Grok 4.6 lands with benchmarks that put xAI back at the frontier table — the clearest sign yet that the model race has gone from a two-horse sprint to a crowded field of frontier labs, Chinese open-weight challengers, and dirt-cheap near-frontier options. Plus: Cognition eyes $40 billion, the neo-clouds can't build fast enough, and Washington rethinks open models.
The frontier is crowded again — and your AI options are expanding fast.
A year ago 'frontier model' meant OpenAI, Anthropic, or Google; a couple of months ago it basically meant the first two. Now Grok 4.6 puts xAI squarely back in the race, Chinese open-weight labs are pushing efficiency and cost, and order-of-magnitude-cheaper models are good enough for more and more work. Even with OpenAI and Anthropic sitting on stronger models they haven't released, it's hard to look around the landscape and not conclude we have more choice, not less.
Bloomberg reports Cognition is in early talks to raise another billion dollars at a $40 billion valuation — up almost 50% in a quarter. The revenue backs it up: sources say the coding-agent company has doubled its run rate to $1 billion since the last round.
Cursor sought $50 billion on $2 billion in annualized revenue before SpaceX bought it for $60 billion in stock. Investors now openly speculate that a hyperscaler preempts Cognition with a $60-100 billion stock offer within a year — Cognition's Sandeep shot back: "We aren't selling."
The $13.3 billion Series C announcement describes a 'software creation platform' for running businesses, not just building apps: nearly 8 in 10 users are building a business or side project they hope to monetize, and more than a third of those are already earning revenue.
CoreWeave doubled revenue year-over-year to $2.6 billion for the quarter — with cash burn also doubling to $5.7 billion — and reported a $104 billion demand backlog, up $25 billion since June. Analysts read neo-clouds as the best indicator of marginal AI demand, and it shows no signs of slowing.
— Arkady Volozh, Nebius CEO, to investors. Nebius posted 454% revenue growth, beat earnings forecasts by 83%, and said its Blackwell compute auctions cleared 15% above the previous Hopper record. Markets rewarded the supply crunch: Nebius jumped 34%, CoreWeave 19%.
Tencent tripled AI infrastructure spend to $7.8 billion, flipped free cash flow negative, and reassured markets it could sell compute but chooses not to — uncannily echoing US hyperscaler narratives from Q1. NLW isn't sure markets have fully accounted for what a Chinese build-out does to the global investment environment.
Per Korean reports, three months of Claude Code integration compressed complex tasks like system-on-chip verification from three months to two days, and let a second-year engineer finish a month-long task in a day. A clean example of the jagged frontier: huge gains on highly customized jobs, and junior employees contributing far beyond their experience.
Wired reports the administration will expand its model testing framework to cover open models once they match frontier capabilities. The logic is pro-open-source: officials worry the framework becomes a stamp of approval, and excluding open models would create a two-tiered system that makes enterprises hesitant to use them.
Treasury Secretary Bessent publicly cheered Meta's open-weight Muse Glimmer as 'another win for American innovation,' while President Trump reportedly insists the framework stay voluntary, believing formal regulation helps China catch up. The administration's safety faction is still pushing for something more formal.
A year ago 'frontier' meant the big three US closed labs; recently it meant OpenAI and Anthropic. Grok 4.6's release — alongside credible Chinese and open-weight entrants — flipped the vibe in weeks; as Nathan Lambert put it, from 'Anthropic is so far ahead' to model competition at all-time highs in about four.
xAI claims Grok 4.6 edges past GPT-5.6 Soul and Fable 5 on GDPVal, and it jumped five points to 61 on the Artificial Analysis index — ahead of Kimi K3, tied with 5.6 Soul, a point or two behind Fable 5 and Opus 5. From OpenAI or Anthropic this would read as 'not quite state-of-the-art'; from a lab many had written off, it's a huge achievement.
Grok 4.6 keeps 4.5's pricing — $2 per million input, $6 per million output, 60% cheaper than GPT-5.6 Soul per token — and it's token-efficient too: Artificial Analysis's run cost 84 cents per task, 32% cheaper than 5.6 Soul and 73% cheaper than Fable.
Some testers are calling Grok 4.6 their new default — one blind bug-bench of 105 hidden bugs went well, and it held up on Martin Casado's technical tests. Others report incomplete work, dangerous mistakes it later tries to cover up, and oddly verbose behavior that 'values completeness above everything, including economics.'
The sober read circulating in AI circles: 4.6 shows xAI is not out of the race but not at the top — middle-top-ish against GPT-5.6 and Fable 5, both several months old and only un-updated because releases now run through government review. State-of-the-art for the public is very different from state-of-the-art inside the top labs.
— Elon Musk, on X. Musk says Grok 4.7 is significantly better than 4.6 and three to four weeks away, with initial training complete and 'a massive amount of SpaceX company data' being added in supplemental training. Notably, the community's reflexive skepticism of Musk timelines is softening — 'I'm taking this seriously now' captured the mood.
After the departures of Hassabis and Jeff Dean, many are counting Google out of the frontier race — but Reuters reports Brin is back as a key day-to-day force, telling engineers it's time to play catch-up and using co-founder power to push resources toward areas like recursive self-improvement. AI training is also relocating from DeepMind London back to Mountain View, where Brin can cut through the bureaucracy.
Sentiment holds that a merely competent Gemini 3.5 Pro — already months late — would now read as failure. Per leakers, teams are shifting to the scaled-up Gemini 4 instead: risky, but it makes sense in context.
Hours after Grok 4.6 launched, leaked numbers put the updated V4 Pro within a point of Fable and GPT-5.6 Soul on Terminal Bench and ahead on CyberGym. But Artificial Analysis scored it just 53 — barely ahead of V4 Flash and behind Kimi K3 — and first impressions skew toward 'benchmark slop,' though at roughly 1/12 Fable's price it stays interesting.
The argument gaining ground: even if Kimi K3, Grok 4.6, and DeepSeek V4 Pro are more benchmark-maxed than Fable and GPT, it doesn't matter — they're an order of magnitude cheaper, and as the frontier advances, fewer people need the bleeding edge and need it less often. DeepSeek's inference reportedly uses about half the GPU time.
Ramp's AI Index shows Fable 5 at just 6% of Anthropic tokens purchased and 11.4% of dollars spent — versus GPT-5.6 Soul's 25% of OpenAI tokens. Ramp's conclusion: more performance is no longer worth the price tag, especially with open models only a few months behind.
First, selection bias: Ramp's numbers come from a spend-management product whose users are literally optimizing away from more-powerful-than-needed models. Second, and more damning: Fable 5 carries a 30-day data retention requirement for US government safety checks, and many employees simply aren't allowed to use it — a detail Ramp's own economist later amplified.
Lurking behind everything: Anthropic and OpenAI both have more advanced models essentially ready, held back by government pressure, internal concern, or simply the absence of competitive urgency. Even so, look around the model landscape right now and it's hard not to feel we have increasingly more choice rather than less.