Grok 4.5

Grok 4.5 is xAI's new flagship model, made publicly available on 8 July 2026. The launch matters less for any single headline number than for what it signals: after a period in which xAI's models sat a clear tier below the leading closed labs, Grok 4.5 puts the company back on the intelligence frontier. Coverage has leaned into the "SpaceXAI" framing, a nod to the organisational gravity behind the release, and described the model, more usefully, as a cheap, Cursor-trained coding model. That second description is the one that should shape your expectations.

On the Artificial Analysis Intelligence Index v4.1, measured at high reasoning effort, Grok 4.5 scores 54. That places it fourth overall, behind Fable 5 (60), GPT-5.6 Sol (59) and Opus 4.8 (56), and a level ahead of Claude Sonnet 5 (53). Fourth place is not a marketing line anyone prints on a homepage, but it is genuinely frontier-tier territory, and it is achieved at a price point none of the three models above it comes close to matching.

So the honest one-sentence summary is this: Grok 4.5 is not the smartest model you can rent, but it is the cheapest way to rent something this smart from a closed lab. The rest of this review unpacks where that trade-off pays off, chiefly in agentic coding and high-volume workloads, and where it does not.

The capability story centres on coding, and specifically on agentic coding: the model operating inside a harness that lets it read files, run commands, edit code and iterate, rather than answering one-shot questions. xAI ships its own harness for this, called Grok Build, and it is within Grok Build that the model's headline coding result was recorded: a Coding Agent Index score of 76.

That 76 puts Grok 4.5 level with GPT-5.5 running in Codex, which is respectable company. More striking is the result on SWE-Atlas-QnA, one of the three evaluations that make up the Coding Agent Index: there, Grok 4.5 ties GPT-5.6 Sol outright. Sol still leads the overall index at 80, so nobody should read this as a changing of the guard, but matching the current coding leader on one of its three constituent evals is a real signal, not noise.

The harness detail deserves emphasis, because it cuts both ways. Scoring a model inside its own vendor-built harness is standard practice, every lab tunes model and scaffold together, but it means the 76 reflects the Grok 4.5 plus Grok Build combination, not the raw model dropped into whatever tooling you already run. If your team lives in a different agentic stack, treat the number as an upper bound until you have run your own trials.

Start with general intelligence. The Artificial Analysis Intelligence Index v4.1 places Grok 4.5 at 54 when run at high reasoning effort. The ordering above it is Fable 5 on 60, GPT-5.6 Sol on 59 and Opus 4.8 on 56; immediately below sits Claude Sonnet 5 on 53. Fourth place, then, but the shape of the table matters. The gap from Grok 4.5 up to Opus 4.8 is two points; the gap up to the leaders is five or six. This is a model within touching distance of the front row, not one watching it from the stands.

On coding, the picture is narrower but arguably stronger. The Coding Agent Index score of 76 (again, recorded in the Grok Build harness) matches GPT-5.5 in Codex, and the tie with GPT-5.6 Sol on SWE-Atlas-QnA is the single most impressive line in Grok 4.5's benchmark record. Sol remains ahead on the index overall at 80, driven by the other two evaluations, so across a full spread of agentic coding work the leader is still the leader.

Two reading notes. First, all figures here are Artificial Analysis measurements from July 2026, and index versions matter: these are v4.1 Intelligence Index numbers and should not be compared against scores from earlier index revisions. Second, benchmark position and practical value are different questions. Fourth on intelligence sounds mid-table until you attach the cost column, which is where Grok 4.5's case actually lives, and where the next section goes.

Here is the number that defines this release: $0.31. That is what it costs, on Artificial Analysis's methodology, to run one Intelligence Index task on Grok 4.5. The equivalent figure is $1.04 for GPT-5.6 Sol, $1.80 for Opus 4.8 and $2.75 for Fable 5. Put differently, the model five or six Index points ahead of Grok 4.5 costs roughly nine times as much per unit of benchmarked work, and even the nearest premium rival costs more than three times as much. Among closed frontier-tier models in the current set, nothing else is close.

