Grok 4.5 Costs $2 a Million Input Tokens. The Benchmark Doesn't Matter.
Grok 4.5 Costs $2 a Million Input Tokens. The Benchmark Doesn't Matter.
SpaceXAI shipped Grok 4.5 yesterday, July 8, 2026. The list price is $2 per million input tokens and $6 per million output. Stop. Read that again. Anthropic's Claude Opus 4.8 is $5 and $25. Claude Fable 5 is $10 and $50. GPT-5.5 and GPT-5.6 sit at $5 and $30. Grok 4.5 is not "competitive on price." It is a different price tier pretending to share a table with the rest of the frontier.
The press coverage will lead with benchmarks because that is what benchmarks are for — generating press. Don't fall for it. The story is not the score. The story is the math, and the math is a body blow to every other vendor pricing their agentic-coding SKU.
The math, because the math is the whole release
Vendor-stated numbers, not my opinion: Grok 4.5 uses 4.2x fewer output tokens than Opus 4.8 on SWE-Bench Pro tasks (15,954 vs 67,020 average) and serves at roughly 80 tokens per second. Multiply that against the per-token price and the cost-per-solved-task gap is enormous:
# Vendor-stated: output tokens per SWE-Bench Pro attempt
opus_48_out_tokens = 67_020
grok_45_out_tokens = 15_954 # 4.2x fewer
# List pricing per token
opus_48 = (5 / 1e6, 25 / 1e6) # $5 in, $25 out per 1M
grok_45 = (2 / 1e6, 6 / 1e6) # $2 in, $6 out per 1M
input_tokens = 50_000 # assume a typical SWE task preamble
def cost(in_price, out_price, out_tokens, in_tokens=50_000):
return in_tokens * in_price + out_tokens * out_price
print(f"Opus 4.8 : ${cost(*opus_48, opus_48_out_tokens):.2f}")
print(f"Grok 4.5 : ${cost(*grok_45, grok_45_out_tokens):.2f}")
# Opus 4.8 : $1.93
# Grok 4.5 : $0.20That is a roughly 10x reduction in cost per SWE-Bench Pro attempt, on a model that also scores competitively on Terminal-Bench 2.1 (83.3% vs Fable 5's 84.3%). For any team running agentic SWE evals at scale, or just shipping an LLM-backed IDE, the cost line on the bill is about to be a tenth of what it was. "About to" if you migrate; the migration is not a small thing, and I will get to the caveats.
The benchmark table, because you will want it
This is the vendor-published table from the Cursor blog, with the annotations Cursor themselves added (they are unusually honest about which numbers they trust):
| Model | DeepSWE 1.1 | Terminal-Bench 2.1 | SWE-Bench Pro |
|---|---|---|---|
| Fable 5 (max) | 70% | 84.3% | 80.4% |
| GPT-5.5 (xhigh) | 67% | 83.4% | 58.6% |
| Opus 4.8 (max) | 59% | 78.9% | 69.2% |
| Grok 4.5 | 53% | 83.3% | 64.7% |
| GLM-5.2 | 44% | 81.0% | 62.1% |
Grok 4.5 is not the smartest. On the harder software-engineering evals it sits squarely in the middle of the frontier pack. Fable 5 is better. Opus 4.8 beats it on SWE-Bench Pro. But the cost story makes those gaps a lot less load-bearing than the leaderboard implies, and the gap that does exist is real and will close — Musk is promising monthly fresh-trained-from-scratch releases for the rest of 2026. We will see.
A small but important disclosure from the Cursor post, which I want to call out because the rest of the field should be doing this and almost none of them do: Cursor's own CursorBench has a contamination problem. An earlier snapshot of the Cursor codebase was accidentally included in Grok 4.5's training data, which is why Grok 4.5 looks unusually strong on CursorBench. They have removed the offending data for future models and are working on a CursorBench refresh. Cursor published this in the same blog post announcing the partnership. Read that again, then think about how many of the other vendors' "internal evals" are quietly benefitting from the same kind of contamination and never saying so.
The flywheel is the real product
Now the part that matters more than any of the above. The headline-grabber is the model; the product is the loop.
SpaceX bought Cursor in mid-June 2026 for $60 billion in stock. Cursor is the AI code editor more than a million developers use daily. Every edit, every agent run, every test, every commit message in Cursor is a training signal. Grok 4.5 was trained on trillions of tokens of Cursor-derived data capturing "a wide range of user interactions with codebases and software tools." The model is now embedded in Cursor on every plan, with double-usage for the first week. So:
1. Cursor users generate the training data. 2. The training data trains the model. 3. The model ships inside Cursor. 4. More users use Cursor because the model is better. 5. Goto 1.
Add a second loop. SpaceXAI's Colossus supercluster in Memphis is the same compute that xAI has been renting to Anthropic and Google while its own models run on the same boxes. That is a vertical integration we have not seen in this market — the model vendor, the IDE vendor, the data source, and (some of) the compute provider are now the same company. The conflict of interest for any other lab paying Colossus rent is a real problem Anthropic will need to address; that is a story for another post, but it is the second-biggest implication of this launch after the price.
What I would actually do
If you are running an agentic coding product, Grok 4.5 is now a serious default to evaluate, not a press-release curiosity. The price moves the unit economics enough that "use Opus" is no longer the obviously-correct default — the right move is to run your private eval, compute cost-per-solved-task, and let that number pick the model. The leaderboard will not pick it for you. It never did.
If you are Anthropic, you have a margin problem and a compute-supplier problem in the same quarter. The Anthropic response to this is the only thing worth watching in July.
If you are OpenAI, GPT-5.6 ships to general availability the same week. The story is not "can GPT-5.6 beat Grok 4.5 on benchmarks." The story is "can OpenAI match SpaceXAI's $2 input price without losing the ability to fund the next training run." That is a much harder question than a benchmark, and the answer probably requires an answer to the first one first.
One more thing, because it is going to come up. The model is not in the EU yet, and the EU AI Act is the reason. If you operate in Europe, plan accordingly and do not let anyone schedule a launch on a date that depends on Grok 4.5 being available there. Mid-July is the target xAI has been quoted on. I would not bet the launch date on it.
Sources
- Cursor — Introducing Grok 4.5 (the official post, including the CursorBench contamination disclosure)
- The Next Web — SpaceXAI ships 'Opus-class' Grok 4.5 to undercut OpenAI and Anthropic
- The Decoder — Grok 4.5 is so cheap compared to Fable 5 and GPT 5.5 that benchmark gaps may not matter much
- Mashable — Elon Musk says 'Opus-class' Grok 4.5 is about to launch
- Digital Applied — Grok 4.5: SpaceX's 1.5T V9 Model Trained on Cursor (private-beta background, Cursor acquisition context)