Grok 4.6 in 2 weeks and Grok 4.7 in 4 weeks

— Elon Musk (@elonmusk) July 24, 2026

Grok 4.6

Parameter count: 2 trillion (2T). Confirmed by Elon as the “2T model” that is “better than our 1.5T in every way.”

Relation to prior model. It is a direct successor to Grok 4.5, which is based on xAI’s 1.5T V9 foundation model.
Training status, Initial pre-training was expected to finish the week of ~July 20–27, 2026. Supplemental training is planned, including SpaceX’s large corpus of engineering data (excluding ITAR-restricted material) to boost engineering/reasoning performance.

Goals / expected characteristics. They are aiming at matching or exceeding Moonshot’s Kimi K3 (reported ~2.8T parameters) in capability, while keeping speed and token efficiency close to Grok 4.5. Elon described it as potentially exceeding Kimi “but with speed and token efficiency close to our 1.5T (aka Grok 4.5).”

Grok 4.5 itself is served at roughly 80 tokens per second, with strong token efficiency (roughly 2× better than comparable leading models on some tasks, and significantly fewer output tokens needed on benchmarks like SWE-Bench Pro variants).

Speculation is that Grok 4.7 would have 4 to 6 trillion parameters.

Cursor Composer 3 has not been released. It is widely expected as Cursor’s next major in-house mode

Composer Model Line

Composer 2 (March 19, 2026).
Major jump. Built on Moonshot AI’s open-source Kimi K2.5 via continued pre-training + large-scale RL in realistic Cursor environments. Frontier-level coding performance at low cost (~$0.50/M input, $2.50/M output tokens; faster variant available). Strong on internal CursorBench and public benchmarks like Terminal-Bench.

Composer 2.5 (May 18, 2026)
Current default specialist. Substantial gains over Composer 2 in sustained long-running tasks, complex instruction following, reliability, and collaboration feel. Same Kimi K2.5 base + significantly more training (scaled RL environments, new methods, ~25x more synthetic tasks). Same pricing structure. Remains available and actively supported as a distinct (smaller) weight class

Improving Cursor and SpaceXAI Grok Already

Joint model training and data sharing. Cursor trained Composer models on SpaceXAI compute. Grok 4.5 was trained alongside Cursor (explicitly noted in the July 2026 launch). This creates a coding-specific data flywheel from real enterprise workloads.

Grok 4.5 launch and deep Cursor integration (July 8–16, 2026). Priced at $2 per million input / $6 per million output tokens (fast variant higher). Available immediately in Cursor (all plans, with doubled usage initially), Grok Build, and the API. Positioned as Opus-class but faster, more token-efficient, and lower-cost for coding/agentic/knowledge work. Musk highlighted the efficiency edge.

Cursor Router (July 22, 2026). Intelligent routing that prefers efficient first-party models (Grok 4.5, Composer 2.5) for routine work and escalates only when needed. Early reports show up to 60% lower cost per commit with no performance drop on Teams/Enterprise. Modes include Auto, Intelligence, Balance, and Cost optimization.

Pricing and packaging evolution. Shifted to usage-based credit pools (tied to actual API costs).

Individual Hobby free, Pro $20, Pro+/Ultra higher.
Teams Standard $40/user (split first-party vs third-party pools with higher headroom)
Premium $120/user (5× usage).
Enterprise: custom, pooled, SCIM, audit, controls.

June 2026 updates increased limits and lowered effective costs for most teams. Grok/Composer now draw from the more generous first-party pool.