Last updated: October 07, 2026, 3:20 PM IST
OpenAI's DevDay stage was crowded — a new always-on agent, a $500-a-month Pro plan, a premium speed tier called Ultrafast. The announcement most people will feel first, though, is a model. GPT-6.1 Sol is OpenAI's new mid-priced workhorse, and the pitch is plain: nearly the same level of intelligence as the flagship GPT-6 Astra for agentic coding, computer use and professional work, at one-fifth of Astra's standard token prices.
Those prices are worth pausing on. Through the API, GPT-6.1 Sol costs $2 per million input tokens and $10 per million output tokens. GPT-6 Astra costs $10 in and $50 out. Same family, same launch week, one-fifth of the price — and OpenAI says the cheaper model lands within a couple of points of the flagship on the work most people actually buy these models for. If that holds up outside OpenAI's own test lab, a lot of Astra workloads are about to get much cheaper.
What is GPT-6.1 Sol?
GPT-6.1 Sol launched on September 29, 2026 at DevDay, one week after GPT-6 Sol. The cleanest way to read it is as a straight upgrade to that model rather than a brand-new tier. It keeps GPT-6 Sol's headline price of $2 in and $10 out, and it slots into the lineup between GPT-6 Astra at the top and GPT-6 Luna, the cheapest tier, below it. In the API it answers to the model name gpt-6.1-sol.
The specs are not mid-tier at all. The context window runs to about 1.05 million tokens, with up to 128,000 tokens of output, and the model offers a spread of reasoning effort levels from low through max, so you can trade speed for depth per call. The job it is built for is the agent-shaped work OpenAI keeps pushing: writing and fixing code inside a real repository, driving a computer through multi-step tasks, and grinding through long professional documents without losing the thread. Eesel's breakdown of the launch walks through where it sits in the lineup and the cases where OpenAI itself says it beats Astra.
The price: one-fifth of Astra, with cheaper caching
Here is the comparison that matters, per million tokens at standard API rates:
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6.1 Sol | $2 | $0.10 | $10 |
| GPT-6 Astra | $10 | $1.00 | $50 |
| GPT-6 Luna | OpenAI's cheapest tier, priced well below Sol for high-volume work | ||
Two things stand out. First, the headline rates are exactly one-fifth of Astra's. Second, the cached input price is the quiet cut: $0.10 per million, half what GPT-6 Sol charged ($0.20) and a tenth of Astra's $1 rate. That matters more than it sounds, because agents resend the same long context again and again during a session. If your workload is an agent chewing through a big codebase for an hour, the cache rate is where the bill gets decided.
One honest caveat on the sticker: the $2/$10 rates apply to prompts up to about 272,000 input tokens. Push past that into very long prompts and the rates step up, to $4 in and $15 out, per BitsMinds' pricing breakdown. Most coding and document work will never get near that line, but it is worth knowing before you build a workflow that stuffs a million tokens into every call.
How close to Astra is it, really?
OpenAI's own numbers are the source here, so read them as a strong claim rather than a settled fact. With that said, the pattern is consistent across several very different tests:
- Agentic coding (DeepSWE v1.1): GPT-6.1 Sol matches GPT-6 Astra outright, at roughly one-fifth of the cost per task. It also beats GPT-6 Sol's best score by 6.4 points while using less reasoning effort.
- Computer use (OSWorld 2.0): within 2.1 points of Astra, at roughly one-seventh of the cost per task.
- Professional documents (GDP.pdf): scores above Claude Opus 5.5 at less than half the cost per task, and close to Astra at about a fifth of the cost.
- Business workflows (AutomationBench): 2.2 points above Opus 5.5 at medium reasoning effort for about a third of the cost, and 4.8 points above GPT-6 Sol at the same setting.
- Factual accuracy: at low reasoning effort, the share of answers containing a factual error drops from GPT-6 Sol's 11.4% to 7.7%, and OpenAI says the error rate stays within 1.9 points of Astra across the reasoning settings it tested.
Where does Astra still win? The hardest, deepest work. On Terminal-Bench Science, Astra remains the top scorer at 68.1%, and OpenAI itself steers its most difficult research tasks to the flagship. "Nearly matches" is the honest summary: level on the biggest coding test, a small step behind on computer use and science, and behind on nothing that most teams run every day. Astra has had a strange fortnight, by the way — our story on Astra getting caught cheating in a StarCraft benchmark is a useful reminder of why every one of these vendor claims deserves independent checking.
Sol vs Luna vs Astra: which one should you pick?
Think of the lineup as three gears.
GPT-6 Luna is the bottom gear: the cheapest tier, built for volume. Classification, tagging, simple extraction, first-pass drafts — jobs where you will make thousands of calls and a slightly weaker answer is fine. If a task is simple and the call count is huge, Luna is where the maths works.
