Grok 4.5 shipped publicly on July 8, 2026. It costs $2 per million input tokens and $6 per million output tokens, runs at 80 tokens per second, and is live in Grok Build, in Cursor on all plans, and through the SpaceXAI console. Elon Musk called it “an Opus-class model, but faster, more token-efficient and lower cost.”
Open the benchmark chart xAI published with the launch and a different picture appears. Across the four benchmarks the company chose to show, Grok 4.5 places third or fourth on every one. It never finishes first.
What was announced
.Grok 4.5 became publicly available on July 8, 2026. It is live in Grok Build as the default model, inside Cursor on all plans, and through the SpaceXAI console API. It is not yet available in the EU.
Pricing is $2 per million input tokens and $6 per million output tokens, served at roughly 80 tokens per second across a 500K context window. Claude Opus 4.8 is priced at $5 input and $25 output. GPT 5.6 Sol, which OpenAI launched hours later the same day, is priced at $5 input and $30 output.
The private beta ran at SpaceX and Tesla from late June. At the time, Musk said the model was built on xAI’s 1.5 trillion parameter V9 foundation with Cursor data added during supplemental training. The launch fills in the detail. Grok 4.5 is a mixture-of-experts model trained jointly by xAI and Cursor on trillions of tokens drawn from real Cursor user sessions, capturing how developers actually work rather than how finished repositories look, on tens of thousands of NVIDIA GB300 GPUs. This is the first public model since the SpaceX and xAI merger closed in February 2026, and SpaceX has a pending $60 billion deal to acquire Cursor outright.
Musk’s framing at launch was narrower than the pre-release claim. He called Grok 4.5 an Opus-class model that is faster, more token-efficient and lower cost, and described it as roughly on par with Opus 4.7. Opus 4.7 is Anthropic’s previous flagship. Claude Opus 4.8 succeeded it, and Claude Fable 5 now sits above both.
xAI published five benchmark comparisons. Grok 4.5 places third on DeepSWE 1.0 at 62.0%, ahead of Opus 4.8 at 55.75% but behind GPT 5.5 at 64.31% and Fable 5 at 66.1%. It drops to fourth on DeepSWE 1.1 at 53%, behind Opus 4.8 at 59%. On Terminal Bench 2.1 it scores 83.3%, a tenth of a point behind GPT 5.5 and a point behind Fable 5. On SWE Bench Pro it resolves 64.7%, behind both Opus 4.8 at 69.2% and Fable 5 at 80.4%.
Then there is the fifth chart, which most launch coverage skipped. On SWE Marathon, which measures long-horizon agentic task resolution, Grok 4.5 finishes first at 29.0%. Opus 4.8 scores 26.0%. Fable 5 scores 24.0%. It is the only row Grok 4.5 wins and the only row where Fable 5 does not lead.
The efficiency number is what xAI leads with. On SWE Bench Pro, the company reports Grok 4.5 resolving tasks with an average of 15,954 output tokens against 67,020 for Opus 4.8 at max setting, roughly 4.2 times fewer. For teams running AI agents at volume, output token count is the line item that compounds, not the sticker price.
Read the caveats before reading the chart. Competitor figures come from published system cards rather than xAI’s own runs. Cursor notes that the SWE Bench Pro and Terminal Bench scores for third-party models are self-reported. Cursor also disclosed that an earlier snapshot of its codebase was accidentally included in Grok 4.5’s training data, that the impact is unclear, and that its own CursorBench results were excluded from the public comparison for that reason. Independent verification of any of these numbers does not exist yet.
Why Grok 4.5 matters for marketers
The first takeaway is not speed of releases. It is cost per completed task.
At $2 input and $6 output per million tokens, Grok 4.5 sits at roughly a quarter of what Claude Opus 4.8 charges on the output side. Stack the token efficiency claim on top and the gap widens again. A model that resolves a task in 15,954 output tokens instead of 67,020 is not 60% cheaper. It is cheaper twice over, once on rate and once on volume. For a content ops agent auditing 400 pages, or a reporting agent parsing weekly platform exports, that compounds into a real line item rather than a rounding error.
