The increment lands where 4.5 was weakest, the price stays at half the frontier, and the most revealing thing in the announcement is a comparison column xAI didn't have to include
Tips, corrections, or questions? support@omniscient.media
Get this every weekday.
The Omniscient Bulletin: consequential AI, explained and evaluated. 5 to 7 items a day with the take, not the recap.
xAI's own benchmark table for Grok 4.6 bolds a rival model on five of its ten headline rows. Anthropic's Claude Fable 5 Max, not Grok, wins the composite intelligence index and three of the four broad agentic-coding benchmarks - the kind of comparison a vendor doesn't have to print, and xAI printed it anyway.
On August 12, xAI - now shipping under the SpaceXAI brand[1] - released Grok 4.6[2], five weeks after Grok 4.5 launched on July 8[3]. The framing is deliberately modest: an increment, not a leap, with the same $2 per million input tokens and $6 per million output pricing as its predecessor, built by extending Grok 4.5's training - a longer supplemental run and reworked SFT and RL stages - rather than pretraining a new base model.
The stated focus is endurance: sustaining work across many steps, whether that's researching a topic, moving through an unfamiliar codebase, or taking a product idea to a finished artifact. xAI reports that as tasks stretch longer, the model increasingly pauses to check its own output before pressing ahead, an emergent habit the company frames as a byproduct of training rather than a built-in feature.[2]
The benchmarks mostly support that framing, and vendor benchmark tables usually have a shape: your model in bold, competitors selected and arranged so the bold column wins. Grok 4.6's table doesn't do that. xAI publishes a four-column comparison - Grok 4.6, Grok 4.5, OpenAI's GPT-5.6 Sol Max, and Anthropic's Claude Fable 5 Max - and bolds the best score per row wherever it falls.[2] The bold marks end up split three ways.
Benchmark | Grok 4.6 High | Grok 4.5 High | GPT-5.6 Sol Max | Fable 5 Max |
|---|---|---|---|---|
AA Intelligence Index | 61 | 56 | 61 | 62 |
GDPVal-AA v2 | 1753 | 1526 | 1728 | 1741 |
CursorBench v3.2 | 69.9% | 66.7% | 67.2% | 70.5% |
DeepSWE v1.1 | 65.9% | 54% | 73% | 70% |
FrontierCode v1.1 (Extended) | 61.3% | 56.6% | 60.6% | 63.6% |
APEX-Agents | 57.5% | 47.1% | 56.7% | 59.2% |
Terminal-Bench v3.0 | 26% | 15.7% | 34.6% | 34.1% |
APEX-SWE | 56.4% | 53.6% | - | 58.8% |
AA-Briefcase | 1577 | 1313 | 1502 | 1574 |
Harvey LAB (Vals) | 15.8% | 12.9% | 2.5% | 11.3% |
Count the wins and the market segmentation falls out of the table. Fable 5 Max leads five rows: the AA Intelligence Index composite and the broad agentic coding middle - CursorBench, FrontierCode, APEX-Agents, APEX-SWE. Sol takes the two hardest agentic coding lines, DeepSWE and Terminal-Bench, by wide margins. Grok 4.6's three wins are all knowledge work: GDPVal-AA v2 at 1753, AA-Briefcase at 1577, and Harvey LAB at 15.8%. Two of those are photo-finishes over Fable - 1753 to 1741, and 1577 to 1574, margins comfortably inside any honest error bar on an Elo-style rating. Harvey LAB, the legal-work eval, is the strange row: Grok's 15.8% against Fable's 11.3% is a real lead, while Sol's 2.5% is low enough to suggest the benchmark measures something Sol simply wasn't tuned for.
The methodological fine print matters, and xAI notes that competitor figures come from the respective developers' published system cards or benchmark leaderboards, taking the best of self-reported or publicly available results.[2] That's a self-assembled comparison, not an independently run one - different harnesses, different scaffolding, different dates. But publishing a mixed scoreboard at all is a departure from the genre, and it buys credibility for the rows where Grok does lead. Whether that's editorial honesty or just the recognition that independent aggregators like Artificial Analysis would surface the gaps anyway, the table ends up reading as a statement of where xAI thinks it sits in the market.
Generation-over-generation, the gains concentrate exactly where Grok 4.5 was weakest. Terminal-Bench v3.0 moves from 15.7% to 26%. DeepSWE v1.1 goes from 54% to 65.9%. APEX-Agents climbs from 47.1% to 57.5%, AA-Briefcase from 1313 to 1577, and GDPVal-AA v2 from 1526 to 1753.[2] Where 4.5 was already respectable, the moves are modest - CursorBench gains three points, APEX-SWE fewer than three.
