Grok 4.5 at $2/$6: The Coding Model That Comes With Its Own Editor

The most interesting number in the Grok 4.5 release isn't a benchmark score. It's a token count.

On SWE Bench Pro, Grok 4.5 solved tasks using an average of 15,954 output tokens where Claude Opus 4.8 (max) used 67,020 - a 4.2x difference (MarkTechPost). In an era where every serious coding tool meters by the token, verbosity is a line item. A model that reaches an answer in a quarter of the words is cheaper than its list price suggests, and Grok 4.5's list price was already aggressive: $2 per million input tokens, $6 per million output, serving at 80 tokens per second.

A sourcing note before the numbers, because we hold ourselves to this: SpaceXAI's own announcement page rejects automated verification, so the pricing and benchmark figures in this piece come from MarkTechPost's launch report, cross-checked against Yahoo Tech's coverage. The $2/$6 pricing appears consistently across outlets; the benchmark tables are the vendor's own numbers relayed through press coverage. Treat them accordingly.

Released July 8 - one day before GPT-5.6 - Grok 4.5 is a model built specifically for coding and agentic work, and it "was trained alongside Cursor," the AI coding editor. That relationship is bigger than a training footnote, and we'll get to why.

The scoreboard, honestly read

The published comparisons (vendor-reported; independent replication pending):

Read the fine print with us: Grok 4.5 beats Opus 4.8 on these tests, roughly ties or slightly trails GPT-5.5, and - by the vendor's own admission - Anthropic's Fable (max) "leads across all four benchmarks tested," with Grok staying closest on Terminal Bench. It also claims the top spot on Harvey's Legal Agent Benchmark, an office-work test that hints at ambitions beyond code.

So the honest positioning is not "new best coding model." It's "near-frontier capability at a fraction of frontier verbosity and price." That's a real category, and right now it's the most contested one in the market.

The $6-output-token knife fight

Step back and look at what July 2026 did to the budget-capable tier. Within nine days, three vendors converged on nearly the same price point:

• Grok 4.5: $2 / $6 (July 8)
• GPT-5.6 Luna: $1 / $6 (July 9)
• Claude Sonnet 5: $2 / $10 intro until August 31, then $3 / $15 (June 30)

A year ago, capable agentic coding meant $15-30 per million output tokens. Today three models sell it for $6-10, each with a different edge: Luna has the lowest input price and OpenAI's ecosystem, Sonnet 5 has Claude Code and the strongest agentic-tooling story, Grok 4.5 has the verbosity advantage - fewer tokens per solved task, which the sticker price doesn't capture.

Here's what that gap means in dollars. If the SWE Bench Pro efficiency ratio holds even loosely on real work, a model charging $6/M that uses 16K tokens per task costs about $0.10 per task in output; one charging $10/M that uses 60K tokens costs about $0.60. Efficiency ratios from one benchmark won't transfer cleanly to your codebase - but the gap is wide enough to be worth measuring on your own workload rather than dismissing.

One practical restriction to know before you plan a trial: Grok 4.5 is not yet available in the EU, either in SpaceXAI's products or the API, with EU access expected in mid-July per launch coverage (Yahoo Tech). European teams are on a short forced wait.

The Cursor relationship is not a partnership - it's an acquisition in progress

When a press release says a model was "trained alongside" a commercial editor, the natural read is a co-marketing deal. The corporate reality here is much bigger, and it's the part a tool chooser actually needs to know.

Start with who released this model. xAI was acquired by SpaceX in February 2026 and is being folded into a combined SpaceXAI (Yahoo Tech) - which is why you'll see both names in coverage. Then, on June 16, SpaceX agreed to acquire Anysphere - the company behind Cursor - in an all-stock deal valuing it at $60 billion, days after SpaceX's Nasdaq debut, with the transaction expected to close in Q3 2026 subject to regulatory approval (TechCrunch).

So Grok 4.5 - about which Cursor's own statement reads "We've partnered with SpaceXAI to train Grok 4.5," and which was available in Cursor from day one (Yahoo Tech) - is not evidence of an open question about whether editor-specific model tuning becomes a moat. It's the answer arriving: the model vendor is buying the editor. Pending close, "which model" and "which editor" stop being independent choices inside this stack - they're the same company's product decisions.

For Cursor users, the near-term read is straightforward: Grok 4.5 has, in effect, practiced in your gym, and trialing it as your session model is the lowest-friction test of whether that shows up in your results. The longer-term read deserves more caution than launch coverage gives it. If you're choosing Cursor today, you're also choosing SpaceXAI's roadmap - including how long rival models stay first-class citizens in an editor whose owner ships a competing model. Nothing has changed yet, and Cursor still runs Claude and GPT models today. But "nothing has changed yet" is a sentence with a shelf life, and our editor comparison will track it.

Should you try it?

For token-metered agent pipelines with real volume, yes - this is the strongest candidate audience, because the efficiency claim, if it replicates on your tasks, compounds with the low price. Benchmark it against your incumbent on your own workload before committing anything. Cursor users are a special case in the same direction: the co-training makes in-editor results potentially better than the public numbers suggest, and the trial costs almost nothing.

If you need maximum capability regardless of cost, this isn't your model - by the vendor's own comparison table, Fable-class models still lead. And if you're standardizing a team this quarter, hold the decision until independent benchmark replications land. Every score above is vendor-reported, launch week, from a lab with a history of confident claims. The price is real today; the scores need a few weeks of outside scrutiny. EU teams don't get a vote yet either way.

The market takeaway is bigger than the model: with Grok 4.5, Luna, and Sonnet 5 all crowding the same band, the cost of good-enough agentic coding fell roughly 60% in twelve months. Whichever of the three you pick, the era of defaulting to the most expensive model out of caution is over. Current per-model numbers live on our Cursor pricing page and the linked model coverage above.