Grok 4.5, reviewed: the coding specialist that got smaller, not bigger

Model updates are supposed to move one direction: more context, more capability, a bigger number somewhere. Grok 4.5 breaks that pattern in one specific way. xAI's own model card lists a 500,000-token context window — half of Grok 4.3's one million — while the price per token went up. That's not a typo, and it's not a downgrade either. It's a specialization, and whether it's the right trade for you depends on what you're actually building.

xAI released Grok 4.5 on July 8, 2026, after running it in private beta at SpaceX and Tesla through late June. The company built it together with Cursor, the AI-native code editor, and shipped it simultaneously inside Grok Build (xAI's own coding agent), inside Cursor on all plans, and through the standard API. That distribution pattern tells you what this model is for before you read a single benchmark: it's a coding and agentic-work model first, a general chatbot second.

A sourcing note before anything else: xAI has not published an official benchmark table for Grok 4.5, the same gap the Grok 4.3 review flagged. The pricing, context window, and modality figures above come straight from xAI's own developer documentation. Every capability score is sourced to the independent Artificial Analysis leaderboard, which is not xAI's own claim and should be read as outside estimate, not a vendor number.

The context window actually shrank — and a lot of write-ups got it wrong

This is worth stating plainly because it's easy to get backwards. Grok 4.3 shipped with a 1,000,000-token context window. Grok 4.5's official model card, at docs.x.ai/developers/models/grok-4.5, lists 500,000 tokens — not a million. Several third-party roundups published after launch describe Grok 4.5 as having a bigger, million-token window, which appears to be a mix-up with Grok 4.3's spec rather than anything xAI actually documents for 4.5. If a context-window number for this model doesn't match the official card, don't build around it.

The price moved in the same "smaller footprint, sharper focus" direction. Grok 4.3 runs $1.25 input / $2.50 output per million tokens flat, no matter how long the context gets. Grok 4.5 runs $2/$6 below a 200,000-token threshold and $4/$12 above it, with cached input at $0.30 and $0.60 respectively. That's a real price increase over the sibling model it's positioned to replace as xAI's flagship default, and it's tiered in a way 4.3 isn't.

Built with Cursor, for coding and agent loops

The detail that actually explains the trade-off is who xAI built this model with. Grok 4.5 was developed together with Cursor and made available inside Cursor on all plans starting day one, alongside Grok Build and the standard API. That's a different launch pattern than a general chatbot release, and it points at what the training and tuning work actually optimized for: multi-step coding tasks, tool use, and agent loops that run inside a code editor rather than a chat window.

xAI frames Grok 4.5 as built for "coding, agentic tasks, and knowledge work" on its own release notes, and the immediate, simultaneous availability in a third-party coding tool most model launches don't get is the clearest signal that coding is the actual target, not a side benefit. Whether that focus produced a better coding model than the broader Grok 4.3 is the real open question, and it's one xAI's own documentation doesn't answer with a published benchmark table.

Read the table as a trade, not a straight upgrade. If your workload is coding and agentic tool use inside Cursor or Grok Build, Grok 4.5's focus and the Cursor integration are the reason to pick it over its cheaper, broader sibling. If your workload needs the largest context window in xAI's lineup, or you want the lower flat rate, Grok 4.3 still does that job and remains available.

What the numbers don't tell you

xAI has not published an official benchmark suite for Grok 4.5, so there's no vendor-reported SWE-bench or agentic score to cite here with a straight face. Independent estimates from Artificial Analysis put its Intelligence Index at 54, similar territory to Grok 4.3's 53 and not a leap ahead of it on that one general measure. Coding-specific claims about token efficiency relative to other frontier models are circulating in secondary coverage of the launch, but the specific multipliers reported disagree between sources, so benchr isn't repeating a precise figure here — treat any single stated ratio as unconfirmed until xAI publishes its own numbers.

The Batch API offers no discount for Grok 4.5, unlike some other xAI tiers, so there's no cheap-and-slow lane to fall back to for bulk work. Reasoning effort on other current Grok 4.x models is configurable (low, medium, high); xAI's model card for 4.5 confirms reasoning support but doesn't spell out the parameter the way the release notes do for 4.3, so don't assume identical controls until you've checked your own API response.

The verdict

Grok 4.5 is a legitimate, narrower model, not a strict upgrade. The Cursor co-development and same-day availability inside Cursor and Grok Build are real, and they matter if coding and agent loops are your actual workload — that's a distribution advantage no benchmark table captures. But the context window is smaller than its sibling's, the price is higher, and there's still no official benchmark table to weigh against the higher cost.

Go with Grok 4.5 if you're building inside Cursor or Grok Build and want a model xAI tuned specifically for that loop. Stick with Grok 4.3 if you need the larger context window or the lower flat rate and don't need the Cursor integration. Either way, treat every context-window claim about this model that isn't docs.x.ai itself with suspicion — this is one case where the internet's second-hand numbers and the vendor's own page don't agree.