xAI releases Grok 4.5 for coding agents at $2/$6 per million tokens

• Grok 4.5 shipped as SpaceXAI's first model trained specifically for coding and agents, with Cursor named in the training loop in SpaceXAI's announcement.
• The headline price is $2 per million input tokens and $6 per million output tokens, and SpaceXAI's pricing post claims half the tokens per task, higher tokens per second, and less than half the cost of comparable models.
• Cursor access is live with a launch perk: ericzakariasson's Cursor post says first-week usage is doubled, while SpaceXAI's availability post says limits were reset so users can start free.
• The strongest creator demo came from DannyLimanseta's Cursor test, where Grok 4.5 planned and coded a game boss fight with explosion particles and animation timing.
• The main benchmark caveat is real: one caveat screenshot says an earlier snapshot of the Cursor codebase was unintentionally included in training, with unclear score impact.

The launch link says Grok 4.5 was trained with Cursor; the benchmark caveat screenshot says a Cursor codebase snapshot leaked into training and was removed for future models. DannyLimanseta's game test is the creator-facing demo: brainstorm, plan, implement, then watch a boss fight and particle explosion run inside Cursor. aakashgupta's app screenshot adds a rollout wrinkle, SuperGrok still told him it was Grok 4.

Grok 4.5 is framed as a coding-and-agent model, not a general chat refresh. SpaceXAI says it was trained with Cursor and built for large codebases, long-running tasks, multiple repositories, hundreds of skills, and varied tools.

Launch surfaces:

• SpaceXAI console
• Grok Build
• Cursor

Launch limits and geography:

• SpaceXAI says limits were reset so users can start free.
• Cursor says usage is doubled for the first week in ericzakariasson's Cursor post.
• EU access is not live yet, with availability expected later this month.

SpaceXAI is selling Grok 4.5 on the cost curve: $2 per million input tokens, $6 per million output tokens, half as many tokens per task, higher tokens per second, and less than half the cost of comparable models.

That price changed at least one buying decision immediately. kiaran_ritchie said they were setting up a GLM account, then switched to Grok after the pricing announcement.

The practical claim is simple: one model can stay in the loop from planning to execution because the per-run bill is lower. DannyLimanseta's cost reply said a big feature produced only a small cost bump.

CursorBench is where the launch pitch gets sharp. The chart places Grok 4.5 High at 66.7% for $1.51 per task, with a visible gap between score and spend.

The fuller table in thekitze's benchmark post gives the comparison:

• Fable 5 Max: 70.5%, $17.32, 103,525 tokens, 72 steps.
• Fable 5 Extra High: 68.4%, $11.73, 64,971 tokens, 56 steps.
• Grok 4.5 High: 66.7%, $1.51, 19,521 tokens, 33 steps.
• Fable 5 High: 66.5%, $8.77, 43,747 tokens, 48 steps.
• Grok 4.5 Medium: 65.4%, $1.54, 18,914 tokens, 34 steps.

nicklaunches pointed at the action count: 33 steps for Grok 4.5 High versus 72 for Fable 5 Max.

The benchmark has a training-data caveat attached. The screenshot says Grok 4.5 had an advantage on CursorBench because an earlier snapshot of the Cursor codebase was unintentionally included in training.

The same note says the exact score impact is unclear and that the data was removed for future models. That makes CursorBench useful for cost-per-task texture, but messy as a clean model-quality read.

The DeepSWE screenshot puts Grok 4.5 at 62% pass@1 within each model provider's harness.

The order shown in minchoi's chart:

• Fable max: 66.1%.
• GPT 5.5 xhigh: 64.3%.
• Grok 4.5: 62%.
• Opus 4.8 max: 55.8%.
• Opus 4.7 max: 54%.

DannyLimanseta tested Grok 4.5 in Cursor for several days and used it to brainstorm, plan, and implement an act-one boss fight for a game. The attached video shows Cursor code generation followed by gameplay with boss mechanics and an explosion effect.

His hands-on notes break into five claims:

• Planning: stronger than Composer 2.5, with more edge-case coverage and better confirmation questions.
• Brainstorming: close to Opus 4.8, while Fable still edges it out for creative suggestions.
• Speed: fast enough to keep the build flow moving.
• Visual coding: better animation movement, timing, and particle effects than Composer 2.5.
• Model switching: cheap enough to use from planning through execution.

DannyLimanseta's follow-up said Grok 4.5 did all the coding for the explosion animation, effects, and boss fight mechanics. In another reply, he called it his daily driver in Cursor.

One small workflow detail matters for creative readers: DannyLimanseta said he prefers Cursor because he is not a fan of terminal work as a non-coder.

That explains why the demo lands differently from a raw API benchmark. The model is being judged inside an editor where planning, code edits, visual iteration, and playtesting happen in one loop.

AIandDesign's reply also described Grok 4.5 in Cursor as "Opus-ish" after trying it.

The rollout was not perfectly legible across surfaces. aakashgupta posted a SuperGrok screenshot where the app self-identified as Grok 4 and said Grok 4.5 would drop publicly tomorrow.

SpaceXAI's own access post says Grok 4.5 is available in the console, Grok Build, and Cursor, but not yet in the EU. thekitze's EU screenshot repeats the EU note: no SpaceXAI products or API console access in the region until mid-July.

DannyLimanseta later described the creative pipeline around the code demo:

• Start with a loose game idea.
• Brainstorm with different AI models to expand it.
• Browse Pinterest for art direction.
• Collect a moodboard in Figma.
• Feed the moodboard into a custom image-generation tool.
• Generate art assets for the game.

That pipeline is the real creator story around Grok 4.5: the model enters after taste-making, moodboarding, and asset direction, then turns the playable mechanics into code.