xAI’s Grok 4.6 just climbed into the frontier tier — scoring 61 on the Artificial Analysis Intelligence Index, neck-and-neck with GPT-5.6 Sol and five points above the Grok 4.5 it replaces — and it did it in about a month. But the closer Grok gets to the frontier, the less it looks like the model people started using it for. That tension is the spine of developer-builder Theo’s testing on the t3.gg podcast, “xAI just caught up (Grok 4.6 is here),” where he runs the release through security audits, framework migrations, stacked pull requests, and a recurring game-port gauntlet. His read in a sentence: the Cursor acquisition was really an acquisition of Cursor’s post-training machinery, Grok 4.6 is the first public proof that it works — and xAI paid for the leap by giving up the two properties that made Grok 4.5 uniquely useful: near-instant responses and class-leading token efficiency.
The Cursor acquisition, translated into a model
Grok 4.6 is not a new foundation model. It is a post-training run stacked on Grok 4.5, and that is the whole story. xAI’s announcement is unusually explicit: “Grok 4.6 builds on Grok 4.5 with a particular focus on long-running agents and more ambitious interactive and visual work.” The company ran a longer supplemental training phase using curated, model-generated data, regenerated its supervised fine-tuning trajectories with Grok 4.5 itself, filtered problematic traces with model-based checks, and then trained the result on a wide range of agentic reinforcement-learning tasks — knowledge work, general coding, kernel optimization, web development and computer-aided design.
Theo’s translation of what changed: the $60 billion acquisition of Anysphere, Cursor’s parent, was not primarily about owning a code editor. It was about getting the reinforcement-learning and post-training machinery that produces exactly these jumps. Grok 4.5 was, in his telling, “the first model that’s good enough from a lab that isn’t Anthropic or OpenAI that I could see myself actually daily driving it.” His standing complaint about previous Grok models was that they “would just write the whole thing and it would work sometimes and then fail other times.” The 4.6 release notes promise more self-testing and verification on long trajectories, aimed squarely at that problem — though Theo’s caveat lands early: “Sadly, that was not really my experience with the game gen stuff I was doing.”
Access is straightforward. The model is live in Cursor and in the Grok Build CLI — which he genuinely likes (“awesome by the way”) — with 2x usage for the first week. His own T3 Code stack already ships native Grok Build bindings, a collaboration with the Grok Build team, and the model later tidied that very integration. xAI also flagged its widest-ever suite of pre-deployment safety testing, a response to the model’s expanded “potential to hack and whatnot.”
Benchmarks say frontier, with an asterisk
On paper, the grading sheet reads like xAI has arrived. The 61 on Artificial Analysis’s Intelligence Index is a five-point gain over Grok 4.5 just one month after that release, and 23 points above Grok 4.3. “The pace of releases from these guys has been nuts lately,” Theo said. On Deep Suite, a long-horizon agentic benchmark, the model jumped from 54% to 66% — “a huge bump for a post-training improvement.” On Frontier Code it went from 56.6 to 61.3, landing “smack dab in the middle between Sol and Fable.”
BenchmarkGrok 4.5Grok 4.6Position after the jumpArtificial Analysis Intelligence Index5661Neck-and-neck with GPT-5.6 Sol; behind Opus 5 and Fable 5Deep Suite (long-horizon agentic)54%66%Just behind GPT-5.6 Sol and Fable 5Cursor Bench——Ahead of GPT-5.6 Sol; behind Fable 5Frontier Code56.661.3Between GPT-5.6 Sol and Fable 5GDP Val—Best-in-class”I don’t care. It’s GDP Val.”AA Briefcase, Harvey Lab—Best-in-classAgentic evals
The asterisks matter. First, the Cursor Bench score is contaminated: xAI “accidentally trained some of Cursor Bench data into the models,” though a newer Cursor Bench 3.2 may have cleaned it up. Second, and more fundamental, aggregate indexes are a poor proxy for usefulness. His exhibit is Opus 5: it “looked and felt really good day one,” but the more he merged the code it wrote, the more cleanup it required — he now calls it “Opus sloppus.”
“I am obligated to say that this shows that these numbers aren’t the best way to measure how smart or useful a model actually is, because for my experience, Opus 5 is significantly worse to use than Fable and Soul.”
The same caution runs in reverse for Grok 4.6’s scores. “We need to see if this holds up,” he said, because if the model really is that close to the frontier “at a fifth the price, things are changing quickly.” The list price does look like a bargain: $2 per million input and $6 per million output, roughly 60% below Opus 5 and GPT-5.6 Sol, with the same 500,000-token context window.
The hidden regression: more money, more minutes
The list price tells a misleading story. “The important detail they seem to be hiding,” Theo said, is the real cost per task, which more than doubled versus Grok 4.5. Grok 4.6 burns over 30% more tokens per run, and the cached-token read price rose from 3 cents to 5 cents per million. Grok 4.5 was one of the few major models more token-efficient than GPT-5.6 Sol; Grok 4.6 no longer is.
