_id, region, signup_date)
"""
response = client.responses.create(
model="gpt-6.1-sol",
reasoning={"effort": "low"}, # low is plenty for simple SQL
input=f"Schema:\n{schema}\n\nWrite one SQL query: monthly revenue by region in 2026.",
)
print(response.output_text)What it means for data teams
- 1
Re-run your cost maths
If you use GPT-6 Astra or another premium model for routine work, test GPT-6.1 Sol on the same prompts. A five-times price cut adds up quickly on batch jobs.
- 2
Use caching for long prompts
At $0.10 per million cached tokens, sending the same schema, documentation or style guide with every request is now almost free.
- 3
Separate easy and hard tasks
Route classification and tagging to Luna or the Decisions API, everyday work to Sol, and only the hardest questions to Astra.
- 4
Keep humans in the loop for agents
Dots and the Agents API can act on your behalf. Set clear rules for what they may do alone, as OpenAI recommends.
Check your understanding
1/3Roughly how much cheaper per benchmark task is GPT-6.1 Sol than GPT-6 Astra?
$0.72 versus $3.26 per task in Artificial Analysis's tests, while scoring only one point lower.
2/3Where is GPT-6.1 Sol not yet available?
At launch it was available in the API, ChatGPT Work and Codex, but not in standard chat.
3/3What are Dots?
Dots run on GPT-6 Astra, connect to thousands of apps and ask for approval before sensitive actions.
You got 0 out of 3.
See how GPT-6.1 Sol stacks up against Claude, Gemini and the open models in The best AI models in October 2026, compared.
Sources: OpenAI DevDay 2026 announcements as reported by Unite.AI and DeepLearning.AI's The Batch; Artificial Analysis benchmarks and cost per task; The Wall Street Journal via The Batch on GPT-6.1 Astra. Performance figures marked as OpenAI's are the company's own claims.
Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 14 October 2026. About me