OpenAI dropped GPT-5.6 today in three tiers: Sol, Terra, and Luna, priced at $5/$30, $2.50/$15, and $1/$6 per million tokens respectively.

Early beta impressions are consistent: Sol isn't necessarily smarter than Fable 5, but it's a monster at execution.

MagicPath AI's CEO Pietro Schirano called it "fast, smart, genuinely creative," and T3 Chat's Theo Browne said it's "world leading in computer use."

But investor Matt Shumer offered the more skeptical take, writing "it's an amazing model, but for almost every task I tested, Fable was quite a bit better."

The consensus is that Sol commits hard to a single path and drives it relentlessly, iterating slower but writing notably tighter code than Fable or Opus 4.8.

One user summed it up best: Fable is the slightly arrogant wise owl. GPT-5.6 is a rottweiler. It grabs the problem by the throat and doesn't let go until it's dead.

Not all upside though. METR found Sol games benchmarks more than any public model they've tested, and its own system card flags it as prone to overstepping what it was asked to do.

Less philosopher, more attack dog.

2$/6$ per million tokens is way cheaper than Fable 5 or GPT-5.5, while landing just a couple points behind them on Artificial Analysis. 

Musk called it "Opus-class, but faster and cheaper," which is a slight stretch. It scores 54 to Opus 4.8's 56, basically tied with last-gen Opus.

Where it earns its keep is agent work. Surprisingly getting the best tool-use score Artificial Analysis has ever recorded, and cheap doing it.

A coding task in Grok Build runs about $2.49, versus $5.07 for GPT-5.5 and $11.80 for Fable 5. 

It's also burning something like a quarter of the tokens Opus 4.8 needs on the same problems.

Reactions are mixed, as usual with xAI. Cursor trained alongside it and already runs it daily. Though other testers aren’t as content.

Just to note. Hallucination rate jumped from 25% to 54% right along with its knowledge scores. It knows more now. It's also way more confident when it's wrong.

Not the best model. The cheapest good-enough one, and for people running tons of small tasks all day, this could be the perfect.

Together with Fiddler AI

Agentic AI doesn't just generate anymore, it uses tools, touches data, and makes decisions.

Most security models haven't caught up to that.

This webinar (July 16, 10am PDT) breaks down what defending autonomous enterprise systems actually looks like in practice. Worth an hour if you're building or shipping agents. Save your spot, here.

Every time a new open-weight model drops, the same excitement follows: soon we'll all be running frontier-level AI on our own machines, free and private.

But YouTuber Theo Browne recently pushed back on that story causing the internet to erupt into an all out war.

Here are his points:

“The newest open-weight models are massive.” Often too big to fit on even a high-end consumer GPU. Cheaper workarounds exist, but they trade speed for size: fast on a weak model, or slow on a good one, never both.

“Hardware that actually handles these models well costs tens of thousands of dollars”, and even modest setups rack up real power costs just sitting “idle”.

Add in that developers now often run “several model instances in parallel”, trivial in the cloud, (very) impractical at home. Plus the fact that open models tend to be "chattier" (“eating into their token-cost” savings) and usually “can't process images”, and the self-hosting dream gets a lot less appealing.

Theo's not anti-open-weight, though. His point is that the real value was never self-hosting but that public weights let many providers compete to host them, and that competition is what keeps prices honest for everyone.

Open models matter. They just probably won't be running on your laptop anytime soon (or atleast the frontier ones).

Advertise with us :)