MetricGrok 4.5Grok 4.6List price (in / out per M tokens)Same as 4.6$2 / $6Cache-read price$0.03 / M$0.05 / MTokens per runBaseline+30% or moreCost per taskRoughly half of 4.6$0.84, “more than double” 4.5Response speed”Ran almost immediately, like creepily fast”Frontier-normal: “send off the prompt and then go do something else”Pareto frontier positionThe golden green corner, cheapest and smartestCostlier side of the median
Speed went the same direction. At 84 cents per real-world task, the model sits right next to Kimi K3, slightly above Gemini’s Flash model and Terra, slightly below GPT-5.6 Sol — and a long way from Fable 5’s $3.14. But on the cost-versus-intelligence Pareto frontier, Grok 4.5 sat in the corner where it was simultaneously the cheapest and the smartest; the 4.6 post-training “yanks them out of that golden area.” Theo’s read is blunt: “They’re clearly trying to rush their way to the frontier.”
“I was so pumped that Grok 4.5 was as efficient as it was. Seeing them lose that with 4.6 is a little sad.”
Long-horizon agentic work: where it earns its keep
“For me, as the way we build with agents changes, the ability for agents to go off for a long time and solve hard problems and come back with working results is more important than ever.” That framing doubles as the test protocol. Grok 4.6 is, in Theo’s assessment, the strongest long-horizon agentic model xAI has shipped.
He pointed it at Lakebed, his still-unreleased cloud platform, with a blunt mandate: “Do a thorough security audit of this project. Are we ready to launch? What should we fix first?” The model found a previous audit in the git history, then produced a sensible findings list of its own: the broker, worker isolation and storage control plane “will get us burned”; lock down endpoints and get the domain onto the public suffix list; finish the admin trust-and-safety UI; stop putting identity tokens in URLs; refuse sandbox-off flags in production. “It did a decent audit here. Didn’t get me any errors or problems when it did it. There are other deeper things it didn’t find, but that’s acceptable.”
The harder test was architectural. T3 Code currently supports Cursor through the “out-of-date” ACP adapter, and the plan is to migrate to the Cursor SDK. Asked to audit the migration path, Grok 4.6 correctly identified why ACP is the wrong host contract for a long-lived GUI host that needs a real model catalog, resume, cancel, images and usage tracking — “these are all real problems that we’ve had with the current cursor bindings.” It also flagged the constraints: the official docs license is no longer MIT, the SDK lacks GUI-level approve and deny, blocking ask-questions won’t work, and agent login cannot be reused, so T3 Code must implement its own login.
Then came a signature failure. Without being asked, the model dropped into plan mode, and the session announced: “Plan approval could not be completed because the client disconnected. Plan mode remains active.” The plan itself was never written anywhere. “That is obnoxious,” Theo said, and he had to turn on legacy plan mode. The messiness extended to Grok Build’s event handling: when he typed “HTML” to trigger a skill, the UI collapsed the output “again because their events are broken.”
When the model works, though, it works at a level of autonomy that matters. He had it turn the whole migration plan into a pull request and “babysit” it. The result was a roughly 1,000-line PR covering gaps in T3 Code’s Grok Build implementation. Automated feedback from Macroscope “tore things to shreds,” and the model replied to the comments using Theo’s own GitHub account — opening each reply with a note identifying which model was responding, as his custom skill requires. Then he gave it his favorite test: search his actual machine history for gaps between the events that occurred and the events T3 Code processes, and file a second PR stacked on top of the first. That requires holding competing context — the old PR, the new PR, the shell history, the skills — and staying on task. “This is one of the things again I thought Grok 4.5 did surprisingly well”: taking multi-step, unrelated work and doing it cohesively in one thread.
Design and games: “old era AI slop”
On creative output, the episode is a slaughter. Reviewing Grok 4.6’s design work on witchai.dev, the comparison site built by Dra that Theo considers essential for this kind of testing, he finds the noise pattern and sharp text over a blurry background incohesive, the purple palette generic, the layouts full of “too many cards… too old school like Tailwind templaty,” and the brutalism and centered-page options boring. His summary: “This feels like old era AI slop. This is GPT 5.0 is how this feels to me.” Claude’s design output is a mix of “sinful and disgusting” ideas and one decent floating-card hover concept, but the reference points are unambiguous: Fable’s versions are “much, much better, like comically”; Opus does “a much better job”; and even GPT-5.6 Sol, “as cringe as it often is,” is meaningfully better.
The game tests are more diagnostic. Fish Slop is the “insane aquarium clone” Theo built with Opus 4.5 in late 2025 and abandoned; his ritual is to point each new model at the old codebase and tell it to rebuild the game from scratch, using it as a reference for end-to-end systems understanding. Grok 4.6’s 2D pass has bad controls, wrong element sizing — fish pellets too big, fish too small — a UI highlight that stays stuck when the cursor moves off it, and a broken game pace where growth is too slow and stage transitions never arrive. “This is like Kimi K3 did a meaningfully better job with this than Grok 4.6 seems to be.”