The counterweight matters just as much. A model that costs less per token but needs a second attempt costs more per finished job. Grok 4.5 finishes fourth on DeepSWE 1.1, behind Opus 4.8. Cheap tokens spent on a wrong answer are not cheap. The only number worth tracking is what it costs to get one correct, usable output, and that number does not appear on any pricing page.
The second takeaway is where Grok 4.5 actually wins. On SWE Marathon, the benchmark measuring long-horizon agentic task resolution, it beats both Opus 4.8 and Fable 5. Sprint benchmarks measure whether a model can fix one bug. Marathon benchmarks measure whether it can hold a plan across dozens of steps without drifting. For marketers, that second thing is the whole ballgame. Campaign QA agents, reporting agents, and content ops workflows are all long-horizon tasks dressed up as chat prompts.
The office work signal points the same direction. xAI says Grok Build can construct multi-sheet Excel models that pull research from the web, build diagrams from native PowerPoint shapes, and write clean prose in Word. Cursor broadened the training mix beyond software engineering to include STEM tasks, research papers, and general knowledge work. That is not a developer story. It is a marketing ops story, and it lands on QBR decks, spend models, and campaign data cleanup before it lands anywhere near a blog post.
This is where the connection to AI prompts for content writing becomes practical. Better models still need better inputs. A stronger model does not fix vague briefs, weak ICP context, missing proof points, or lazy prompts. It just produces cleaner-looking output faster, which is not always a win.
In other words, the useful work happens behind the scenes. Less “write me a viral post” and more “find the 32 pages where our positioning is inconsistent and draft updates by ICP.”
The third takeaway is AI search visibility. Grok is tied closely to X, and that makes it different from models sitting inside standalone chat products. If Grok-powered answers keep expanding across X and related surfaces, marketers need to think harder about how brand narratives, product claims, founder posts, and customer proof get read by AI systems. That connects directly to AI Search SEO. The brands that win AI search will not be the ones repeating the same keyword 19 times. They will be the ones with clear, consistent, well-cited content across pages, social platforms, and third-party sources.
None of this stays still for long. Musk has said xAI plans to ship new foundation models, trained from scratch, roughly every month through the end of 2026. Whatever you conclude about Grok 4.5 has a shelf life measured in weeks.
How Grok 4.5 compares with alternatives
The comparison is not about picking one winner. That is usually how teams waste three weeks and call it “AI strategy.”
Here is what xAI published on launch day, reproduced as posted. Grok’s column is the one to read across, not down.
| Model or platform | Current signal | Marketing relevance | What is still unclear |
|---|---|---|---|
| Grok 4.5 | Publicly available since July 8. $2 input and $6 output per million tokens, roughly 80 tokens per second. Beats Opus 4.8 on three of five benchmarks xAI published, including first place on SWE Marathon | Long-horizon agentic work: campaign QA, reporting pipelines, content audits. Cheap enough to leave running | Every published number is vendor-reported. No independent verification exists yet |
| Claude Opus | Anthropic positions it for coding, agents, enterprise workflows, and complex knowledge work. Now sits below Claude Fable 5 in Anthropic’s own lineup | Strong fit for research, long-form content, planning, and review-heavy workflows | Higher cost than lighter models, and teams still need human review |
| ChatGPT | Broad adoption across content, research, coding, and general productivity. GPT 5.6 Sol shipped the same day as Grok 4.5 | Strong everyday tool for ideation, campaign planning, and fast drafting | Model choice and workflow discipline matter more than brand familiarity |
| Gemini | Strong Google ecosystem connection | Relevant for SEO, analytics, docs, and Google-native workflows | Best use depends on how deeply a team works inside Google’s product stack |
The comparison is not about picking one winner. That is usually how teams waste three weeks and call it “AI strategy.”
Most launch coverage reported four benchmarks and concluded Fable 5 leads all of them. There are five. On SWE Marathon, which measures long-horizon agentic task resolution, Grok 4.5 scores 29.0% against 26.0% for Opus 4.8 and 24.0% for Fable 5. It is the only row Grok 4.5 wins, and the only row Fable 5 does not lead. It is also the row that most resembles the work marketing agents actually do.