That pattern - biggest jumps on the worst lines - is the signature of targeted post-training remediation rather than a general capability lift. It's also the correct strategy for a five-week release cycle: identify the benchmarks where the model embarrasses itself, build environments that exercise those failures, and grind. The result is a model whose floor rose much faster than its ceiling. Grok 4.6 is still behind the frontier on the hardest agentic coding evaluations, but it's no longer far behind, and on knowledge work it has rows it can claim outright.
The qualitative claims track the same theme. xAI says a first attempt at visual and interactive work now typically lands further along than it did with 4.5: handed a concrete product brief, the model sketches out both the structure and the look of an app in a single pass, then tightens it up over subsequent rounds of feedback.[2] That's the "product idea to finished artifact" pitch made concrete, and it's aimed squarely at the vibe-coding workflow that Grok Build and Cursor monetize.
The training account is unusually specific about lineage. Instead of pretraining a new base model, xAI put Grok 4.5 through an extended second training pass, feeding it synthetic examples skewed toward reasoning and hard technical problems, on top of a retuned optimizer setup. The more consequential detail is what came next: Grok 4.5 was pressed into service generating its own successor's training data, redoing the supervised fine-tuning trajectories across different reasoning efforts and agent scaffolds, spanning STEM, coding, and knowledge work, with the model itself, not a human reviewer, screening out the bad traces. The reinforcement-learning side drew on a spread of agentic environments, including chip-level optimization, web app building, and engineering design tasks.[2]
This is a self-distillation flywheel: each generation rewrites the curriculum for the next. The appeal is obvious - trajectory data from a strong model is cheaper and more consistent than human-authored demonstrations, and regeneration lets you refresh the entire SFT corpus every cycle instead of accreting stale data. The risk is equally obvious: when the model generating the data, the model filtering the data, and the model being trained are all siblings, errors that pass the family's shared blind spots compound rather than wash out. xAI's answer is the model-based filtering, which is itself subject to the same critique. The claimed emergence of self-verification on long trajectories is the encouraging counter-evidence, though xAI reports it as an observation and publishes no eval that quantifies it.
Grok 4.6 ships at $2/$6 per million tokens - unchanged from 4.5, roughly half what rival frontier models charge, with a fast variant at double the rate.[2] Distribution is the other half of the launch: day one inside Cursor[4], plus Grok Build, the SpaceXAI API, OpenRouter, Vercel, and Cloudflare, with 2x included usage in both Cursor and Grok Build for the first week.[2] That "half the frontier" framing is a snapshot, not a fixed advantage: rival pricing has kept moving this same week, and holding $2/$6 steady only counts as a competitive edge for as long as competitors don't reprice under it.
Read the scoreboard and the price list together and the strategy is coherent. xAI isn't claiming the best model; it's claiming a model within one to three points of the best on most lines, at half the cost, surfaced inside the editor where agentic coding tokens actually get burned. For a team running long-horizon agents - where output tokens compound across thousands of trajectories - a model that lands within roughly three points of the leader on eight of ten benchmark rows, and seven to nine points back on the two hardest agentic-coding lines, at half the price, is a stronger pitch than a benchmark crown, and it's a pitch the mixed scoreboard makes more credible. (One clarification worth making, since the figure often circulates in headlines as a bare "1753 ELO" with the benchmark name dropped: that's the GDPVal-AA v2 score specifically, an Elo-style rating for knowledge-work deliverables that Artificial Analysis itself uses[5], and the table's narrowest outright win.)
On safety, xAI says safeguards were recalibrated to the model's capabilities and claims its "widest-ever" run of pre-deployment tests for capability and safeguard calibration, backed by post-deployment monitoring and third-party review. The announcement is specific about where it wants the model to stay helpful - vulnerability patching, accelerating the engineering design cycle, augmenting AI research, all classic dual-use territory - but no external evaluator is named, and no system card figures appear.[2]
That's thinner than the current disclosure norm. When OpenAI shipped GPT-5.4 in March, the system card and a chain-of-thought controllability evaluation were part of the launch package. For a model explicitly positioned for long-running autonomous work, the operating mode where the attack surface is largest, a superlative in a blog post isn't the same disclosure as a published document. If the testing suite is as extensive as claimed, publishing it would cost xAI little and would pressure the rest of the field. Until then, readers are being asked to extend the credit the benchmark table earned to a testing process nobody outside xAI has seen.
Grok 4.6 is what a maturing second-place strategy looks like: a five-week increment that repairs the weakest lines, a training pipeline that turns the previous model into a data factory, honest-enough benchmarks, and a price designed to make "almost as good" the rational choice for high-volume agentic workloads. None of that produces a headline, but it's the kind of release that produces revenue.
The two things to watch are the flywheel and the paperwork. If self-regenerated training data keeps compounding cleanly, xAI's release cadence becomes a structural advantage; if it starts amplifying the family's blind spots, the benchmark floor stops rising. And if the "widest-ever" safety suite ever becomes a published system card, Grok's increments will be easier to trust at the horizon lengths xAI is now explicitly selling.
