SpaceXAI's Grok 4.5: Cheaper Than Opus, GPT-5.6 Rival
11 Jul 2026
SpaceXAI has launched Grok 4.5, a new AI model the company is positioning as its smartest yet — built specifically for coding, agentic tasks, and knowledge work — and priced notably below rivals from Anthropic and OpenAI.
What's new
SpaceXAI published its Grok 4.5 announcement in a blog post on a Wednesday, describing the model as a "workhorse" for coding, app-building, office and clerical work, research, writing, and routine knowledge tasks. Grok 4.5 became publicly available the following day, per a statement from Elon Musk, who said the release followed "strong positive feedback" from a customer beta test program. Musk described the model as "roughly comparable to Opus 4.7, but much faster."
The model is now the default in Grok Build and is available today in Cursor (on all plans) and through the SpaceXAI console. Notably, Grok 4.5 was trained alongside Cursor, using datasets spanning coding, science, engineering, and math. Training reportedly ran across tens of thousands of NVIDIA GB300 GPUs, with reinforcement learning covering hundreds of thousands of tasks.
Pricing and performance claims
SpaceXAI is pricing Grok 4.5 at $2 per million input tokens and $6 per million output tokens — undercutting both Anthropic's Opus 4.7 ($5/$25 per million tokens) and OpenAI's Sol ($5/$30 per million tokens). It's priced closer to OpenAI's Luna ($1/$6 per million tokens). The model is served at fast-model speeds of 80 tokens per second, and SpaceXAI claims it delivers twice the token efficiency of leading competing models.
It's worth noting these performance and efficiency claims — including the "Opus-class" comparison and the 2x token efficiency figure — are self-reported by SpaceXAI, and no independent benchmark data is available to verify them.
Timing and competitive context
The launch lands just one day before OpenAI's planned release of GPT 5.6 on Thursday. Sources indicate GPT 5.6's prior release had been limited by the Trump administration over government security concerns — though the specific nature of those concerns isn't detailed in available material. That said, this suggests AI models in this performance class may continue to face regulatory scrutiny going forward.
Geographically, Grok 4.5 is not yet available in the EU through any SpaceXAI products or the API console; EU availability is expected by mid-July, though the reasons for the delay aren't specified.
Why founders should care
For early-stage teams evaluating AI infrastructure costs, Grok 4.5's pricing could plausibly reduce spend compared to Opus 4.7 or Sol — particularly for coding and agentic workloads where token volume adds up quickly. Teams already using Cursor may find the integration path faster, since Grok 4.5 was trained alongside that tool and is available on all Cursor plans starting today.
The claimed 2x token efficiency, if it holds up in real-world use, could further compound cost savings for compute-heavy workflows — though founders should treat this as an unverified claim until independent testing emerges. Speed (80 TPS) combined with lower per-token cost may also make the model appealing for latency-sensitive or high-volume applications, such as agent-based products.
European founders, however, should factor in the current unavailability of Grok 4.5 in the EU, with access not expected until mid-July — a timeline that could affect roadmap planning for teams building in that region.
Finally, with GPT 5.6 launching the day after Grok 4.5, founders comparing providers have a narrow window to evaluate both models nearly simultaneously — a useful moment to benchmark pricing, speed, and task performance before committing to a single vendor.
The bottom line
Grok 4.5 enters the market as a lower-cost, faster-served alternative to Opus 4.7 and GPT-5.6, with tight integration into developer tools like Cursor. The pricing advantage is clear and verifiable; the performance claims are not yet independently confirmed. Founders weighing a switch — or a first adoption — may want to test the model directly against their own workloads rather than relying solely on SpaceXAI's self-reported benchmarks.