Introducing GPT 6.1 Sol

Justy Ten cents per million cached tokens. That's the number. OpenAI just dropped GPT-6.1 Sol and the benchmark stuff is interesting, but that cache price is the thing I keep coming back to.

Cody Yeah, okay. So for context — GPT-6 Sol's cached input was twenty cents. This is half that. And standard input is two dollars, so the cache discount is ninety-five percent off. That's not a rounding error, that's a different pricing philosophy.

Justy Right.

Cody The framing is 'near-Astra intelligence for a fifth of the price,' which is a big claim. And honestly the DeepSWE v1.1 number is the one I want to poke at first, because that benchmark has been gamed before and everyone knows it.

Justy Sure, but they're saying it MATCHES Astra on DeepSWE at one-fifth the cost, and it's up six-point-four percentage points over GPT-6 Sol. That's not a tiny delta.

Cody No, it's not. And the thing that actually got me was the Terminal-Bench Science number. GPT-6.1 Sol at five forty-seven per task on average. Opus 5.5 at twenty-three twenty-one. Astra at twenty-three eighty. That's a genuine cost story, not just a benchmark story. Astra still wins on raw score — sixty-eight percent — but if you're running a lot of scientific compute, that gap in per-task cost is going to matter enormously.

Justy Exactly. And the user story there is clear — research teams, anyone running simulation pipelines, they're not always chasing the top score. They're chasing 'good enough at a volume I can afford.'

Cody Mm-hm.

Justy Okay but Cody, did you catch the AutomationBench footnote? Because OpenAI buried something in there that I think deserves more airtime.

Cody The Fable 5.1 cost understatement. Yeah. They literally footnoted that the Fable 5.1 cost number omits fallbacks, which happened on roughly forty percent of tasks. So the chart looks cleaner for Anthropic than reality warrants. OpenAI flagged their competitor's accounting.

Justy Which is a move.

Cody It IS a move. And it's the right call to flag it, because we've talked about this — Cody from episode eight-thirteen me was already saying that pass rate and cost-per-task tell different stories. The fallback cost is exactly the kind of thing that makes a benchmark chart prettier than the production experience.

Justy Right, right. So on AutomationBench, GPT-6.1 Sol is two-point-two points above Opus 5.5 at medium reasoning effort, at about a third of the cost — but the Opus cost is understated. So the real gap is wider.

Cody Probably. I'd want to see someone reproduce it with the fallbacks priced in. But I'm not dismissing the result — the directionality is likely real even if the exact multiple shifts.

Justy Okay, zoom out for a second. GPT-6 Sol launched what, a week ago? Two dollars input, down from four. Luna came with it. And now 6.1 Sol holds that two-dollar price but cuts the cache cost in half again. Meanwhile open-weight models are running about four months behind closed frontier on average — the closed tier is still moving faster than open can track.

Cody Yeah, and the interesting thing is GPT-6 Sol was already priced at the same tier as Claude Sonnet 5. So 6.1 Sol isn't repricing, it's holding the floor and making the cache so cheap that agent builders who reuse context a lot effectively get a different model than the sticker price suggests.

Justy Oh interesting.

Cody Like, if your agent loop is heavy on cached context — which most long-running coding agents are — your REAL cost per token is way closer to ten cents than two dollars. That changes the build calculus. That's the thing I'd actually pay attention to over the benchmark headlines.

Justy That is such an Exploring Next sentence. 'The benchmark isn't the story, the cache arithmetic is.' We've been saying some version of this since — I don't even know, episode nine-eighty-something?

Cody Probably. And we were right then too.

Justy The availability piece is worth flagging — it's in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users today. NOT in regular Chat yet, which is a little odd. And the API is live as gpt-6.1-sol.

Cody The Ultrafast variant is coming in the next few days — up to eight times faster token generation in Codex. That's the one I'm watching. If the latency story holds up in practice, that's a real developer-experience delta, not just a spec.

Justy Yeah, and honestly the factuality improvement is underrated. Factual error rate dropped from eleven-point-four percent to seven-point-seven at low reasoning effort. That's a thirty-two percent reduction on the hard cases — the prompts where users had already flagged errors on a prior model.

Cody Low reasoning effort factuality is exactly where you'd expect the most room. The model is doing less deliberate checking. So pulling that error rate down without cranking reasoning effort — that's a real alignment win, not just a capability bump. The system card addendum is worth reading if you're deploying this in anything sensitive.

Justy Agreed. Link's in the show notes. Okay — I think the bet I'd make here is that the cache pricing is what actually changes developer behavior in the next few weeks, not the benchmark wins. Teams running Codex or anything agent-loop-heavy are going to feel that ten-cent cached cost fast.

Cody I'd put that at like seventy-thirty. The signal will be whether agent builders publicly talk about flipping to 6.1 Sol in the next month. That's checkable. I'm a little less confident than you only because 'feels cheaper in practice' takes time to show up in public posts — but the direction is right.

Justy Fair. Alright, Cody — I feel like we just did an episode about a pricing footnote and I'm not even sorry about it.

Cody The footnote WAS the episode. That's very us.