On July 8, 2026, xAI released Grok 4.5 (API model ID grok-4.5), generally available on the xAI API and positioned as the company's frontier model "for coding, agentic tasks, and knowledge work." Elon Musk's lab priced it aggressively and leaned hard into the agentic-coding narrative. So the question for anyone building with AI is simple: is Grok 4.5 actually a frontier model, or a fast, cheap challenger wearing a frontier badge?
We did what we always do: pulled the primary docs and the independent Artificial Analysis data, and split it into verified facts, company framing, and our analysis. The honest headline: Grok 4.5 is genuinely good, genuinely cheap, and genuinely fast, but on independent intelligence it lands #8, not #1, and its context window quietly shrank. Here are the numbers.
The verified spec sheet
The facts that are not in dispute, from xAI's docs and Artificial Analysis.
| Grok 4.5, verified specs | Detail |
|---|---|
| API model ID | grok-4.5 (note the dot, not a dash) |
| Released | July 8, 2026, generally available on the xAI API |
| Positioning | xAI's frontier model for coding, agentic tasks and knowledge work |
| Context window | 500k tokens |
| Modalities | Text and image input, text output (multimodal in, text out) |
| Reasoning | Reasoning model with configurable reasoning_effort (low / medium / high, default high) |
| Output speed | ~89.5 tokens per second (Artificial Analysis) |
One detail the launch glossed over: that 500k context is a step down. Earlier Grok builds advertised a one-million-token window, so the flagship actually shrank on context. Not fatal for most work, but worth knowing before you architect a long-document pipeline around it.
The price: this is the real story
Grok 4.5's API pricing is $2 per million input tokens and $6 per million output tokens, with cached input at just $0.50 (a 75% discount). Artificial Analysis computes a blended rate of $1.35 per million tokens. For a model pitched at the frontier, that is cheap, materially cheaper than the top Claude tiers, which is exactly where Grok 4.5 makes its case.
Output price per million tokens (lower is cheaper)
Sources: xAI docs and Anthropic pricing docs, July 2026 (verified). Grok 4.5 output is a fraction of the top Claude tiers; its blended rate is about $1.35 per million tokens.
Where it actually ranks: #8, not #1
Here is where the marketing meets the scoreboard. On the independent Artificial Analysis Intelligence Index (a composite of evals like GPQA Diamond, Humanity's Last Exam, Terminal-Bench and SciCode), Grok 4.5 scores 54 and ranks #8 out of 186 models. That is a strong score, but it sits behind the current frontier: Claude Fable 5 leads at 60, GPT-5.6 is at 59, and Claude Opus 4.8 is at 56.
Artificial Analysis Intelligence Index, July 2026 (higher is better)
Axis starts at 50 to show the spread. Source: Artificial Analysis Intelligence Index, July 2026 (independent). Grok 4.5 = 54, ranked #8 of 186. Strong, but not the frontier leader.
What this means in the frontier race
What follows is our analysis.
Read the two charts together and Grok 4.5's real position snaps into focus. It is not the smartest model on the market, that is currently Claude Fable 5 and GPT-5.6. But it is one of the best intelligence-per-dollar deals at the high end: an Index of 54 at $6 output, when the models scoring a handful of points higher cost $25 to $50. For the token-hungry reality of running agents and coding loops, where you care about cost-per-completed-task and latency as much as peak IQ, that trade can be very attractive.
The caveats are real, and a launch post will not lead with them. The context window went down to 500k. The individual benchmark numbers xAI likes to headline (ARC-AGI and the rest) were not something we could independently confirm at publication, so we are not printing them. And Grok carries a well-documented history of controversial and unpredictable outputs, which is a governance question for any brand putting it in front of customers.
The through-line, the one we keep returning to across this whole model wave, is that the model is a swappable component, not a religion. Grok 4.5 is a strong new option to slot into a multi-model stack: use it where its price-performance wins (high-volume agentic and coding work), keep Claude or GPT for the hardest reasoning, and switch between them with a config change rather than a rebuild. That is exactly how we build AI features for clients, chosen per task on capability, cost and governance, over infrastructure you control (see our work). If you want an AI stack that can adopt a release like Grok 4.5 on day one and drop it just as fast if it misbehaves, tell us about your project or get in touch, and we reply within 48 hours. For the rest of the fast-moving race, read our breakdowns of Claude Sonnet 5, GLM-5.2, the best open-weights LLM, and the Fable 5 suspension saga.
Key numbers (as of July 8, 2026)
- July 8, 2026 Grok 4.5 launches, GA on the xAI API, model ID
grok-4.5. - $2 / $6 per million input/output tokens, cached input $0.50, blended about $1.35.
- 54, ranked #8 of 186 on the independent Artificial Analysis Intelligence Index, behind Fable 5 (60), GPT-5.6 (59) and Opus 4.8 (56).
- 500k tokens context window, a step down from earlier Grok builds.
- ~89.5 tokens/second output speed, positioned for fast agentic and coding work.
