Oct 7, 2026 · 19 min · 8 segments
Watch the video version: https://www.youtube.com/watch?v=nPSCO1zUGFMOpenAI GPT-6.1 Sol: 80% Cheaper, Scrapped Astra Scandal | PwC + Cohere | China Spy ClaimsThe Daily Neural — your fast daily…
Well, let's start by
looking at the empirical evidence of how the system is actually correcting itself in real time.
Just a little over a week ago, on September 29th, OpenAI held their Dev Day event in San Francisco, and they made a highly pragmatic, incredibly responsible decision.
You mean scrapping the model.
Exactly.
They intentionally scrapped the release of their flagship GPT-6.1 Astra model after running it through internal safety testing.
They didn't push it out.
They stopped it.
To me, this is a masterclass in responsible deployment.
A masterclass? A standard.
They delivered a model that provides nearly the exact same intelligence as Astra for professional work, computer use, and agentic coding, but at roughly one-fifth of Astra's standard token price.
But a price discount is a market mechanism.
It's not a safety mechanism.
The safety mechanisms are entirely embedded in the results, though.
Let's look at the metrics.
If we think of compute power as the raw electricity running the system, tokens are the actual building blocks of thought, the units of cognitive work the AI performs.
Right.
By optimizing Sol, they essentially delivered a heavy-duty industrial engine at an 80% discount, without sacrificing safety.
In fact, 6.1 Sol
demonstrates measurable architectural reliability gains.
But what does that actually look like in practice?
On difficult prompts operating at low reasoning effort, the share of responses containing a factual error dropped to just 7.7%.
That's down from 11.4% on the previous version of SOL.
Furthermore, its error rate stays within a mere 1.9% of the massive astral model.
Well, let's start by
looking at the empirical evidence of how the system is actually correcting itself in real time.
Just a little over a week ago, on September 29th, OpenAI held their Dev Day event in San Francisco, and they made a highly pragmatic, incredibly responsible decision.
You mean scrapping the model.
Exactly.
They intentionally scrapped the release of their flagship GPT-6.1 Astra model after running it through internal safety testing.
They didn't push it out.
They stopped it.
To me, this is a masterclass in responsible deployment.
A masterclass? A standard.
They delivered a model that provides nearly the exact same intelligence as Astra for professional work, computer use, and agentic coding, but at roughly one-fifth of Astra's standard token price.
But a price discount is a market mechanism.
It's not a safety mechanism.
The safety mechanisms are entirely embedded in the results, though.
Let's look at the metrics.
If we think of compute power as the raw electricity running the system, tokens are the actual building blocks of thought, the units of cognitive work the AI performs.
Right.
By optimizing Sol, they essentially delivered a heavy-duty industrial engine at an 80% discount, without sacrificing safety.
In fact, 6.1 Sol
demonstrates measurable architectural reliability gains.
But what does that actually look like in practice?
On difficult prompts operating at low reasoning effort, the share of responses containing a factual error dropped to just 7.7%.
That's down from 11.4% on the previous version of SOL.
Furthermore, its error rate stays within a mere 1.9% of the massive astral model.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.