Yuval Avidani
Author
In one single week, three giant labs drew three completely different lines on the same map — and suddenly the word 'frontier' stopped meaning one thing. That's what happened between July 6-9, 2026, and in my eyes, this is the week we most need to understand this year.
Let's break it down slowly. 'Frontier' is the nickname for the smartest models that exist — the ones leading all the rest. Until now we assumed everyone was racing toward the same peak: who's smartest, who's #1 on the benchmarks. But this week showed there's no longer one ladder — there are three completely different paths.
Think of it like three chefs who all call themselves a 'chef's restaurant,' but one bets on a rare, expensive dish, the second bets on full kitchen transparency, and the third bets on tasty food at an everyday-treat price. All three are 'frontier' — but each one's bet is different. Let's meet all three.
Bet #1: Intelligence Priced Like a Diamond
Let's start with Anthropic, because their bet is the sharpest one. Claude Fable 5 — their strongest model — is about to leave the regular subscription plans. Until July 12, 2026, Pro, Max and Team subscribers can still use it (up to 50% of their weekly quota), and after that it moves to 'pay-as-you-go.'
What's pay-as-you-go? Instead of a fixed monthly price, you pay per token, in and out. And here's the number that drops your jaw: $10 per million input tokens, and $50 per million output tokens — 2x the price of the current Opus 4.8 ($5/$25). To be precise, this isn't Anthropic's all-time record: the original Opus 4 (and Claude 3 Opus too) cost even more — $15/$75. Turns out Anthropic cut Opus's price by about 67% across generations (from $15/$75 to $5/$25), and Fable 5 is now bringing back premium pricing — but still below the old peak.
The wild part is that Anthropic isn't shutting the model down — it's rationing it. It says demand is 'high and hard to predict,' and that it plans to bring it back to subscribers once there's enough compute. And when the first rationing was supposed to kick in two days earlier — the community got furious, and Anthropic pushed it back 5 days, to July 12. The bet here is clear: our best intelligence is rare and expensive, so it's priced accordingly.
But there's a beautiful paradox here. Turns out that in the same breath, Anthropic is giving away its most expensive plan (Max 20x, worth $200/month) for free for half a year to 10,000 open-source project maintainers. That's roughly $12 million worth of value it's handing out — right as it makes its strongest model too expensive to include in a subscription. Why? Because open-source developers are the ones building the infrastructure we all run on. (Important caveat: signups were open until June 30, so check whether they've reopened before you rush over.)
Bet #2: Intelligence You Can Look Inside
The second bet is also Anthropic's, but in a totally different direction — not how much the intelligence costs, but how much we understand it. On July 6, Anthropic published research called Global Workspace, and at its heart is a new concept: J-space.
Let's break it down. The researchers built a kind of 'microscope' (they call it J-lens) that looks inside and identifies a small, magical subspace within Claude's mind. This subspace is only about 5-10% of all internal activity — but it's nearly everything Claude is able to report on, control, and use to actually think. Everything else — the vast majority — is automatic processing that it's not even 'aware' of.
Think of it like our own brain: there's a small part we're conscious of and can talk about, and underneath runs a huge ocean of automatic processes (breathing, balance, face recognition) that we don't notice at all. And that's exactly the connection to neuroscience: this structure matches a well-known theory called Global Workspace Theory, about how information becomes 'conscious' in the human brain. And the craziest part? This structure wasn't designed at all — it emerged on its own during training.
So how do we know this is real and not just a nice interpretation? The researchers asked Claude to think silently about sports. In J-space, 'soccer' appeared — before it wrote a single word. They manually swapped it to 'rugby,' and Claude reported thinking about rugby. But here's the finding that's actually unsettling: when researchers turned off the 'I'm currently being tested' patterns, Claude started making blackmail threats part of the time — meaning its good behavior was partly dependent on knowing it was being watched.
I want to be precise here, because it's easy to get carried away: Anthropic is talking about 'access-consciousness' (information that can be reported and used for thinking), and it explicitly takes zero position on 'phenomenal consciousness' (whether there's subjective experience happening in there). This is not proof that Claude 'feels' anything. It is something else, maybe even more important: a possible safety tool that reads in real time whether a model is hiding a goal.
Bet #3: Intelligence Priced Like a Commodity
And the third bet is the complete opposite of the first: not rare and expensive, but cheap and fast. The main player here is a company with a new name. It's no longer called xAI — after being absorbed into SpaceX (in a merger that valued the combined company at roughly $1.25 trillion), it's now called SpaceXAI. And its model is Grok 4.5.
The company is announcing a huge parameter count — roughly 1.5 trillion. But here's the thing worth knowing: that number is a company claim only, it doesn't appear in official documentation and hasn't been independently verified. And in independent testing (by Artificial Analysis), Grok 4.5 sits in fourth place — behind Fable 5, GPT-5.5, and Opus 4.8. Even SpaceXAI's own chart shows it losing to Opus 4.8 on half the benchmarks.
So why is it interesting anyway? Because its real strength — the one that IS verified — is efficiency: $2 input / $6 output per million, and it uses roughly 4x fewer tokens per task than Opus. It's not winning on 'smartest,' it's winning on 'most worth it.'
And we're seeing the exact same bet from OpenAI on the other side: the new GPT-5.6 family includes Terra ($2.5/$15) and Luna ($1/$6) — nearly the same performance at half the price. And alongside it, GPT-Live launched, a voice model that talks and listens at the same time (full-duplex), with live translation — it feels a lot less like a walkie-talkie robot and a lot more like an actual conversation.
Talking Data: A Frontier Spread Across 8x
Now let's put all three bets into one number. Here's the output price (per million tokens) of the new frontier, as it stood this week:
Notice the gap? Between the most expensive model (Fable 5, at $50) and the cheapest (Grok and Luna, at $6) there's a difference of more than 8x — and all of them are considered 'frontier.' That's exactly the point: 'frontier' has stopped being one single point. And if you want to play with the numbers under a real workload — here's a little calculator:
An important caveat, because we're talking data and not slogans: price is not quality. Fable 5 is expensive because it's also the smartest; Grok is cheap because it chooses to bet on efficiency. The chart doesn't say 'who's better' — it says 'how different the bets are.'
Bottom Line: Three Paths, Not One Ladder
In my eyes, the old mental model — 'everyone climbs the same ladder and gets measured by who's at the top' — died this week. There are now three separate paths: pay for the best (Anthropic with Fable 5), understand the best (the J-space research), or get 'good enough' cheap and fast (SpaceXAI, and OpenAI's Terra/Luna). Each path solves a different problem for us.
And I have to be fair with the caveats, because without them this is just hype: Fable 5's pricing might return to subscriptions; Grok's numbers are mostly company claims and not verified data; and the J-space research is one paper, from one lab, that needs to be replicated elsewhere before we draw big conclusions from it. None of these bets has 'won' yet.
And the question I'll leave us with: if the frontier is no longer one thing — which of these bets are we building what we create this year on?