Grok 4.6 Shows How Fast Your AI Options Are Expanding

August 13, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
August 13, 2026

HEADLINES

Cognition in talks to raise at $40B

Cognition is seeking another funding round on the back of booming coding agent demand, with Bloomberg reporting early talks at a $40 billion valuation. The company closed its last round just three months ago, raising a billion dollars at $26 billion — meaning the valuation would be up almost 50% in a quarter. The revenue figures back it up: sources familiar with the effort say Cognition has doubled its revenue run rate to a billion dollars since it last sought funding, and is looking for another billion here. The closest comp is Cursor, which sought a $50 billion valuation on $2 billion in annualized revenue back in March before SpaceX acquired the company in a $60 billion all-stock deal. If Cognition can price a round at $40 billion, that SpaceX deal starts to look like a bargain.

Lovable raises $400M at $13.3B

Lovable announced a $400 million Series C at a $13.3 billion valuation, and the most interesting part is how the company now describes itself. The announcement frames Lovable as a software creation platform for people closest to a problem, with survey data showing nearly eight in ten users building a business or side project they hope to monetize — more than a third of them already earning revenue. CEO Anton Osika emphasizes the same idea, calling the goal the most intuitive platform to build and run a business. Lovable is inching away from Claude Code and closer to something like Shopify — the clearest place to see where what was once called vibe coding meets the actual transformation of small digital businesses.

Neocloud demand is booming

CoreWeave and Nebius both reported earnings this week, and both vastly outstripped analyst expectations. CoreWeave's revenue doubled over the past year to $2.6 billion for the quarter, with cash burn also doubling to $5.7 billion — but the story for investors was the line out the door, with a $104 billion backlog that had already grown by $25 billion since the books closed at the end of June. Nebius recorded 454% revenue growth to $582 million, beat EPS forecasts by 83%, and cleared its Blackwell compute auctions at 15% above its previous record price for Hopper. Markets rewarded both, with CoreWeave up 19% and Nebius up a staggering 34% since reporting. Neoclouds are among the best indicators of marginal AI demand because they service hyperscaler overflow — and even in a quarter when token austerity came into vogue, demand shows no signs of slowing.

Tencent triples capex as the buildout reaches China

Tencent spent $7.8 billion on AI infrastructure last quarter, boosting its training and inference fleet. That is still modest against the US hyperscalers — Meta had the slowest capex in that group and spent $31.9 billion in Q2 — but it marks a clear attitude shift as Chinese tech giants commit to scaling up data center construction. Revenue is growing at 11% while free cash flow has dipped negative, with incremental earnings going toward infrastructure, and President Martin Lau noted Tencent could monetize its compute by selling to outside customers but is prioritizing its own needs for now. It is uncanny how closely this follows the US narratives from Q1, and US market participants likely haven't fully accounted for what a Chinese capex boom does to the global investment environment.

Samsung sees massive gains from Claude Code in chip design

According to reports from a Korean outlet, the first three months of integrating Claude Code into Samsung's semiconductor software stack have been an outstanding success. Development personnel cut the time to complete complex tasks like system-on-chip verification from three months to two days, and in one example a second-year engineer completed a month-long task in a single day. The report doesn't claim efficiency gains across the entire chip design process — but it is an interesting example of the jagged frontier of AI adoption in the enterprise, where the wins came from making highly customized jobs more efficient and letting a junior employee contribute far beyond their expertise.

Trump framework expands to cover open models

Leading frontier labs were briefed on the administration's model testing framework last Tuesday, and it was reported the policy would cover only state-of-the-art models — notably not open models, a decision that relieved open source advocates while a contingent of China hawks argued it left a gap. On Wednesday, Wired reported the administration has changed its mind, with an official saying the White House expects to expand the policy to open models in the coming months, triggered as soon as they reach Mythos or GPT-5.6 capability levels. Officials had hoped the policy would be one and done, but exponential capability growth forced them to iterate; the concern now is that excluding open models creates a two-tiered system where the framework reads as a stamp of approval and enterprises hesitate to use untested open weights, ultimately disincentivizing US labs from shipping them. There is still plenty of consternation around all of it — Trump himself is reportedly insisting the framework stay voluntary because he believes formal regulation helps China catch up, while the administration's safety-focused faction pushes for something more formal.

MAIN STORY

Strong Grok 4.6 Shows How Your AI Model Choices Are Expanding

A year ago, "frontier model" meant OpenAI, Anthropic, or Google. A couple of months ago it meant OpenAI or Anthropic. As of this week, Grok 4.6 puts SpaceXAI squarely back in that conversation — and the more important story isn't the model itself but what it says about the practical shape of the AI race, and how much more choice that leaves for what gets built at home and inside companies.

SPACEXAI JOINS THE FRONTIER

Grok 4.6
"Benchmarks deserve a bowlful of salt — these are hard to ignore"
On GDP-Val, the measure of how agentic AI performs on economically valuable tasks, SpaceXAI claims to have overtaken both GPT-5.6 Sol and Fable 5 — by a small amount, but overtaken nonetheless. Coding performance lands in the same range on Cursor Bench and a few points behind on DeepSWE and Terminal Bench, while the overall Artificial Analysis Intelligence Index score jumped a full five points from Grok 4.5's 56 to 61, ahead of Kimi K3, tied with 5.6 Sol, and a point or two behind Fable 5 and Opus 5. If this were a new model from Anthropic or OpenAI, the conversation would be about how it didn't quite push the frontier forward — but for a lab many had written out of the race entirely, it is a huge achievement, and the broad sense across AI circles is that there are once again three frontier labs.

