My token spend got cheaper overnight.
That’s the short version of what happened when GPT-6 Sol and Luna landed on September 22, 2026, priced at half of what their GPT-5.6 predecessors cost. For anyone who ships bots for a living, that sentence matters more than any benchmark chart. I’ve spent the last few years watching the cost side of agent architectures quietly eat project budgets. A pricing cut of that size changes what you can build, not just what you can afford.
Let me lay out what we actually know, because the rumor mill around “GPT-6.1 Sol” is running hot and the sources don’t back it up yet.
What shipped, and what didn’t
The GPT-6 family has three members: Astra, Sol, and Luna. Astra is the flagship. Sol and Luna are the ones positioned as near-frontier intelligence at significantly reduced prices. Astra arrived September 3, 2026. Sol and Luna followed on September 22.
On the numbers side, Sol posted 68.8% on DeepSWE and carries a 1.05M-token context window. Astra’s pricing went to $10 input and $50 output per million tokens, which is 2.5 times more than GPT-5.6 Sol cost across the tiers that existed before.
As for a GPT-6.1 Sol or any further release, the sources don’t give a definitive answer. Polymarket traders are betting on it, which tells you about sentiment, not about shipping dates. I’d treat the prediction market odds as market noise until OpenAI says something. Building a roadmap around a bet is how you end up rewriting your agent loop for a model that never arrives.
The DeepSWE number is the one I care about
68.8% on DeepSWE isn’t a marketing stat. It’s a software engineering benchmark, which means it’s testing something close to what most of us actually ask these models to do: read a repo, understand context, make a change that doesn’t break three other things.
For bot builders, that translates directly. If your agent writes and patches code as part of its job, a model that clears roughly two thirds of real engineering tasks is a different tool than one clearing a third. The gap between “I need a human reviewing every diff” and “I need a human reviewing flagged diffs” is where the actual productivity lives.
The honest caveat: benchmark performance and production performance diverge. Your prompts, your retrieval quality, your tool definitions, and your error handling will move your results more than the model swap will. I’ve seen teams gain more from fixing a retrieval step than from upgrading a model tier.
What a million-token window actually changes in your architecture
1.05M tokens of context is enough to stop treating context as a scarce resource in a lot of designs. That’s the interesting part. Some of the chunking and summarization machinery we’ve been writing for years exists purely to work around small windows.
Things worth reconsidering if you’re on Sol:
- Aggressive conversation summarization. If a support bot’s entire session history fits comfortably, the summarizer is a lossy step you may not need.
- Multi-hop retrieval chains built to fetch minimum viable context. You can often widen the retrieval and let the model sort it out.
- Splitting long documents across multiple calls, then stitching results. One call with the whole document is simpler and usually more coherent.
Two warnings from experience. First, a big window is not a free window — you still pay per token, and stuffing 900K tokens into every request will make the price cut evaporate fast. Second, attention quality across a very long context isn’t uniform. Position your most important material deliberately rather than dumping and hoping.
How I’d tier a real system right now
The three-model lineup invites a routing strategy, and I think that’s the right read. Sol and Luna sit at near-frontier capability for less money. Astra sits above them at a steep premium.
My default plan looks like this. Route the bulk of traffic to Sol. Classification, extraction, summarization, routine tool calls, most code edits — that work doesn’t need the flagship. Reserve Astra for the tasks where a wrong answer is expensive and hard to detect: complex multi-step planning, ambiguous user intent, anything touching money or permissions.
Then measure. Build an eval set from your own traffic before you route anything, because the only comparison that matters is Sol versus Astra on your workload. Half the time the cheaper model wins outright and you’ve just cut a line item.
The practical takeaway
Price drops of this size reset assumptions. Features you shelved because per-request cost didn’t pencil out deserve a second look, and that’s true whether or not a 6.1 release ever shows up.
Start with pricing math on your top three highest-volume endpoints. That’s a cheap afternoon of work with a clear payoff, and it doesn’t depend on a single prediction market bet coming in.
🕒 Published: