Grok 4.5 benchmark comparison chart

xAI shipped Grok 4.5 on July 8. The headline is not that it beats anyone on benchmarks. It doesn't. The headline is that it costs $2 per million input tokens and uses 4.2 times fewer output tokens than Opus 4.8 to solve the same task. For a small team running agentic coding day in and day out, that arithmetic changes everything.

The model is a mixture-of-experts system co-trained with Cursor on trillions of tokens of developer-agent interaction data. Not scraped public repos. Not synthetic code. Real sessions where developers prompted, iterated, read errors, fixed bugs, and verified results. That feedback loop is what gives Grok 4.5 its token efficiency, and it's something no other frontier model can replicate without access to that same dataset.

How Cursor's data changed the math

When SpaceX acquired Cursor (then Anysphere) for $60 billion in June, the immediate assumption was that xAI would get access to Cursor's codebase. What actually happened was more interesting: they got access to the training signal. Trillions of tokens of how developers actually work through problems with AI assistance. The model doesn't just know what correct code looks like. It knows the messy middle part where you try three approaches, read a stack trace, and eventually land on something that works.

Cursor's own post says Grok 4.5 is their "most intelligent model and the first we've built for more than software engineering." That's the strategic shift. Composer 2.5 was a coding specialist. Grok 4.5 was deliberately trained on a broader mix: STEM tasks, research papers, knowledge work. The result is a model that handles coding, data science, finance, and legal work in one weight class.

The pricing table tells the story:

Model Input Output
Grok 4.5 $2/M $6/M
Grok 4.5 Fast $4/M $18/M
GPT 5.5 $5/M $30/M
Claude Opus 4.8 $5/M $25/M
Claude Fable 5 $10/M $50/M

The 4.2x token efficiency claim comes from SWE-Bench Pro: Grok 4.5 resolves tasks with roughly 15,700 output tokens where Opus 4.8 uses 67,020. Cheap price multiplied by fewer tokens is where the real gap opens up. A task that costs $0.40 on Grok 4.5 costs $1.68 on Opus 4.8. Run that across a month of agent runs and the gap becomes thousands of dollars.

The benchmark honesty problem

Here's where the marketing gets tricky. xAI frames Grok 4.5 as "around Opus 4.7 level" performance. That's one generation behind the current frontier. On the four main coding evals, Grok 4.5 sits mid-pack. Fable 5 leads all four. Opus 4.8 and GPT 5.5 beat it on several. On DeepSWE 1.1, it's fourth of five.

Cursor also disclosed that an earlier snapshot of its own codebase accidentally ended up in training, giving Grok an unfair edge on CursorBench. They removed the data and are rebuilding the benchmark. Credit for owning it publicly, but it does raise questions about training-data contamination in general.

The independent Thomas Wiegold review tested Grok 4.5 against GPT 5.5 and Claude on three coding tasks. His verdict: "Grok 4.5 is a genuinely good coding model now, and at $30 a month on SuperGrok it's an absurd amount of value." He also noted it caught a bug that no other model had flagged for him, including the ones he rates higher.

The trust problem nobody's talking about

The HN thread hit 672 points and 1,077 comments. A significant chunk of the discussion was not about benchmarks or pricing. It was about whether developers will use an xAI model regardless of quality.

One commenter: "Even without the politics, Elon has shown that he will weaponize his platforms against people/companies he personally doesn't like. Using Grok is therefore a supply chain risk."

Others pushed back: "Americans are 4% of the world's population, and even among those 4% at least half don't give a shit. The rest of us give even less of a shit, we don't have the luxury to be principled."

The pricing also raised sustainability questions. xAI reported $2.5 billion in operating losses last quarter. Some theorize the aggressive pricing is possible because xAI has excess compute capacity from their massive GPU build-out sitting partially idle. Others wonder how long $2/M input pricing can last when the infrastructure costs don't add up.

What this actually changes

The real disruption isn't the model. It's the vertical integration. SpaceXAI now owns both the IDE (Cursor) and the weights (Grok). That's a tighter feedback loop than any competitor has. The same company that sees how developers work can train the model that assists them. The same company that serves the model can optimize the inference stack around it.

For developers, the immediate question is whether Grok 4.5's broader training makes it worth the cost premium over Composer 2.5. Several HN commenters noted that Composer 2.5 handles most coding tasks well and is much cheaper to run. The answer probably depends on whether you need the cross-domain capabilities or just pure code generation.

The EU availability gap is worth noting. Grok 4.5 is not yet available in Europe, expected mid-July. For teams with EU-based developers, this is a non-starter regardless of the pricing advantage.

xAI's approach is different from the "beat everyone on benchmarks" playbook that OpenAI and Anthropic run. They're betting that cost-per-task matters more than leaderboard position, and that training data from real developer workflows creates an efficiency advantage that benchmarks don't capture. Whether that bet pays off depends on whether the market cares more about the last two percentage points on SWE-Bench or about the invoice at the end of the month.

Sources