Grok 4.5 undercuts Opus 4.8 by two-thirds and burns fewer tokens getting there. The catches are a context window that shrank to 500K and a surcharge xAI won't put a number on.
xAI shipped Grok 4.5 on July 8 at $2 input and $6 output per million tokens, its first model built specifically for coding and agent work. Against Opus 4.8, Fable 5, and GPT-5.6 Sol it is 68 to 84 percent cheaper per token, and because it spends far fewer tokens per task, the real gap is wider than the rate card shows. Two things keep it honest: the context window dropped to 500K, and the long-context surcharge is a rate xAI mentions but never prints. Here is what the price actually buys.
Photo by Zoha Gohar on Unsplash
Two dollars in, six dollars out
| Line | Per 1M tokens |
|---|---|
| Input (up to 200K tokens) | $2.00 |
| Cached input | $0.50 |
| Output | $6.00 |
| Input above 200K tokens | Higher, unpublished |
USD per million tokens, from the xAI API docs and Artificial Analysis, which had pre-release access. Cached input reads at $0.50, a 75 percent discount off the input rate. Context window is 500K tokens. Server-side tools bill on top of tokens: web and X search run about $5 per 1,000 calls.
Cheaper than the frontier, more expensive than the floor
The pitch xAI wants you to hear is Opus-class capability at a fraction of Opus-class price, and on a straight per-token basis that holds. Price a mid-size coding-agent month at 30M input tokens and 6M output, no cache. Grok 4.5 comes to $96. The frontier it is aiming at is three to six times that: Opus 4.8 at $300, GPT-5.6 Sol at $330, Fable 5 at $600. If your bill today reads Opus or Sol, Grok 4.5 is a two-thirds cut on the token line alone.
| Model | Input / output | 30M in / 6M out |
|---|---|---|
| Grok 4.3 (predecessor) | $1.25 / $2.50 | $52.50 |
| GPT-5.6 Luna | $1 / $6 | $66.00 |
| Grok 4.5 | $2 / $6 | $96.00 |
| Gemini 3 Pro | $2 / $12 | $132.00 |
| GPT-5.6 Terra | $2.50 / $15 | $165.00 |
| Claude Opus 4.8 | $5 / $25 | $300.00 |
| GPT-5.6 Sol | $5 / $30 | $330.00 |
| Claude Fable 5 | $10 / $50 | $600.00 |
Read the top of the table, though, and the marketing gets more careful. Grok 4.5 is not the cheapest token you can buy. GPT-5.6 Luna matches it on output and beats it on input, landing at $66 for the same month. Grok's own last-generation model, Grok 4.3, is cheaper on both lines and finishes at $52.50. So the accurate sentence is narrow: Grok 4.5 is the cheapest way into Opus-class coding, not the cheapest model on the board. If raw price is the only axis you care about, you were already shopping below it.
What makes the $2/$6 interesting is not where it sits on this list. It is that the list understates the gap, because a fixed token budget assumes every model spends tokens the same way. They do not.
The real advantage is how few tokens it spends
The per-token table above pays every model for the same 6M output tokens. In practice, Grok 4.5 finishes the same task in far fewer. Artificial Analysis clocked it at roughly 14,000 output tokens per Intelligence Index task, and on SWE-bench Pro it closes the same tickets using about 4.2 times fewer output tokens than Opus 4.8, near 16,000 against 67,000. Output tokens are the expensive line, so spending a quarter of them compounds directly onto the price you already saw.
Fold efficiency and rate together and the honest comparison is cost per finished task, not cost per million tokens. On Artificial Analysis's own harness that lands Grok 4.5 near $2.50 for a representative coding task, against about $5 for GPT-5.5 and close to $12 for Fable 5. Same work, a quarter to a fifth of the outlay. That is the number xAI is actually selling, and unlike the surcharge, it is one an independent lab measured.
| Model | Cost per coding task | Why |
|---|---|---|
| Grok 4.5 | ~$2.50 | Cheap output plus ~4x fewer tokens per task |
| GPT-5.5 | ~$5.07 | Higher rate, heavier token spend |
| Claude Fable 5 | ~$11.80 | Top capability, most tokens and highest rate |
A caveat worth keeping: token efficiency is workload-dependent. A model that writes terse patches on SWE-bench may still ramble on your codebase, and a reasoning-heavy prompt can erase the gap. The safe read is that Grok 4.5's per-task advantage is real and independently seen, but the exact multiplier is yours to measure, not to assume.
