xAI's July 2026 coding and agent model is cheaper and faster than the frontier model, and a step behind it in capability. Here, you will get to know what it is, what it really costs, where the numbers disagree, and a security issue in its coding CLI.
Grok 4.5 was trained alongside Cursor, the AI coding editor, using real developer-session data. Cursor describes it as tuned on long-running agent work inside codebases, not isolated snippets. xAI says training ran across tens of thousands of NVIDIA GB300 GPUs, with reinforcement learning over hundreds of thousands of multi-step engineering tasks, aimed at getting more out of each token.
Claims that it uses a new "V9" base of roughly 1.5 trillion parameters circulate in secondary coverage, and some pages describe it as a mixture-of-experts model. xAI's own model card publishes no parameter count or architecture, so treat these as unconfirmed.
Grok 4.5 Benchmarks: Competitive, Not Dominant
xAI's prose says Grok 4.5 beats comparable leading models. Its own chart is more mixed: Claude Fable 5 tops all four coding benchmarks xAI published.
| Coding benchmark | Grok 4.5 | Fable 5 (max) | GPT-5.5 (xhigh) | Opus 4.8 (max) |
| DeepSWE 1.0 (pass@1) | 62.0% | 66.1% | 64.31% | 55.75% |
| DeepSWE 1.1 (mini-swe-agent) | 53% | 70% | 67% | 59% |
| Terminal Bench 2.1 | 83.3% | 84.3% | 83.4% | 78.9% |
| SWE-Bench Pro (resolve rate) | 64.7% | 80.4% | 58.6% | 69.2% |
Grok 4.5 beats GPT-5.5 on SWE-Bench Pro but trails Opus 4.8 and Fable 5 there, and is closest to the leader on Terminal Bench 2.1. Figures are from xAI's launch chart as reported by MarkTechPost, with rival numbers taken from system cards and leaderboards.
Independent testing agrees it sits behind the leaders. On the Artificial Analysis Intelligence Index (current version v4.3.2), the guide reports a score of 39, ranking 63rd of 225 models as of October 6, 2026, and notes the site now lists the model as deprecated. Artificial Analysis's earlier testing also flagged a higher hallucination tendency than its predecessor, even with better factual knowledge. For legal, financial, or client-facing work, that is a reason to verify answers.
One more caution: the model was co-trained with Cursor on Cursor's workflows, and Cursor left its own CursorBench out of the launch comparison after flagging a training overlap. That doesn't taint the other benchmarks, but it's a reason to run your own trials.
Where Sources Disagree
- Intelligence score. The guide cites 39 on index v4.3.2. Other tracking pages show about 54 (one lists 53.8). The guide says scores from older index versions aren't comparable, which probably explains the gap, but check the version before quoting any number.
- Speed. The guide gives about 67 tokens per second now and about 80 at launch. One aggregator lists 88.
- Cached input price. The guide cites $0.30 per million. One aggregator says sources quote between $0.30 and $0.50, and advises checking the console.
- Grok 4.6 date. Pre-release pages said August 7. The guide and later coverage say August 12.
Price Comparison
The 2/6 rate only holds for prompts under 200,000 tokens. At 200,000 tokens and above, the whole request bills at a higher tier: $4 input, $0.60 cached input and $12 output. This is a cliff, not a marginal tier: a 201,000-token request costs roughly double a 199,000-token one. So the advertised rate covers only the bottom 40% of the 500K window. That window is also smaller than the multi-million-token windows some earlier Grok variants advertised.
| Model | Input / 1M | Output / 1M | Context |
| Grok 4.5 | 2(4 at 200K+) | 6(12 at 200K+) | 500K |
| Claude Opus 5.5 | $4 | $20 | 1M |
| GPT-6.1 Sol | 2(4 above 272K) | 10(15 above 272K) | 1.05M |
| GPT-6 Luna | 0.10(0.20 above 272K) | 0.50(0.75 above 272K) | 1.05M |
Per the guide, Grok 4.5 is half Opus 5.5's input price and about 70% cheaper on output, and matches GPT-6.1 Sol's input with 40% less on output. GPT-6 Luna undercuts it heavily at a twentieth of the input price.
Token rates aren't the whole bill:
- Built-in tools (web search, X search, code execution) cost $5 per 1,000 calls, and the agent decides how many calls to make.
- Hidden reasoning tokens bill at the $6 output rate.
