OpenAI has been on a run of new launches this pursuit of speed that OpenAI has and if it is costing them any intelligence previously as well. Now during their DevDay 2026, they have announced GPT-6.1 Sol exactly seven days after the original GPT-6 Sol announcement. As powerful as OpenAI’s models have been, they are replacing them with newer versions a little too quickly where we haven’t had much time to sit with even the first iteration. The question now is what all has OpenAI changed in this upgrade and how much of an improvement GPT-6.1 Sol can be from GPT-6 Sol. It’s an upgrade to GPT-6 Sol, the middle tier of the GPT-6 family, sitting between top-shelf Astra and budget-friendly Luna. It looks like a routine point release. It’s closer to a repricing of what counts as “good enough” for serious work.
Also read: OpenAI Dots explained: An AI coworker that never logs off
| Model | Pitch | Input (per 1M tokens) | Output (per 1M tokens) | Cached input (per 1M tokens) |
| GPT-6 Astra | Most intelligent, best results | $10 | $50 | $1 |
| GPT-6.1 Sol | Near-Astra intelligence, a fifth of the price | $2 | $10 | $0.10 |
| GPT-6 Luna | Fast, efficient everyday work at scale | $0.10 | $0.50 | $0.01 |
The cached input price is the sleeper. At $0.10 per million tokens, it’s 95% cheaper than Sol’s standard input and half of what GPT-6 Sol charged for cached tokens. Agents that keep re-reading the same codebase or the same 200-page PDF are the ones that win here.
Begin with coding, where the news is strongest. DeepSWE v1.1 sees GPT-6.1 Sol outclassing Astra on performance by around one-fifth of the price, and beating the best score that GPT-6 Sol has ever achieved on the benchmark by 6.4 percentage points, while doing all of this with less reasoning effort. Computer use shows the same pattern. On OSWorld 2.0, Sol is only 2.1 percentage points off Astra’s score at maximum effort and does this for one-seventh of the price per task, beating GPT-6 Sol’s score by seven percentage points.
Also read: SURI: The Seventh Note wants you to feel the beat through the PS5 DualSense
It’s business workflows where it faces a competing model. On AutomationBench, GPT-6.1 Sol beats Opus 5.5 by 2.2 percentage points at medium reasoning effort, at one-third of the cost. It’s science where Astra holds its title. On Terminal-Bench Science 0.1, Sol costs $5.47 per task on average, while Opus 5.5 costs $23.21 and Astra costs $23.80, yet Astra holds the top score at 68.1 percent. Even OpenAI recommends using Astra on the toughest research tasks. Factuality is improved, too. At low reasoning effort, the percentage of responses containing factual errors falls from 11.4 percent to 7.7 percent.
Looks like a clean sweep, but it isn’t a neutral scoreboard. OpenAI ran its own evaluations, and the competitor numbers come from public reports. The factuality test uses deliberately difficult prompts pulled from conversations where users flagged an earlier model’s mistake, so 7.7% is not what you’ll see in daily use. I am treating every figure as a vendor claim until I have tested the model myself.
The safety section is worth a skim too. OpenAI says Sol is better than before at admitting when a search tool is broken, failing to disclose it in 2.1% of test cases versus 4.9% for GPT-6 Sol. Astra still does better at 1.5%. Luna misses it in 28.7% of cases, which tells you what the cheapest tier costs you in reliability.
GPT-6.1 Sol is live for Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex. It’s not in standard Chat yet, so casual users won’t see it in the usual window. Developers can call it through the API as gpt-6.1-sol. A faster variant is on the way. OpenAI says GPT-6.1 Sol Ultrafast will arrive in the coming days, promising up to 8x faster token generation in Codex. For Indian developers and startups, where API bills hit harder, the cached-token pricing is the number to watch. Run your own workloads before trusting anyone’s chart, OpenAI’s included.
Also read: Best 43-inch TV deals on Amazon Great Indian Festival sale 2026