Grok 4.5 Cost Per Task: The 4.2x Efficiency Test

SpaceXAI shipped Grok 4.5 on July 8, 2026 at $2 per million input tokens and $6 per million output, roughly half the per-token rate of Claude Opus 4.8 and a fifth of Claude Fable 5. The launch headline was a “4.2x token efficiency” claim and an Opus-class capability pitch. One of those holds up. The other depends entirely on whose benchmark harness you read.

What Grok 4.5 actually costs

Start with the half of this that is not contested. Grok 4.5 is priced at $2 per million input tokens and $6 per million output, per the SpaceXAI announcement (Introducing Grok 4.5, July 2026). It carries a 500,000-token context window and serves at a claimed 80 tokens per second, which Artificial Analysis independently measured at 91.3 output tokens per second. The model was trained jointly with Cursor on trillions of tokens of real coding-editor interaction, as reported by InfoWorld, which is why the pitch is aimed squarely at agentic software work.

Against the two Anthropic models this site tracks, the sticker gap is large. Claude Opus 4.8 bills $5 per million input tokens and $25 output; Claude Fable 5 bills $10 and $50. Grok 4.5 undercuts Opus on output by more than 4x and Fable by more than 8x. That is the same pattern documented in why Chinese AI models are so cheap: get close enough on capability to be credible, then force the comparison to be about price. VentureBeat framed the launch exactly that way, as a half-price bet aimed at rattling Anthropic and OpenAI.

But a per-token rate is not a cost per task, and the whole point of this site is that the two diverge. A model that thinks in fewer tokens can be cheaper to finish a job even at a higher rate, and a model that rambles can be expensive even when it looks cheap on the rate card. The only number that matters to a budget is dollars per finished task.

The cost-per-task math

There are two honest ways to price a task, and they agree on direction.

The first is independent and all-in. On the Artificial Analysis Coding Agent Index, which runs each model inside a real agent harness against a fixed task set, an average task costs $2.49 in Grok Build, $5.07 for GPT-5.5 in Codex, and $11.80 for Claude Fable 5 in Claude Code (figures via InfoWorld). Grok is the cheapest of the three by a wide margin.

Cost per task by agent harness (Artificial Analysis Coding Agent Index)

Tool	Cost per task	Multiple of baseline
Grok 4.5 (Grok Build)	$2.49	1.0x
GPT-5.5 (Codex)	$5.07	2.0x
Fable 5 (Claude Code)	$11.80	4.7x

The second way is a transparent reconstruction, the kind of modeled cost this site prefers because anyone can reproduce it. On SWE Bench Pro, Grok 4.5 resolved the average task using 15,954 output tokens; Claude Opus 4.8 (max) used 67,020. Multiply each by its output rate:

Model	Output tokens / task	Output rate	Output cost / task
Grok 4.5	15,954	$6 / Mtok	$0.096
Claude Opus 4.8	67,020	$25 / Mtok	$1.68

That is a 17.5x gap in output cost per task, far wider than the 4.2x token ratio alone. The reason is that the two multipliers stack: Grok writes about 4.2x fewer output tokens, and each of those tokens is priced about 4.2x lower than Opus. Efficiency and rate compound. This is output cost only, so absolute task cost is higher once input and tool tokens are added, but the ratio between the two models is what a switching decision turns on.

So the cost claim survives scrutiny. If the only question is “does Grok 4.5 cost less per task than Opus 4.8,” the answer is yes, and by a lot. The harder question is what you give up to get there.

Why the efficiency claim is harness-contingent

Here is where the launch narrative gets slippery. “4.2x more efficient” and “Opus-class” are presented together, as if the second follows from the first. It does not, and the benchmark table shows why.

Coding benchmarks now come in two flavors: runs where each vendor uses its own agent scaffolding, and neutral runs where every model uses the same harness. The distinction is the entire subject of DeepSWE vs FrontierCode: two ways to grade AI code, and it decides who wins here.

Grok 4.5 vs Opus 4.8 across provider-harness and neutral DeepSWE runs

Item	DeepSWE 1.0 (xAI harness)	DeepSWE 1.1 (neutral)
Grok 4.5	62%	53%
Opus 4.8	55.75%	59%

On DeepSWE 1.0, the provider-harness run, Grok 4.5 posts 62.0 against Opus 4.8 at 55.75, and it also edges ahead on Terminal Bench 2.1 (83.3 vs 78.9). Read only those, and “Opus-class” looks conservative. But on DeepSWE 1.1, the neutral run where the scaffolding advantage is removed, Grok falls to 53 while Opus rises to 59, per the benchmark table reported by MarkTechPost. On SWE Bench Pro resolve rate, another neutral measure, Opus leads 69.2 to 64.7. And on the Artificial Analysis Intelligence Index, an aggregate, Grok scores 54 to Opus 4.8’s 56, ranking fourth overall behind Fable 5, Opus 4.8, and GPT-5.5.

