Grok 4.5 Review — Benchmarks, Pricing, and the Agentic Angle
On July 8, 2026, xAI released Grok 4.5, and the launch had a distinct emphasis: rather than lead with academic leaderboards, xAI put the model’s agentic tool use and long-session coding front and center. On the independent Artificial Analysis Intelligence Index it scores 54, ranking fourth overall — behind Claude Fable 5, Claude Opus 4.8, and GPT-5.5 — but it tops the field on agentic tool use, the axis xAI clearly optimized for. This review covers what Grok 4.5 is, which of its numbers are measured versus estimated, and where it actually fits.
TL;DR verdict
| Grok 4.5 | |
|---|---|
| Type | Frontier-adjacent general LLM (agentic + coding focus) |
| Context window | 500,000 tokens · up to ~32,000 output |
| Modalities | Text, vision |
| License | Closed (xAI API) |
| Pricing | $2 in / $6 out per 1M · $0.50 cached input (≤200K prompts) |
| Headline numbers | SWE-bench Pro 64.7 · Terminal-Bench 2.1 83.3 · AA Intelligence Index 54 (#4) |
| Availability | xAI API, OpenRouter, Grok apps |
| Best for | Tool-heavy agents, long-session coding, intelligence-per-dollar |
| Caveat | Published agentic/coding numbers only — broad-reasoning cells are estimates |
If you skip the rest: Grok 4.5 is one of the strongest agentic models you can run per dollar this month. It published hard agentic and coding numbers — SWE-bench Pro and Terminal-Bench 2.1 rather than the friendlier Verified or Lite tiers — priced at $2/$6 with a 500K context. The honest asterisk is that xAI led with tool-use and coding benchmarks, not the full standard suite, so its general-knowledge ranking is still interpolation.
What it is
Grok 4.5 is a multimodal (text + vision) model with a 500,000-token context window and up to about 32,000 tokens of output — double the usable context of the Grok 4.3 predecessor and enough to hold a substantial codebase slice plus the running history of a long agentic session. The framing xAI leaned on at launch is that Grok 4.5 was tuned on real developer session data and evaluated on benchmarks designed to measure what a model can actually do inside a real codebase over a long session, not what it can recall on a one-shot exam.
That positioning matters because it explains the benchmark choices. xAI did not try to top MMLU-Pro or GPQA leaderboards; it optimized for the axis where an agent’s value shows up — planning, iterating, and using tools to finish multi-step command-line and coding workflows — and then published the benchmarks that measure exactly that.
What the launch numbers say
Here is where Grok 4.5 makes its case. The published, third-party-visible numbers are agentic and coding evaluations, and they are strong:
| Benchmark | Grok 4.5 | What it measures |
|---|---|---|
| AA Intelligence Index | 54 (#4 overall) | Blended cross-domain intelligence (Artificial Analysis) |
| Terminal-Bench 2.1 | 83.3 | Planning + tool use to finish real terminal workflows |
| SWE-bench Pro | 64.7 | Hard, contamination-resistant software-engineering tasks |
Two things make this card more credible than a typical launch deck. First, the benchmarks are the hard, agentic ones — SWE-bench Pro, not the easier Verified or Lite, and Terminal-Bench 2.1, which drops a model into a real terminal and scores whether it can actually finish the job. Second, the Intelligence Index is an independent aggregate, not a vendor-selected slice, and it places Grok 4.5 fourth overall while calling out its first-place finish on agentic tool use. That is a coherent story: a model that trades a little raw exam intelligence for a lot of get-things-done-with-tools capability.
What is not in the card: MMLU-Pro, GPQA Diamond, SWE-bench Verified, HumanEval, MATH-500, MMMU, Aider Polyglot, tau-bench — the full standard public suite. That is why, in our models leaderboard, all eight of Grok 4.5’s benchmark cells are conservative estimates anchored to Grok 4.3 (MMLU-Pro 87.0, GPQA 90.0, HumanEval 92.0, SWE-bench 73.0, MATH-500 97.0, MMMU 78.0, Aider 78.0, tau-bench 74.0), nudged up on the coding and agentic axes consistent with the confirmed tool-use gains and left flat where there is no signal of a jump. Read those cells as directional placement, not measured fact, until third-party runs of the standard suite land.
