Grok 4.6 and @Bot: the agent fight moves into knowledge work, priced to win on tokens

Grok 4.6 shipped alongside @Bot, a teammate product powered by it and built by a group latent.space describes as the team that went from Cursor to SpaceX, to what the newsletter calls very positive reviews ^[1][3]. That pairing is the point: latent.space has spent the year tracking coding agents breaking containment into knowledge work, and frames the teammate and multiagent category as the next battleground, with Claude Tag arriving to mixed reviews and Block's Buzz needing a more technical user ^[4][5].

The model underneath is a confirmed 1.5T parameter system that, per xAI, builds on Grok 4.5 with a focus on long-running agents and more ambitious interactive and visual work ^[6]. The training disclosure is short enough that latent.space reproduces all of it: a longer supplemental training run than 4.5, curated model-generated data for reasoning and advanced technical concepts, high-quality engineering data, and an improved optimizer and recipe ^[7][10]. Grok 4.5 then regenerated the SFT trajectories across reasoning efforts, agent harnesses, and domains including STEM, software engineering and knowledge work, with problematic traces filtered by model-based checks ^[8]. Agentic RL followed on knowledge work, general coding, and domain-specific environments for kernel optimization, web development and computer-aided design ^[9]. Note the phrase "agent harnesses" ^[8]. If a vendor trains against harnesses, harness fit is a variable in your evaluation, not a detail.

The independent numbers are mid-frontier, not top. Artificial Analysis puts Grok 4.6 at 61 on its Intelligence Index, roughly in line with GPT-5.6 Sol Max and behind Claude Opus and Fable ^[11], with 88.4% on Terminal-Bench v2.1 and 1753 Elo on GDPval-AA v2 ^[12][13]. Code Arena's early data slots it near GPT-5.6 Sol and Claude Fable on webdev ^[19]. The claimed edge is cost: $2 and $6 per million input and output tokens, which Artificial Analysis calls materially below frontier peers, with competitive AA-Briefcase results at far lower cost ^[14][15]. latent.space calls it arguably the second best knowledge work model in the world, citing both Cognition and Elon Musk, and surely the top on efficiency ^[17]. It is available in Devin, and practitioners including Pawel Huryn immediately framed it as the default for coding and bug-finding ^[16].

Efficiency is relative to whom you compare. On the same day, DeepSeek's V4 Pro went GA at roughly $0.435 per million input and $0.87 output ^[20][29], which makes Grok 4.6 about 4.6x that on input and 6.9x on output ^[22]. Cline put V4 Pro at roughly 57x cheaper than Fable 5 ^[21]. So the honest reading is that Grok 4.6 is cheap against Claude and OpenAI, not cheap in absolute terms, and long-running agents are exactly where per-token price compounds.

Two things to instrument if you test it. xAI reports more self-testing behavior during long tasks ^[18]; measure whether that cuts retries or just adds output tokens you pay for. And treat GDPval and Briefcase as proxies, not as your workload ^[13][14]. The demo layer offers little help here: SpaceXAI's Grokathon crowned Nova, a 12-hour binary-to-C decompiler prototype, but published no technical documentation or independent accuracy measurements with the results, and closed the thread pointing at its careers page ^[24][25][27].

Watch Grok 4.7, which Musk says has finished initial training with supplemental training on SpaceX internal data planned ^[23]. That is a bet that the next increment of knowledge-work capability comes from proprietary corpora, and it is the kind of thing that is very hard to reproduce or audit from outside.