The pattern is consistent: Grok 4.5 wins the runs xAI controls and loses the runs it does not. That is not an accusation of cheating; a vendor tuning its own harness for its own model is normal, and Grok was literally trained on Cursor agent traces, so its scaffolding is genuinely good. But it means the token-efficiency number is partly a property of the harness, not the raw model, and readers deciding whether the capability claim is real should weight the neutral runs. This is the same skepticism the are AI benchmarks reliable pillar applies to every leaderboard: a score is only as trustworthy as the setup that produced it, and you can check where each model lands on the live AI model leaderboard and model tracker.

The price-as-weapon pattern

Strip out the harness noise and the strategy is clear. Grok 4.5 is a near-frontier model, a couple of index points behind Opus 4.8 on neutral evals, priced to make those couple of points feel irrelevant. It is the same move that turned the Chinese labs into a pricing problem for the incumbents, now run by a US company with SpaceX behind it. The bet is that most buyers will not pay a 4x rate premium for a benchmark edge they cannot feel in daily work.

For a lot of agentic coding, that bet is sound. When a model finishes the task, the cheaper one wins the invoice. The risk lives in the tasks it does not finish: a 64.7 percent resolve rate against 69.2 means Grok bounces off more hard problems, and a bounced task can cost more in retries and human cleanup than the token savings returned, the hidden cost of AI-generated code that never shows on the rate card. The right frame is the same one that governed the Fable 5 vs Opus 4.8 cost-per-task comparison: cheaper per attempted task is not the same as cheaper per solved task, and the gap between them is your failure rate.

Bottom line

Grok 4.5 is genuinely, verifiably cheaper per task than Claude Opus 4.8, by a multiple, not a margin. That claim survives both the independent index and the modeled math. Use it for high-volume agentic coding where tasks are well-scoped, retries are cheap, and the invoice is the constraint.

What does not survive is the implication that cheaper comes free of capability cost. On every neutral benchmark, Opus 4.8 is a step ahead, and Grok’s flashiest wins come from runs xAI tuned itself. If your work lives in the hard tail, where a failed task is expensive and a near-frontier model that misses is worse than a frontier model that lands, the 4x price gap is buying you something real. Price the failure rate, not just the token rate, and the answer usually picks itself.

Sources

SpaceXAI (2026). Introducing Grok 4.5. x.ai (vendor announcement). https://x.ai/news/grok-4-5

Artificial Analysis (2026). Grok 4.5 model page and Coding Agent Index. Artificial Analysis (independent evaluation). https://artificialanalysis.ai/models/grok-4-5

InfoWorld (2026). SpaceXAI launches Grok 4.5, touts lower coding-task costs than AI rivals. InfoWorld (reporting). https://www.infoworld.com/article/4194895/spacexai-launches-grok-4-5-touts-lower-coding-task-costs-than-ai-rivals.html

MarkTechPost (2026). SpaceXAI Releases Grok 4.5, a Cursor-Trained Model for Coding, Agentic Tasks, and Knowledge Work at $2/M Input. MarkTechPost (reporting). https://www.marktechpost.com/2026/07/08/spacexai-releases-grok-4-5/

VentureBeat (2026). SpaceX’s Grok 4.5 launches at half the price of rivals. VentureBeat (reporting). https://venturebeat.com/technology/spacexs-grok-4-5-launches-at-half-the-price-of-rivals-heres-why-that-could-rattle-anthropic-and-openai

Anthropic (2026). Claude API pricing (Opus 4.8 at $5 / $25 per Mtok). claude.com (primary). https://claude.com/pricing

Frequently asked questions

How much does Grok 4.5 cost?

Grok 4.5 is priced at $2 per million input tokens and $6 per million output tokens, with a 500,000-token context window. That is about half the per-token rate of Claude Opus 4.8 ($5 / $25) and a fifth of Claude Fable 5 ($10 / $50).

Is Grok 4.5 really cheaper per task than Claude Opus 4.8?

Yes. On the independent Artificial Analysis Coding Agent Index a task costs $2.49 in Grok Build. A modeled comparison on SWE Bench Pro output tokens puts Grok at about $0.10 per task versus $1.68 for Opus 4.8, because Grok uses roughly 4.2x fewer output tokens and each token is priced about 4.2x lower.

Is Grok 4.5 as capable as Claude Opus 4.8?

Not quite, on neutral benchmarks. Grok 4.5 leads Opus on the xAI-run DeepSWE 1.0 harness (62.0 vs 55.75) but trails on the neutral DeepSWE 1.1 run (53 vs 59), on SWE Bench Pro resolve rate (64.7 vs 69.2), and on the Artificial Analysis Intelligence Index (54 vs 56). The Opus-class claim depends heavily on which harness is used.

What does the 4.2x token-efficiency claim mean?

On SWE Bench Pro, Grok 4.5 resolved the average task using 15,954 output tokens against 67,020 for Opus 4.8, about 4.2x fewer. Fewer output tokens plus a lower output rate is what makes Grok cheaper per task, but the efficiency figure comes partly from the xAI Cursor-trained agent harness, not the raw model alone.