What it costs
On xAI’s API, Grok 4.5 uses a two-tier prompt-length pricing structure:
- Prompts ≤ 200K tokens: ~$2.00 input · ~$0.50 cached input · ~$6.00 output per 1M
- Prompts > 200K tokens: ~$4.00 input · ~$1.00 cached input · ~$12.00 output per 1M
Our leaderboard records the standard ≤200K tier ($2/$6) as the durable rate, since that is the band most workloads live in; teams routinely feeding 200K-plus prompts should budget the higher tier. Either way, the pricing is the core of the pitch. At $2/$6 the model sits well below the closed frontier leaders — Claude Opus 4.8 at $5/$25, GPT-5.5 at $5/$30 — while topping them on agentic tool use. That is the intelligence-per-dollar argument xAI is making, and on the published agentic benchmarks it holds up. The calculator on the leaderboard lets you plug your own token mix against the closed leaders to see where the crossover lands.
How it compares
Grok 4.5’s real competition is the value-priced agentic tier, not the absolute reasoning frontier. Against its own Grok 4.3 predecessor it is a clear step up on tool use and long-session coding, with double the context (500K vs 256K on Grok 4) and lower pricing than the older Grok 4 line’s $3/$15. Against the freshly generally-available GPT-5.6 family — which shipped Sol, Terra, and Luna on July 9 with their own strong Terminal-Bench and SWE-bench Pro numbers — Grok 4.5’s Terra-tier rival lands at $2.50/$15, so Grok undercuts it on output pricing while GPT-5.6 posts slightly higher published agentic scores. Against the open coding tier — GLM-5.2, Kimi K2.7-Code, DeepSeek V4 Pro — Grok 4.5 is closed rather than self-hostable, but it competes directly on the “frontier-adjacent, priced to undercut” axis and brings vision that the text-only open coders lack.
The caution worth stating plainly: independent evaluator commentary this month flagged elevated benchmark-gaming behavior across the newest frontier releases, so treat headline agentic scores — Grok 4.5’s included — as a reason to run your own evaluation, not a substitute for one. A model tuned on developer-session data and scored on agentic benchmarks is exactly the case where your own repository is the only benchmark that counts.
Who should care
- Teams building tool-heavy agents: This is the headline use case. Grok 4.5’s first-place agentic-tool-use finish and strong Terminal-Bench 2.1 score, at $2/$6, make it a serious candidate for function-calling and command-line agents. See multi-agent pipelines for where a strong, cheap tool-user slots into a larger workflow.
- Cost-sensitive coding workloads: If you are paying top-tier closed-model rates on a high-volume coding pipeline, Grok 4.5’s pricing plus its SWE-bench Pro 64.7 is worth a pilot — A/B it against your current coder on real pull requests.
- Long-context agents: The 500K window (with the higher pricing tier above 200K) suits agents that carry large repository slices or long session histories. Budget for the two-tier pricing if your prompts routinely exceed 200K tokens.
- Anyone ranking on broad reasoning: Wait for third-party standard-suite numbers before you place Grok 4.5 above the frontier leaders on general knowledge. Its published strengths are agentic; its reasoning placement is still an estimate. The LLM Benchmark Comparison 2026 covers how to weigh that trade.
FAQ
What is Grok 4.5? xAI’s frontier-adjacent flagship, released July 8, 2026. A multimodal (text + vision) model with a 500,000-token context window, up to ~32,000 tokens of output, tuned for agentic tool use and long-session coding.
How much does it cost? About $2 per 1M input tokens, $0.50 cached input, and $6 per 1M output for prompts up to 200K tokens; prompts above 200K are billed at $4 / $1 / $12 per 1M.
Is it better than Grok 4.3? On the agentic benchmarks xAI published, yes — a clear step up on tool use and long-session coding, with double the context. On broad academic reasoning the two are close, and xAI did not publish standard-suite numbers for 4.5, so those comparisons rest on estimates.
Does it have published benchmark scores? Only agentic and coding ones: SWE-bench Pro 64.7, Terminal-Bench 2.1 83.3, and an Artificial Analysis Intelligence Index of 54 (#4 overall, #1 on agentic tool use). The standard public suite was not published, so those cells are estimates anchored to Grok 4.3.
Is Grok 4.5 open weights? No. It is a closed model available through the xAI API, OpenRouter, and the Grok apps — not downloadable for self-hosting.
Continue reading
- AI Models Leaderboard — Grok 4.5 versus 65+ models on benchmarks, pricing, and context window, with a cost calculator.
- GPT-5.6 Review (Sol, Terra & Luna) — OpenAI’s three-tier family that went generally available the day after Grok 4.5, and the pricing Grok undercuts.
- GLM-5.2 Review — the open-weights coding flagship competing in the same value-priced tier.
- Claude Opus 4.8 Review — one of the frontier leaders Grok 4.5’s intelligence-per-dollar pitch is measured against.
- LLM Benchmark Comparison 2026 — how to read agentic benchmarks and self-reported numbers without getting fooled.
- All Reviews — index of every head-to-head review on the site.