The 3D pass is where it collapses. On first open, the game renders a black screen — the first model in his testing history to outright fail to produce a 3D environment, a bar “even GPT-4 was able to get” over. After he fed it a screenshot (pausing to praise the Grok Build CLI’s rendering of images, which is “genuinely really nice”), the fix inverted left and right controls, placed ground objects “comically” wrong, and produced fish models that are “the worst 3D pass I’ve seen on this in a long time… last year I would have expected quality like this.”
ModelFirst openCore mechanicsStandout issueGrok 4.6Black screenLeft/right inverted; wrong ground placementCrude fish models; “worst 3D pass in a long time”Muse 12FunctionalMovement and mechanics betterUp/down tilt issuesKimi K3FunctionalBest of the three; best fish modeling any model has producedSlightly broken lighting
Muse 12, which costs far less, delivered a functional port with only tilt issues. Kimi K3’s port is a “generational gap” ahead — the best 3D fish modeling any model has produced in his testing — and it is an open-weight model. “So yeah, there’s your comparison. Grok is not even unimpressive. It’s like last generation. It’s so far ahead here.”
Scorecard: xAI traded its two moats for a faster catch-up
The gap between strong agentic engineering and dated creative output is exactly what Theo’s final scorecard captures. He rates the field across five categories, and the pattern makes the strategic trade legible.
CategoryFable 5GPT-5.6 SolGrok 4.5Grok 4.6Cost”Way too expensive””Significantly better”Best; “way better”~6/10, “a regression”IntelligenceBest”Surprisingly good, not quite”Below frontier~6.5/10Speed”Very slow””Meaningfully faster””Flew,” especially fast mode~5.5/10Thoroughness”Thoughtful, not thorough””Incredibly thorough… writes too much code”Lacked itSlight bumpOrchestrationBest”Surprisingly close””Better than most”~6.5/10
Read top to bottom, every category where Grok 4.5 led — cost and speed — regressed in 4.6, while the categories that improved moved it toward the incumbents rather than past them. “The only things I would consider Grok previously to be a leader in were the speed and the cost. And we saw regressions in both of those categories with this release,” Theo said.
“The thing I liked Grok 4.5 for wasn’t that it was as good as Frontier at specific things. It’s that the speed and the cost were far enough away from Frontier that it felt uniquely useful in those ways. And I feel like we are losing that with Grok 4.6 a little bit.”
His counterweight is cadence. Elon Musk has already tweeted that Grok 4.7 is “significantly better” and should be ready in three to four weeks, with initial training complete and “a massive amount of SpaceX company data” being added in supplementary training; Musk expects it to be “something special.” Theo reads the teaser as the jump that lands xAI fully in the frontier range, and says the strategy explains the messaging: “It makes a lot of sense why Elon’s talking more about 47 than 46 right now.”
“It’s no longer that magic efficient small fast cheap model that I wanted it to be,” he said. His net position, in his own closing: “Is it my favorite model? No. Am I going to use it a whole lot after this video? Probably not. But does it have me excited for Grok 4.7? Absolutely yes.”
What to watch
The episode’s deepest claim is that intelligence is getting cheaper fast — Theo opens with the observation that “intelligence seems to be getting cheaper, my payroll isn’t” — and that the binding constraint is no longer pre-training compute but post-training capability and release cadence. The Cursor acquisition was an admission of that; Grok 4.6 is the proof of concept; Grok 4.7, with SpaceX data in the mix, is the first test of whether xAI can convert cadence into a durable frontier position. Theo expects 4.7 to fix the speed, cost and thoroughness regressions of 4.6 while holding the benchmark gains, landing the model fully inside the frontier.
Two tensions are worth tracking. First, benchmark parity and real-world usefulness keep diverging: Grok 4.6 scores like a frontier model and, in Theo’s agentic testing, behaves like one — except when it crashes into plan-mode disconnects and broken event handling, and except for design and 3D output that suggests the underlying capability distribution is still a step behind. He expects the same divergence he saw with Opus 5, where “every measure seemed solid” and daily use told a different story. Second, xAI’s strategy is internally inconsistent: 4.6 abandons the cheap-and-fast corner that made the brand distinctive, betting that 4.7 will be good enough to compete on quality rather than price. That is a bet on Grok catching up “very fast.”
For everyone paying for tokens, the question is no longer whether Grok 4.6 is the best model in August 2026. It is whether the four-week release cycle that produced it makes the best models cheaper within the year. The pressure is already materializing: Grok 4.6 launched the same day DeepSeek shipped V4 Pro at $0.87 per million output tokens, roughly a seventh of Grok’s price, and xAI paired the release with a new autonomous agent product, Grok Bot. The labs’ incentives are colliding — Anthropic is reportedly preparing a public offering, OpenAI has filed confidentially, and Musk has told staff he expects AI revenue to overtake the rest of SpaceX, including Starlink, as soon as September. Theo’s closing call is less about any single model than about the market structure that produces them: “We need more competition. We need more good models and we need more people fighting to make them cheaper.”