Read the caveats before the numbers. Cursor notes that SWE Bench Pro and Terminal Bench scores for third-party models are self-reported. Cursor also disclosed that an earlier snapshot of its own codebase was accidentally included in Grok 4.5’s training data, and excluded CursorBench from the comparison as a result. xAI’s chart shows Fable 5 at 84.3% on Terminal Bench 2.1, while the independently verified leaderboard puts it at 88.0%.
So the Grok 4.5 vs Claude Opus question resolves less cleanly than either camp wants. “Opus-class” is a fair label for a model that splits five head-to-heads three to two. “Beats Opus” is not, and that is closer to how the quote has been repeated all week.
For most marketing teams, the better move is a model map. Use Claude Opus for deep reasoning and review-heavy workflows where a wrong answer is expensive to catch. Use ChatGPT where speed and broad familiarity matter. Use Gemini where Google context matters. Use Grok 4.5 for high-volume, long-horizon agentic work where errors surface quickly and cost little to correct.
If your team already works across SEO, paid media, and content, this model map should sit inside your broader performance marketing operating system, not outside it as a shiny side project.
What marketing teams should watch next
- The first thing to do is test it. Grok 4.5 is free in Grok Build and inside Cursor on all plans for a limited window, which means the evaluation costs you hours, not budget. Give it a content brief, a messy paid media export, a competitor positioning task, and a search visibility audit. Then compare the output against Claude Opus, ChatGPT, and Gemini on accuracy, time saved, review effort, and usefulness. Run the same four tasks every time a model ships and the comparison becomes a habit instead of a project.
One constraint before you plan around it. Grok 4.5 is not yet available in the EU. For B2B SaaS teams with EU data residency requirements, that rules it out of client-facing workflows until availability opens up.
- The second thing to watch is reliability. Grok has had public controversy around safety, misinformation, and output quality. For marketers, that matters because one bad AI-assisted claim in a comparison page, ad brief, or landing page can create legal and brand problems. The cleanup is never as fast as the draft. Independent testing of the launch benchmarks does not exist yet, and the one dataset issue that has surfaced so far came from Cursor’s own disclosure rather than an outside audit.
- The third is whether Grok 4.5 becomes useful for paid media automation. If it can reason well across structured data, messy platform exports, and audience signals, it could help with campaign diagnostics, ad variation planning, and spend analysis. The SWE Marathon result is the reason to bother testing this. A model that holds a plan across dozens of steps is a better bet for a reporting pipeline than one that wins a single-shot benchmark. But until that is proven in real accounts, treat it as something to test, not something to trust blindly.
- The fourth is SaaS GTM strategy. AI model releases are becoming positioning events. Musk framed Grok 4.5 against Claude Opus, then published a chart showing Grok losing four of five head-to-heads to Claude Fable 5. That is either unusual transparency or an unforced error, and it is worth deciding which before you copy the playbook. Anthropic frames Claude around enterprise reliability and honesty. OpenAI and Google are building their own distribution advantages. Every B2B SaaS founder should notice the GTM lesson: the model is the product, but the comparison is the campaign.
OneMetrik has covered similar AI infrastructure and model-positioning shifts before, including the Anthropic and SpaceX partnership and Claude Opus 4.8. Grok 4.5 belongs in that same bucket: less of a chatbot update, more of a signal that the AI tooling layer marketers depend on is still being rewritten in public.
OneMetrik takeaway
Grok 4.5 is not a benchmark story. It is a price story.
Ignore the leaderboard question. Track cost per completed task instead. A model that costs 60% less per token but needs a second attempt is not cheaper. A model that finishes third on four benchmarks while resolving each task in a quarter of the output tokens might be. Neither of those numbers appears on a pricing page, and neither shows up in a benchmark chart. You get them by running your own work through both models and counting.
The real advantage will not go to the team that tries every new model first. It will go to the team that knows exactly what to test when a better model arrives.