GPT-6.1 Sol lands one point behind Astra at under a quarter of the cost per task, and the flagship gets harder to justify

A week after GPT-6 Sol arrived as OpenAI’s cheaper tier, GPT-6.1 Sol replaces it at the same list price and closes most of the distance to the flagship. OpenAI’s pitch is “near-Astra” performance for coding and professional work. Independent measurements largely back that up, and they suggest the real saving is larger per completed task than the per-token price alone implies.

What OpenAI shipped

GPT-6.1 Sol is available in the API as gpt-6.1-sol on the Chat Completions, Responses, and Batch endpoints, and in Codex and ChatGPT’s paid work plans. It lists at $2 per million input tokens and $10 per million output tokens, unchanged from GPT-6 Sol, with cached input cut in half to $0.10. The context window is 1.05 million tokens with up to 128,000 output tokens, and it exposes five reasoning effort levels from low to max, with medium as the default.

OpenAI frames the model as approaching GPT-6 Astra on coding, debugging, document understanding, and multistep workflows, at one fifth of Astra’s per-token price. The price ratio is not new: GPT-6 Sol already sat at a fifth of Astra’s $10 and $50 rates. What changed is how much capability that cheaper tier now buys.

For developers already on GPT-6 Sol, this is close to a free upgrade: the same rate card with a stronger model behind it, available by changing one model ID.

What the independent numbers show

Artificial Analysis, which benchmarks models independently of their vendors, scores GPT-6.1 Sol at maximum effort at 52 on its Intelligence Index. GPT-6 Astra at maximum effort scores 53, and GPT-6 Sol scores 48. On that aggregate, the new model recovers most of the gap between OpenAI’s two tiers in a single point release.

The cost data is where the story sharpens. Artificial Analysis measured GPT-6.1 Sol at $0.72 per Intelligence Index task, against $3.26 for Astra and $1.05 for GPT-6 Sol. GPT-6.1 Sol used 67 million output tokens to complete the index, somewhat more than Astra’s 60 million but fewer than GPT-6 Sol’s 77 million. So the model is not winning by being terse. It spends slightly more tokens than the flagship, but at a fifth of the rate, which still leaves it at roughly 4.5 times cheaper per task. It is also faster in these tests, at about 66 output tokens per second against Astra’s 51.

The comparison with GPT-6 Sol matters too. The new model is about 30 percent cheaper per task than its predecessor at the same list price, a result of spending about 13 percent fewer output tokens and paying half as much for cached input.

Per-token price tells you the rate; tokens per task tell you the bill, and on both counts 6.1 Sol moves in the right direction.

What it means for developers

The practical decision is about routing. Teams that defaulted to Astra for coding agents because GPT-6 Sol fell noticeably short now have a cheaper candidate that independent testing places within a point of it. For agent loops that run many steps per task, a gap of roughly 4.5 times in cost per task changes budgets, not just line items.

The caveats are real. The Intelligence Index is an aggregate, and a one-point gap on average can hide larger differences on specific workloads, such as long refactors across a large repository or tasks where Astra’s extra reasoning pays off. OpenAI’s coding-specific claims come from its own evaluations. The sensible move is to run GPT-6.1 Sol against the flows where Astra is currently the default, measure pass rate and cost per completed task, and keep Astra only where the difference shows up in your own results.

For teams on GPT-6 Sol, the case is simpler: test the new ID on existing evals and switch if nothing regresses.

What to watch

The open questions are whether the near-parity holds on coding-specific and agentic benchmarks as independent results accumulate, whether OpenAI adjusts Astra’s pricing or positioning now that its own cheaper tier crowds it, and how quickly coding tools make 6.1 Sol their default. The broader pattern from the past month continues: the frontier is increasingly a price and efficiency contest, and the flagship tier has to earn its premium on specific workloads rather than by default. Stackmaven’s follow-up will revisit around December 29.

Sources cited

1. GPT-6.1 Sol model reference (OpenAI API docs) developers.openai.com
2. Introducing GPT-6.1 Sol (OpenAI) openai.com
3. GPT-6.1 Sol (max): intelligence, performance and price analysis (Artificial Analysis) artificialanalysis.ai
4. GPT-6 Astra (max): intelligence, performance and price analysis (Artificial Analysis) artificialanalysis.ai
5. GPT-6 Sol (max): intelligence, performance and price analysis (Artificial Analysis) artificialanalysis.ai
6. OpenAI launches GPT-6.1 Sol, says it nearly matches GPT-6 Astra and costs less (TechCrunch) techcrunch.com