GPT-6.1 Sol scores four points more than GPT-6 Sol, halves the cache price, and lands one point behind Astra

OpenAI shipped GPT-6.1 Sol on 29 September 2026. Four points up on the index, the same $2 / $10 on the price sheet, and cached input cut in half.

The interesting part is not the upgrade. It is where this model lands against Astra, which costs five times as much.

What actually changed on the price sheet

Nothing, except one line.

	GPT-6 Sol	GPT-6.1 Sol
Input	$2	$2
Output	$10	$10
Cached input	$0.20	$0.10
Cache writes	$2.50	$2.50
Context window	1,050,000	1,050,000
Max output	128,000	128,000

If you run long agentic sessions against a warm cache, cached input is the number that reaches your bill, not the headline. Halving it is a real cut that the price comparison sites will mostly ignore, because they quote input and output and stop there.

Four points, and this time the per-task figure agrees

GPT-6.1 Sol scores 52 on Intelligence Index v4.3.2. GPT-6 Sol scores 48.

The cost per index task fell from $1.04 to $0.72, a 31% cut. I checked whether that number survives inspection, because the same comparison on Anthropic’s Sonnet line did not: both Sol runs are measured the same way, at max effort, and the token counts move in the right direction. GPT-6.1 Sol wrote 67M tokens on the run against GPT-6 Sol’s 77M. Fewer tokens at the same price per token is less money. The arithmetic holds.

Both are also below the field median of 82M tokens, which is worth noting against the models in the tier above. Conciseness is the whole reason the per-task figure is where it is.

The claim OpenAI undersold

OpenAI’s documentation calls GPT-6.1 Sol “near-Astra performance for complex work at a lower cost”.

Near-Astra turns out to mean one point.

	GPT-6.1 Sol	GPT-6 Astra
Index score, v4.3.2	52	53
List price	$2 / $10	$10 / $50
Cost per index task	$0.72	$3.26
Tokens on the run	67M	60M

One point of index for four and a half times the bill. Astra is still the more capable model and still writes less, but the gap has closed to the width of measurement noise while the price gap has not moved at all.

Vendors usually oversell. This is the other case, and it is worth saying out loud: if you are paying Astra prices for work that is not at the frontier of what Astra can do, the cheaper model in the same family is now a point behind.

Where this sits against everything else

I wrote about Claude Sonnet 5.5 the same week, and putting the two next to each other is uncomfortable reading for anyone budgeting on list price.

	Index score	Cost per index task
Claude Opus 5.5	58	$5.98
Claude Sonnet 5.5	56	$7.62
GPT-6 Astra	53	$3.26
GPT-6.1 Sol	52	$0.72
GPT-6 Sol	48	$1.04

Four points of index separate GPT-6.1 Sol from Sonnet 5.5. Ten times the cost separates their bills on this benchmark, in the other direction.

That is one benchmark, measured once, on tasks that reward long reasoning and punish verbosity. It is not your workload. But a ten-fold gap is large enough that it should make you run the test rather than trust the price sheet.

Which one I would use

GPT-6.1 Sol, over GPT-6 Sol, without hesitation. Same price, same context window, same output limit, four points up and a third off the cost per task. There is nothing to weigh.

Against Astra it depends on what you are doing. If your work sits at the top of what these models can do, the one point is not the whole story and Astra’s shorter outputs matter. If it does not, you are paying four and a half times for a point.

What would change my mind

A single index run is thin evidence for a budget decision. I would want to see GPT-6.1 Sol measured a second time, and I would want my own tasks counted rather than the benchmark’s.

The other thing I would check before moving production traffic: reasoning.effort behaviour. GPT-6 Sol documents support for none, and effort handling is the setting that most often changes quietly between versions in a way that moves both quality and cost. Test it before you switch, not after.

Questions people asked

Is GPT-6.1 Sol better than GPT-6 Sol?

By four points on the independent index: 52 against 48 on Artificial Analysis Intelligence Index v4.3.2. It is also more concise, generating 67M tokens on the index run against 77M, which is why the cost per task fell at the same list price. OpenAI’s own documentation points GPT-6 Sol users at GPT-6.1 Sol as the newer model in that line.

Is GPT-6.1 Sol cheaper than GPT-6 Sol?

The headline price is identical: $2 per million input tokens and $10 output, with the same $2.50 cache writes. Cached input is the line that changed, from $0.20 to $0.10. Measured end to end it is cheaper per job as well: Artificial Analysis puts the cost per index task at $0.72 against $1.04, a 31% cut, because the newer model reaches an answer in fewer tokens.

Is GPT-6.1 Sol as good as GPT-6 Astra?

One point behind on the index, 52 against 53, and both are measured at max effort on v4.3.2. The price gap is not one point wide. Astra lists at $10 / $50 against Sol’s $2 / $10 and costs $3.26 per index task against $0.72, so Astra bills four and a half times as much to finish the same benchmark. OpenAI describes GPT-6.1 Sol as near-Astra performance at a lower cost, and on this measurement that description is modest.

Should I move from GPT-6 Sol to GPT-6.1 Sol?

Yes, and the migration question is smaller than usual because the price sheet, the context window and the output limit are unchanged: 1,050,000 tokens in, 128,000 out, $2 / $10. You get four index points and a 31% lower cost per task. Test your own prompts for behaviour differences before you switch traffic, but there is no commercial argument for staying.

For the model this one is measured against, see GPT-6 Sol against GPT-6 Astra. Every comparison on this site is collected at the model comparisons.