The raw API pricing tells the same story in simpler terms: $2 per million input tokens and $6 per million output tokens. Those are the rates behind our AITR Value For Money Index, our own derivation of intelligence divided by cost per task, on which Grok 4.5 scores 174, the best result of any closed model we track. On our intelligence-versus-cost scatter, it sits on the edge of the most attractive quadrant: above-median intelligence delivered at below-median cost. Very few closed models have ever occupied that corner at launch.

The practical consequence is that Grok 4.5 changes the default calculus for cost-sensitive deployments. Workloads that previously forced a choice between a frontier model you ration and a mid-tier model you run freely can now, plausibly, get frontier-adjacent intelligence at mid-tier economics. That will not matter for teams whose model spend is a rounding error. For everyone running agents at volume, it is the whole story.

The most load-bearing phrase in the launch coverage is "Cursor-trained". Grok 4.5 was reportedly shaped around the kind of work developers actually do in an AI-native editor: multi-file edits, iterative fixes, tool calls, test-and-retry loops. That lineage shows in the benchmark record, the Grok Build result and the SWE-Atlas-QnA tie both measure exactly this style of work, and it suggests the model's natural home is inside a coding agent or IDE integration rather than as a general-purpose chat assistant.

Deployment logistics look sensible. On OpenRouter, serving endpoints were observed offering a 500k-token context window, with four live endpoints as of our 10 July 2026 snapshot. Half a million tokens is enough headroom for large-repository work (long agent transcripts, sizeable codebases held in context, extended tool-use sessions) without the aggressive pruning smaller windows force on agent builders. Multiple live endpoints at launch also matters practically: it means real routing choice and less exposure to a single provider's rate limits.

Who should shortlist it? Teams running agentic coding at meaningful volume, where the per-task economics compound daily. Startups that want frontier-adjacent capability without frontier invoices. Builders of background agents (CI triage, automated refactoring, batch code review) where a nine-fold cost difference against Fable 5 decides whether the product is viable at all. Teams whose hardest problems demand the absolute best reasoning available should still look one row up the leaderboard, and should expect to pay accordingly.

The most obvious limitation is the one the leaderboard states plainly: fourth is not first. At 54 on the Intelligence Index, Grok 4.5 trails Fable 5 by six points and GPT-5.6 Sol by five. On the hardest reasoning-heavy work, the tasks where those top-of-index points actually bite, the premium models remain the better tools, and no cost argument changes that for teams whose bottleneck is capability rather than budget.

The coding score needs its asterisk kept attached. The Coding Agent Index of 76 was recorded in Grok Build, xAI's own harness, and GPT-5.6 Sol leads the index overall at 80. The SWE-Atlas-QnA tie is genuine, but it is one of three evaluations, and Sol's lead comes from the other two. Anyone planning to run Grok 4.5 outside Grok Build (in their own agent framework, or a third-party editor) should validate performance in that environment before committing, because the published number does not measure it.

Two further cautions. The model has been public only since 8 July 2026, so the track record on reliability, regression behaviour and long-tail failure modes is a week old at the time of writing; our OpenRouter observations are a single 10 July snapshot and endpoint counts and context limits can shift. And while 174 is the best Value For Money Index score of any closed model on our books, it is worth remembering the index's ceiling: open-weight models reach far higher, with MiMo topping the chart at 1,400. "Best value among closed models" is a real distinction, but it is a qualified one.

Grok 4.5 is the most interesting closed-model release of the summer, not because it wins anything outright, but because of where it lands on the map. Fourth on intelligence at 54, level with the coding leader on one of three agentic evals, and priced at $0.31 per Index task against rivals charging $1.04 to $2.75: that combination has no precedent among closed frontier models, and our Value For Money Index score of 174 reflects it.

The recommendation splits cleanly. If your workload is agentic coding, background automation, or anything you run at volume, Grok 4.5 should be on your shortlist today, and quite possibly at the top of it: $2 per million input tokens and $6 per million output, a 500k context window, and a Cursor-trained disposition toward exactly this kind of work make it the obvious value pick. If instead you need the strongest available reasoning for a small number of hard problems, Fable 5 and GPT-5.6 Sol remain the better instruments, and the price difference will not matter to you.

Our verdict: Grok 4.5 does not move the frontier of what AI can do, but it moves the frontier of what frontier-adjacent AI costs, and for most working teams, that is the more consequential frontier. It earns a strong recommendation as a value pick, with the usual caution owed to any model a week into public life.