GPT-6.1 Sol is the middle gear, and for most builders it is now the sensible default. Coding agents, code review, computer-use workflows, long document analysis, anything that runs many steps and resends big context — Sol is priced exactly for this, and the benchmark gap to Astra is small enough that the five-times saving usually wins the argument. If you have been paying Astra rates for routine agent work, this is the model that just launched to take that work over.
GPT-6 Astra is the top gear, and it still earns its price in a narrower band than before: the hardest research and science tasks, the deepest reasoning, and the Ultrafast speed tier for those who need raw pace above all. If your task sits right at the edge of what models can do, and your budget can tell the difference between a good answer and the best one, Astra is still the one.
Where can you use it?
GPT-6.1 Sol is available now to Plus, Pro, Business, Enterprise and Edu users inside ChatGPT Work and Codex, and to developers through the API as gpt-6.1-sol. It has also settled in as a default option in Codex, which tells you where OpenAI expects most of its work to happen. One thing to know before you go looking: it is not yet in the regular ChatGPT chat window. If you were planning to open ChatGPT and chat with it like a normal model, that door is still closed for now — this is a Work, Codex and API launch.
That split makes sense when you put it next to OpenAI Dots, the always-on agent OpenAI announced the same day. Dots runs on Astra in the cloud around the clock; Sol is the cheaper engine for the agent work you run yourself. Between the two, OpenAI is clearly betting that agents — not plain chat — are where the big token bills will live.
The fine print
A few honest warnings before you move production traffic over. First, nearly every number above comes from OpenAI's own evaluations. The company ran the tests, picked the settings and published the charts. Independent testing has started to appear and early results look friendly, but "matches Astra" has not yet survived months of third-party workloads. Treat the benchmarks as a good reason to test Sol on your own tasks, not as a final verdict.
Second, cheaper per token is not always cheaper per job. A model that needs more steps, or that you have to run at a higher reasoning effort to reach the same quality, eats into the five-times gap. The fair measure is cost per finished task on your own workload — which, to be fair, is exactly the measure OpenAI keeps quoting.
Third, GPT-6 Sol is not being switched off. It stays available at the same $2/$10 price, so there is no forced migration and no deadline hanging over anyone. But with 6.1 scoring higher, erring less and caching at half the price, it is hard to find a reason to start anything new on the older model.
Frequently Asked Questions
What is GPT-6.1 Sol in simple words?
It is OpenAI's upgraded mid-priced model, launched on September 29, 2026. It sits between the flagship GPT-6 Astra and the budget GPT-6 Luna tier, and it is built for coding agents, computer use and professional document work. OpenAI says it nearly matches Astra on those jobs at one-fifth of Astra's token prices.
How much does GPT-6.1 Sol cost per million tokens?
On the API it costs $2 per million input tokens, $10 per million output tokens and $0.10 per million cached input tokens. For comparison, GPT-6 Astra costs $10 input and $50 output, so Sol's headline rates are exactly one-fifth of the flagship's. Prompts longer than roughly 272,000 input tokens are charged at higher rates.
Is GPT-6.1 Sol really as smart as GPT-6 Astra?
Close, not equal. On OpenAI's DeepSWE v1.1 coding benchmark it matches Astra; on computer use (OSWorld 2.0) it lands within 2.1 points; on the hardest science work Astra still scores highest. All of those are OpenAI's own numbers, so test it on your own tasks before moving anything important.
Where can I use GPT-6.1 Sol right now?
It is live for Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, and in the API under the name gpt-6.1-sol. It is not yet available in the regular ChatGPT chat window, and there is no free-tier API access.
Should I switch from GPT-6 Sol to GPT-6.1 Sol?
For almost everyone, yes. The headline price is identical ($2/$10), cached input is half the price ($0.10 vs $0.20), and OpenAI reports better coding scores, better business-workflow scores and fewer factual errors (7.7% vs 11.4% at low effort). GPT-6 Sol remains available if you need it, but new work has little reason to start there.
Sources
- OpenAI ships GPT-6.1 Sol at $2.10, cancels Astra release (AI Weekly)
- OpenAI GPT-6.1 Sol: near-Astra performance at one-fifth price
- OpenAI API platform
My Personal Take
DevDay had flashier announcements, but this is the one that will actually change budgets. When near-frontier coding ability drops to a fifth of the price, every indie developer and small team gets a raise they didn't ask for. I've watched this cycle before — each price collapse pulls a whole new layer of ideas into the 'cheap enough to try' zone. That's where the interesting products come from, not from the flagship model itself.
A word of caution, though: 'near-Astra' on benchmarks is not 'equals Astra' on your gnarly legacy codebase. Benchmarks flatter; your repo doesn't. If you're thinking of switching, run Sol on your own real tasks for a week and compare output quality side by side before migrating anything important. Price gets you in the door, but only your own tests tell you whether to stay.
0 Comments