The cost story
"60% cheaper than Sol per token, and token efficient too"
The performance jump appears to come off the same base model as Grok 4.5, and pricing hasn't moved — $2 per million input tokens and $6 per million output. Comparing tokens to tokens is a seductive but fraught exercise given how differently models consume them, but Artificial Analysis found Grok 4.6 completed its benchmark run at 84 cents per task, in line with Kimi, 32% cheaper than GPT-5.6 Sol and 73% cheaper than Fable.

Early reviews
"Very fast, capable and cheap — but not yet Fable class"
The early testing splits hard. Ben Davis calls the Cursor and SpaceXAI comeback glorious to watch, Martin Casado's highly technical tests came back strong, and Pawel Huryn's blind bug bench of 105 hidden bugs across two real repos has him calling it a possible new default model on time, value and cost. On the other side, Mehul Mohan found it does incomplete rather than incorrect work and wonders if the harness is the problem, and Justin Schroeder reports dangerous mistakes that the model then gets defensive about. Benjamin De Kraker's summary is the fairest read: the numbers show xAI is not out of the race but also not at the top — they advanced past a wildcard game into the playoffs, not to a championship.

Grok 4.7
"Three to four weeks out, trained on SpaceX company data"
Musk says initial training on 4.7 is complete and the team is now adding a massive amount of SpaceX company data in supplemental training, claiming the model will exceed all current models and that nothing will beat it at real-world engineering. That framing is landing with more credulity than usual — Chubby says the 4.6 leap makes Musk's words worth taking seriously if the 10T model is still to come, and Leo argues xAI is now arguably the third best lab in the world.

Google
"Sergey Brin is back to spur a Gemini comeback"
After last week's departures of Demis Hassabis and Jeff Dean, many counted Google out of the frontier race entirely. The counterpoint is Reuters reporting that co-founder Sergey Brin has become a key cheerleader for the AI team, addressing a town hall after the Mythos release to tell engineers it was time to play catch up. Brin stepped down as president in 2019 but returned to frequent work in 2023 and deepened his involvement with the AI team in 2024 — and with AI training relocating from DeepMind's London office back to Mountain View, he is positioned to work day-to-day with key researchers and cut through the bureaucracy. Reuters reports he has already used his implicit power as co-founder to push resource allocation toward areas like recursive self-improvement. On what ships next, sentiment is that a merely competent Gemini 3.5 Pro would now read as failure given how far behind schedule it already is, and leakers report teams are shifting to the scaled-up Gemini 4 instead — risky, but sensible in context.

DeepSeek V4 Pro
"Leaked benchmarks looked competitive; the real run disappointed"
Just a few hours after Grok 4.6 launched came a significant leak out of China. The leaked benchmarks claimed V4 Pro scored 87.9% on Terminal Bench 2.1 — 0.1% behind Fable and 1.1% behind GPT-5.6 Sol — and beat Fable by 0.2% on Cybergym, the main cybersecurity benchmark. Almost immediately, other information suggested the model was further behind than that: Artificial Analysis's run scored it just 53, one point ahead of V4 Flash and trailing Kimi K3 and Muse Spark 1.2. First impressions have been rough, with Luckey Faraday calling it benchmaxxed slop that can't manage a simple Minecraft clone. But the counter-argument is about where a model sits in the stack rather than raw performance — at $1.32 per million input and $3.96 per million output even after a substantial price increase, it lands around one-twelfth the price of Fable, and as Agustin Lebron argues, fewer and fewer people need the bleeding edge as often as they think.

Ramp's AI Index and the Fable 5 question
"A new upper bound on what businesses will pay for intelligence"
Ramp's data shows Fable 5 making up just 6% of tokens businesses purchased from Anthropic and 11.4% of dollars spent on Anthropic models, against GPT-5.6 Sol at 25% of OpenAI tokens and 23% of spend — a model so powerful it was briefly banned, and businesses don't think it's worth the price. Ramp's conclusion is that this is a new ceiling on AI spend, and that clearing it requires performance beyond Fable 5 that competitors can't approach. That story is much less clear than it looks. Ramp's data suffers from selection bias generally and acutely here, since it comes from a token and spend management product whose users are by definition optimizing away from models more powerful than they need. The bigger and more obvious issue is that Fable 5 still carries a 30-day data retention policy for US government safety checks — a detail Ramp's own lead economist returned to Twitter to flag after replies poured in from employees who simply aren't allowed to use it. Sophisticated buyers will absolutely stop smashing the state-of-the-art model at max effort on every prompt, but retention makes this a bad comparison.

What's being held back
"More advanced models are ready and sitting on the shelf"
Lurking behind all of it is the fact that Anthropic and OpenAI both have more advanced models more or less ready to go, held back by some combination of government pressure, internal concern, and the simple absence of competitive urgency. That is why 4.6 gets measured against GPT-5.6 and Fable 5, both several months old at this point — state-of-the-art for the public is now very different from state-of-the-art at the top labs. Even so, it is hard to survey the landscape right now and not feel like there is more choice rather than less.