The context window quietly halved
Here is the detail most launch coverage skipped. Grok 4.3 shipped with a 1M-token context window at a flat rate. Grok 4.5 comes in at 500K, half the room, and adds a threshold at 200K tokens above which xAI says input bills at a higher context rate. The company mentions the surcharge and declines to print the multiplier. Secondary trackers fill the blank with roughly double, so $4 input and $12 output past the line, but that figure is aggregated from third parties, not lifted from xAI's own page. Until the docs confirm a number, it is a guess wearing a decimal point.
For most chat and short-agent work this never bites, because you never cross 200K in a single request. For long-repo coding sessions, RAG over big corpora, or anything that stuffs a whole codebase into context, it matters twice: you lost half the window versus the model you might be upgrading from, and the part you kept can cost more than the sticker. Price the flat $2/$6 for planning and treat sub-200K as the regime where the headline number is the real number.
The benchmarks come with a Cursor asterisk
The capability read is better sourced than most launch-day models get, because Artificial Analysis had Grok 4.5 before release. They put it at Intelligence Index 54, ranked fourth behind Fable 5, GPT-5.5, and Opus 4.8 at 56, and a full 16 points above Grok 4.3. On the coding benchmarks it trails the very top: SWE-bench Pro 64.7 percent against Opus 4.8's 69.2 and Fable 5's 80.4, but it clears GPT-5.5, and Terminal-Bench 2.1 lands it at 83.3, a hair off the leaders. For a model priced where it is, losing the benchmark by a few points and winning the invoice by two-thirds is a trade a lot of teams will take.
| Benchmark | Grok 4.5 | Field |
|---|---|---|
| AA Intelligence Index | 54 | #4, behind Fable 5, GPT-5.5, Opus 4.8 (56) |
| SWE-bench Pro | 64.7% | Opus 4.8 69.2, Fable 5 80.4, GPT-5.5 58.6 |
| Terminal-Bench 2.1 | 83.3% | Fable 5 84.3, GPT-5.5 83.4, Opus 4.8 78.9 |
The asterisk is on the coding scores specifically. Grok 4.5 was trained on real Cursor developer session data and launched jointly with Cursor, and Cursor disclosed that an earlier codebase snapshot was accidentally folded into training. That contamination can quietly inflate Cursor-style coding evals in ways nobody has cleanly measured yet. It does not touch the price or the token-efficiency numbers, both of which stand on their own, but it means the SWE-bench line deserves the same "run it on your own repo" skepticism you would give any vendor-adjacent benchmark. Musk's framing, that this is "an Opus-class model, but faster, more token-efficient and lower cost," is roughly what the independent numbers say, minus the coding asterisk.
Move if the bill is the problem
If you are running an agent or coding workload on Opus 4.8, GPT-5.6 Sol, or Fable 5 and the bill is the thing that hurts, Grok 4.5 is the clearest downgrade-in-price, hold-the-capability move on the market this month. The token rate is a fraction of what you pay, the efficiency compounds it, and the one benchmark gap is small enough to pilot against real work. Point non-critical traffic at it, diff a week of outputs against your incumbent, and let the patches decide.
Stay put in two cases. If you are already on a budget tier, Luna, Grok 4.3, or a Chinese open-weight coder, Grok 4.5 is a step up in price, not down, and you should switch only for the capability, not the savings. And if your workload lives past 200K tokens per request, the halved window and the blank surcharge line make this a model to price carefully before you commit. Drop your real input-output mix into the cost calculator and compare it head to head on the pricing page, because at this rate the only question that matters is whether it holds up on your tasks, not whether it saves money. It does.
Sources
- - xAI Grok 4.5 announcement: x.ai/news/grok-4-5
- - xAI API models and pricing: docs.x.ai
- - Artificial Analysis evaluation: artificialanalysis.ai
- - The Decoder (pricing and token efficiency): the-decoder.com
- - DevOps.com (coding-agent pricing): devops.com
- - Related: Grok 4.3 pricing and benchmarks
- - TokenCost pricing page: tokencost.app/pricing