- One aggregator notes Grok 4.5 is excluded from the batch discount that applies to Grok 4.3, and lists file attachments at $10 per 1,000.
OpenRouter lists the same 2/6 and 500K context if you'd rather not sign up with xAI directly.
Speed and Token Efficiency
Efficiency is the strongest part of the case. xAI claims about twice the token efficiency of comparable leading models. On SWE-Bench Pro, Grok 4.5 averaged about 15,954 output tokens per resolved task against roughly 67,020 for Opus 4.8, about 4.2x fewer (xAI's figure).
Output tokens are the costly half of the bill, so the savings stack on the lower rate. The guide concludes that the effective cost per coding task is close to an order of magnitude lower than Opus 4.8, provided prompts stay under the 200K threshold. It is not the smartest model, but it may be the cheapest route to frontier-adjacent work.
How to Access It
- Grok Build, xAI's terminal coding agent (Grok 4.7 is now offered there too).
- Cursor, on all plans.
- The SpaceXAI console and API.
- SuperGrok and X Premium for chat use.
- Gateways such as OpenRouter, Vercel, and Cloudflare.
Cursor's free promotional window ended in July, and xAI has run free trial usage in Grok Build since, so check current offers. Before connecting client or regulated data, read the plan terms: on free and consumer tiers, conversations may be used to improve the models, and only Business and Enterprise plans guarantee exclusion from training.
The Grok Build repo-upload Problem
This concerns the coding CLI, not the model or the API. In mid-July 2026, a researcher publishing as cereblab captured traffic from Grok Build 0.2.93 and showed it uploading the developer's entire tracked repository, with full Git history, to an xAI-controlled storage bucket.
- On a 12 GB repo, about 192 KB went to the model endpoint while 5.10 GiB went to storage.
- A canary file the model was told never to open was recovered verbatim from the uploaded bundle, so the upload didn't depend on what the model read.
- Credentials in a tracked .env file were stored unredacted.
- With "improve the model" switched off, uploads continued.
xAI shipped a server-side fix on July 13, 2026, and uploads stopped in repeat testing. Musk confirmed the uploads happened and said prior user data would be deleted, and xAI documented a zero-data-retention policy and added a privacy endpoint.
Two caveats remain: the upload code reportedly stays in the binary behind a server-side flag that could be flipped back, and the response came through social posts rather than a formal security advisory, with no independent audit and no disclosed count of affected accounts.
Before pointing any coding agent at a sensitive repository, check what it uploads.
What it Means for AI Search and Your Brand
This section reflects the guide author's business. Grok answers questions using live X posts and the open web, so it forms a view of a company and repeats it. A cheaper, faster model gets embedded in more products, which means more AI answers more often. The guide's advice is the usual one: publish pages that state key facts cleanly enough to be lifted into an answer, earn third-party mentions that models treat as corroboration, and measure whether engines mention you at all.
Frequently Asked Questions
Is Grok 4.5 Available in Europe?
Yes. It was excluded across the EU at the July 8 launch while xAI completed EU AI Act evaluations, and xAI confirmed full European availability on July 16, 2026.
How Much Does Grok 4.5 Cost?
$2 per million input tokens and $6 per million output tokens, with prompts of 200,000 tokens or more billed at double across the whole request. Tools bill separately and hidden reasoning tokens bill at the output rate.
Is it Really "Opus-Class"?
That's Musk's phrase. Independent testing places it behind frontier leaders, and his own precise comparison was to Opus 4.7, a generation behind the Opus 4.8 current at launch.
Is it Better Than Grok 4.3?
For coding and agentic work, yes. It lists at a higher per-token rate, has a smaller context window than some earlier Grok variants, and showed a higher hallucination tendency in early testing, so test before switching if you rely on long context or factual reliability. It has since been followed by Grok 4.6 and 4.7.
Does it Hallucinate a Lot?
Early Artificial Analysis testing flagged a higher tendency than its predecessor. Enable web search to ground answers, and verify anything high-stakes.
Source: "Grok 4.5: Specs, Benchmarks, Pricing & the Honest Verdict (2026)," GEO Toolbox, published July 11, 2026, updated October 6, 2026. Cross-checked loosely against pricing and model-tracking sites (pricepertoken, LLM Stats, Swfte, LM Market Cap). Benchmarks are largely xAI's own; figures may have changed. Reworded and reorganized.
Related Searches: