Model comparison · 2026-09-22

Grok 4.5: why a two-generation-old model is still in the catalog, and when to pin it

Covers that xai/grok-4.5 is two generations behind the current one (Grok 4.7), why it can still be worth considering for an integration wanting version stability, and its cheaper cache-read price versus Grok 4.6.

Comparison diagram showing Grok 4.5 carrying a cheaper cache-read price than Grok 4.6, with only three reasoning_effort levels.

The current generation is Grok 4.7, but 4.5 is still in the catalog

xai/grok-4.5's catalog description states directly that the current generation is Grok 4.7; so this identifier is now two generations behind (4.5 → 4.6 → 4.7). Still, the model page stays open and `isPublic: true` — this repo doesn't remove models, it retires them or, as here, keeps them callable; consistent with the soft-retirement principle in rule 39.

Unlike Grok 4.6's four reasoning_effort levels (low/medium/high/xhigh), Grok 4.5 carries only three: low, medium, high. A task flow needing the xhigh level won't find that option on this identifier.

A concrete difference: Grok 4.5's cache-read price is cheaper

Comparing the two models' pricing cards reveals an interesting detail: Grok 4.5's standard-context cache-read price ($0.30 per million tokens) is cheaper than Grok 4.6's ($0.50); input/output prices are the same ($2/$6). A similar gap exists at the long-context tier too (4.5: $0.60 cache read, 4.6: $1). That means a flow making heavy use of a repeated context prefix can see a slight price advantage on 4.5 in that one line item — but the difference is small and usually doesn't outweigh other factors like access to the xhigh level or the most current model behavior.

The practical takeaway: this small cache difference alone shouldn't be the reason to choose 4.5; the real decision should rest on whether an existing integration is already tested against this identifier, or whether you need the xhigh level.

  • Grok 4.5: only low/medium/high; Grok 4.6 additionally has xhigh.
  • Grok 4.5's cache-read price is slightly cheaper than Grok 4.6's (standard and long context).
  • Input/output unit prices are the same on both models.

When staying on this identifier makes sense

If an existing integration is already tested against Grok 4.5, the three-level reasoning_effort range is sufficient for your task set, and you don't need xhigh, moving to a newer generation without a functional gain adds an unnecessary regression risk. If you're building a new integration, starting with the current generation (Grok 4.7) usually makes more sense.

Frequently asked questions

Is Grok 4.5 cheaper than Grok 4.6?

Input/output unit prices are the same; only the cache-read price is slightly lower on Grok 4.5. The overall cost difference is small in most flows.

Can I use the xhigh level on Grok 4.5?

No, the model page lists only low, medium, and high levels; if you need xhigh, consider Grok 4.6 or a more current identifier.

Related posts