SpaceXAI released Grok 4.6 on Wednesday, less than a month after Grok 4.5. The company says Grok 4.6 can research unfamiliar topics, work its way through large codebases and turn a product idea into a functioning app. SpaceXAI also found that it checked its work more often on longer tasks, although those results haven’t been independently verified.But Grok 4.6’s training reflects a shift in the coding-model race. Producing the right code on the first try still matters, but so does catching mistakes, fixing them and continuing without losing sight of the original task.Producing the right code on the first try still matters, but so does catching mistakes, fixing them and continuing without losing sight of the original task.How agents learn failureSpaceXAI subjected Grok 4.6 to a longer supplemental training run than its predecessor, combining model-generated reasoning and technical material with engineering data. The optimizer and training recipe used to update the model’s weights were also changed. It then used Grok 4.5 to regenerate supervised fine-tuning trajectories spanning different reasoning settings, agent harnesses and domains including STEM, software engineering and knowledge work. Problematic trajectories were removed using model-based checks.Reinforcement learning extended that work into general coding, kernel optimization, web development and computer-aided design. The model was rewarded for completing the larger task rather than simply drafting a plausible block of code.SpaceXAI says Grok 4.6 became more likely to pause during longer tasks and check whether its work actually held up before continuing. It also produced stronger early versions of visual and interactive applications, catching and correcting its own mistakes while it was still working.Benchmarks tell a mixed storyThe benchmarks support a considerable improvement over Grok 4.5 more clearly than they support claims that Grok 4.6 is now the best coding model.Grok 4.6 scored 69.9% on CursorBench v3.2, up from 66.7% for Grok 4.5. That put it ahead of OpenAI’s GPT-5.6 Sol Max at 67.2%, although Anthropic’s Fable 5 Max led with 70.5%. Its DeepSWE v1.1 score rose sharply from 54% to 65.9%, but GPT-5.6 Sol Max and Fable 5 Max still led with 73% and 70%. Grok fared better on FrontierCode v1.1 Extended, where its 61.3% score narrowly beat Sol’s 60.6% but remained below Fable’s 63.6%.Terminal-Bench v3.0 was a tougher test for Grok. Its score rose from 15.7% to 26%, but Sol and Fable remained well ahead at 34.6% and 34.1%.Because model selection is closely tied to the surrounding environment, these differences are important for developers selecting a model for their agents. CursorBench evaluates performance within an environment created by Cursor, a company involved in Grok’s training. DeepSWE and Terminal-Bench assess various combinations of repository work, terminal usage, and extended problem solving.Grok achieved one of its largest improvements on APEX-Agents, increasing from 47.1% to 57.5%. This placed it just ahead of Sol Max at 56.7%, but still behind Fable 5 Max at 59.2%. On APEX-SWE, Grok’s score rose from 53.6% to 56.4%, while Fable reached 58.8%.Its Artificial Analysis Intelligence Index score rose five points to 61, tying GPT-5.6 Sol Max but trailing Claude Opus 5 and Fable 5. The biggest gains appeared on longer tasks, while Sol and Fable still performed better on several coding and terminal tests.A model with a low token price may still incur high costs if it repeatedly calls tools, rereads files, or restarts tasks.Token price versus task costGrok 4.6 retains the same headline API pricing as Grok 4.5: $2 per million input tokens and $6 per million output tokens. A faster variant costs twice as much.This pricing makes the standard model less than half the API cost of GPT-5.6 Sol Max, before considering differences in reasoning tokens, caching, retries, and tool usage. For long-running agents, token rates represent only part of the total cost.A model with a low token price may still incur high costs if it repeatedly calls tools, rereads files, or restarts tasks. Conversely, a more expensive model may be more cost-effective if it completes tasks in fewer steps. As the industry moves away from unlimited AI coding budgets, measuring the cost to complete an entire job matters more than comparing token prices alone.Artificial Analysis places Grok 4.6 on its Intelligence-versus-Cost-per-Task Pareto frontier at $0.84 per task. Several models, such as GPT-5.6 Luna, GLM-5.2, and Meta’s Muse Spark 1.2, have lower costs per task but do not match Grok 4.6’s Intelligence Index score.The same testing revealed significant differences in resource usage on AA-Briefcase, a private benchmark for long-running professional tasks. Grok 4.6 completed the evaluation in about 53 turns, consuming approximately 0.5 billion input tokens. Claude Opus 5 Max required around 103 turns and 2 billion input tokens.Where developers find GrokGrok 4.6 is available through the SpaceXAI API, OpenRouter, Vercel, and Cloudflare. It is also accessible via Grok Build and Cursor, which SpaceX recently acquired.Grok Build, SpaceXAI’s counterpart to Anthropic’s Claude Code and OpenAI’s Codex, is included with the $30-per-month SuperGrok plan. SpaceXAI is offering double the included Grok 4.6 usage in Cursor and Grok Build during the model’s first week.This release follows the July launch of Grok 4.5 and comes one day after the introduction of Grok Bot, a system for assigning ongoing tasks to persistent agents.Collectively, these releases indicate that SpaceXAI is moving beyond a chatbot model toward building the infrastructure needed to support long-running agents. Collectively, these releases indicate that SpaceXAI is moving beyond a chatbot model toward building the infrastructure needed to support long-running agents. The post SpaceXAI trained Grok 4.6 on something most AI labs throw away appeared first on The New Stack.