The headline going around right now is that GPT-6.1 Sol delivers almost-Astra intelligence for a fifth of the cost. I’ve been digging through what’s actually verifiable, and I have to be blunt: there is no confirmed 6.1 Sol. What exists, and what I can price out, is GPT-6 Sol, which shipped alongside Luna on September 22, 2026. The interesting part isn’t a version number nobody has shipped. It’s that the cheap tier already won, and most people building with these models haven’t adjusted their defaults yet.
Run the arithmetic before you run the benchmarks
OpenAI released GPT-6 Astra on September 3, 2026 as the flagship, and it came with a price increase, not a cut. Input went from $4 to $10 per million tokens. Output went from $20 to $50. That’s 2.5 times the cost of GPT-5.6 Sol across the tiers people were already budgeting for.
Then Sol and Luna landed a few weeks later at half the price of their GPT-5.6 predecessors. Per Artificial Analysis, GPT-6 Sol in its non-reasoning configuration sits at $2.00 input and $10.00 output per million tokens, with cached input at $0.20.
Line those up and the “fifth of the price” claim holds, just not for the model people are attributing it to:
- Astra: $10 input, $50 output
- Sol: $2.00 input, $10.00 output, $0.20 cached
Exactly one fifth on both sides of the ledger. That ratio is the story. A 5x price gap between two models in the same family, released nineteen days apart, is the kind of spread that should force a hard look at every API call you’ve hardcoded to the flagship.
What Sol actually brings to the table
Two numbers matter for anyone wiring Sol into a toolkit. The first is 68.8% on DeepSWE, released with Sol on that September date. The second is the 1.05M-token context window, the same figure reported for Astra’s context.
That second point deserves emphasis because context length is where cheap tiers usually get quietly nerfed. You pick the budget model and discover it caps out at a fraction of the flagship’s window, which means your document pipeline or your repo-wide code task breaks in ways that are annoying to diagnose. Sol matching Astra’s 1.05M figure removes that particular trap.
Latency on Sol’s non-reasoning config comes in around 0.93 seconds per the same Artificial Analysis listing, and it handles text and image input. For a model at a fifth of flagship pricing, none of that reads like a compromise tier. It reads like the tier most builds should default to.
Where the flagship still earns the premium
I’m not going to pretend the 5x gap means Astra is a bad buy. Astra shipped with native computer-use capability, which is a different category of work than text generation and code assistance. If your product depends on an agent driving an interface, that’s a reason to pay flagship rates, and Sol’s pricing doesn’t change that calculus.
The honest framing is that Astra is now a specialty purchase rather than the obvious starting point. Reach for it when you need what only it does. Reach for Sol when you need good output at volume, which for most of the tools I review is the actual workload.
About that 6.1 rumor
Polymarket has an active market on which lab releases the next Sol-class model in the 6.1+ range, and September 2026’s rollout pace is what’s driving trader interest. Three releases in under a month will do that.
But a prediction market is a market, not a changelog. The sources I can check do not confirm a GPT-6 Sol 6.1 release, or any specific future release in that family. If you’re seeing benchmark tables and pricing for 6.1 Sol circulating right now, treat them as unsourced until OpenAI publishes something. I’ve watched too many tool roundups get rewritten after a rumored model failed to materialize on the rumored schedule.
What I’d actually do this week
Audit which calls in your stack are hitting the flagship out of habit. The 5x spread between Astra and Sol is large enough that the audit pays for itself quickly, especially if your workload has cacheable prefixes, where Sol’s $0.20 cached input rate does real work on the bill.
Then test rather than trust. A DeepSWE figure of 68.8% tells you something about Sol’s coding capability in aggregate, but it doesn’t tell you how it handles your codebase, your prompt structure, or your tolerance for retries. Run your own evals on the tasks you care about and compare the output quality against what you’re currently paying Astra rates for.
The version number people are excited about may or may not arrive. The 5x price gap is already here, already documented, and already worth acting on.
🕒 Published: