Summary: SpaceXAI released Grok 4.6 on August 12, 2026, and Artificial Analysis’s same-day independent evaluation scored it 61, five points above Grok 4.5 and level with GPT-5.6 Sol (max). The evaluation placed it on the cost-performance Pareto frontier for every agentic test, with about 53 turns and 0.5 billion accumulated input tokens versus 103 turns and 2.0 billion for Claude Opus 5 (max). Pricing stayed at $2/$6 per million input/output tokens, while Artificial Analysis estimated $0.84 per task.
SpaceXAI released Grok 4.6 on August 12, 2026, a little more than a month after Grok 4.5 — a quick turnaround for a frontier lab. Artificial Analysis posted its independent evaluation that same day: 61 on the Intelligence Index, five points above Grok 4.5 and 23 points above Grok 4.3, level with GPT-5.6 Sol (max), behind only Claude Opus 5 (max, 63) and Claude Fable 5 (62), and slightly ahead of Kimi K3. Between the launch post and the benchmark write-up, the Hacker News threads pulled in close to a thousand points combined. On the strength of that one modest, fast-cycle update, SpaceXAI is back among the frontier group.
What the Score Says: Frontier Again
Where does 61 land? On the same tier as GPT-5.6 Sol (max), one to two points behind Claude’s two flagships — and the top of this index has not been packed this tightly at any point this year. The practical frontier on the benchmark is now a cluster of models rather than one leader. Among the sub-evaluations, GDPval-AA v2 Elo came in at 1753, second only to Claude Opus 5 and a statistical tie with Claude Fable 5 and Qwen3.8 Max. The two scorecards also disagree on method, which is worth noticing: the eval table x.ai published itself claims leads on GDPval-AA, CursorBench v3.2, AA-Briefcase, and Harvey LAB, yet the same table puts DeepSWE behind GPT-5.6 Sol Max (65.9% vs 73%) and shows the model trailing FrontierCode and APEX-SWE against Fable 5 Max. A vendor flagging its own weak spots inside a launch table stays rare among 2026 frontier releases, and it makes the third-party numbers easier to trust.
Agentic Work Is Where It Wins
Artificial Analysis states Grok 4.6’s strength without hedging: few models stay competitive across knowledge work, customer service, and terminal tasks at the same time, and this one “sits on the cost-vs-performance Pareto frontier for every agentic evaluation.” The sub-scores bear that out — 50.7% on tau3-Banking puts it in the top two alongside Qwen3.8 Max at 51.3%, and an AA-Briefcase Elo of 1577 places it in Claude Fable 5 territory. Turn efficiency is the deeper part of the story: the index tasks took roughly 53 turns and about 0.5 billion accumulated input tokens on average, against about 103 turns and 2.0 billion tokens for Claude Opus 5 (max). Turns and context accumulation, not sticker price, are what drive agent cost, so a model finishing the same job in half the turns produces bills well below what its per-token rates imply — and Grok 4.6’s flat pricing is where that compounds.
The Point of Flat Pricing
Rates hold at Grok 4.5 levels: $2 per million input tokens, $6 per million output tokens, cache hits moved from $0.30 to $0.50 per million, and a 500k-token context window. Artificial Analysis puts the measured cost per task at $0.84, matching Kimi K3 and more than 60% below Claude Opus 5 ($5/$25) or GPT-5.6 Sol ($5/$30). A frontier model delivering a full generation of intelligence gains with no price increase is an unusual event in 2026; at the frontier, more capability normally arrives with a hike. Because output token price is what dominates the cost of reasoning-heavy workloads, the $6 output rate does quiet, heavy lifting here, and it feeds straight into model-selection calls for any team running agents at volume. That Kimi K3 matches the per-task cost is notable on its own: the open-weight camp now sets the price reference frontier labs have to beat.
Availability and Ways In for Developers
Day-one availability covers Cursor and Grok Build, the API with keys from console.x.ai and documentation at docs.x.ai, and partners including OpenRouter, Vercel, and Cloudflare. On the technical side, x.ai points to a longer supplemental training run, SFT trajectories regenerated from Grok 4.5 and filtered by model, and agentic reinforcement learning spanning coding, knowledge work, kernel optimization, web development, and CAD. The model, the company says, is better at “checking its own work before moving on,” with improved one-pass results on visual and interactive projects; x.ai also ran its “widest-ever suite of pre-deployment testing” plus third-party and post-deployment evaluation. For the first week, Grok Build and Cursor carry a doubled-usage promotion, while the “fast” variant is priced at double the standard rates.