Grok 4.5 has one number worth interrupting a model-stack review for: the Grok API lists grok-4.5 at $2 per million input tokens and $6 per million output tokens as of July 10, 2026. My call: test it now for agentic coding and terminal automation, but keep the Opus-class migration claim behind a cost-per-accepted-task gate.
SpaceXAI is the company name used in Genαi's research after the July 6, 2026 rebrand from xAI. Older docs and search behavior still say xAI, so xAI Grok 4.5 and SpaceXAI's Grok 4.5 refer to the same launch.
In practical terms, Grok 4.5 is a proprietary hosted coding and agentic model with a 500K-token context window, image input, function calling, code execution, and low/medium/high reasoning effort under one model string, according to the xAI API overview. The price is real. The broad Opus claim still needs your own benchmark.
TL;DR
Grok 4.5 is commercially interesting because its official API price is low and Genαi's research records an 83.3% Terminal-Bench 2.1 score. The same research puts it behind Claude Opus 4.8 on reported SWE-Bench Pro and neutral DeepSWE evidence, with Grok 4.5 ranked #4 of 168 on the Artificial Analysis Intelligence Index at score 54.
Treat Grok 4.5 as a candidate for cheaper coding agents, terminal copilots, and Cursor-heavy workflows. Treat it as a proven Opus replacement only after it wins on accepted-task cost, retries, review burden, and failure severity inside your own workload.
Key takeaways
- Genαi's launch-watch research verifies July 8, 2026 developer access and July 9 public availability using SpaceXAI's launch page, Reuters launch coverage, and TechCrunch's report.
- Grok 4.5 costs $2/M input and $6/M output through the official API, with $0.50/M cached input and $5 per 1,000 Web/X search tool calls in the model docs.
- Musk's quoted “Opus-class” framing is partly a price-performance claim. It is not settled by the benchmark package in the supplied research.
- Grok 4.5 ranks #4 of 168 on Artificial Analysis with score 54, behind Fable 5, GPT-5.5, and Claude Opus 4.8, according to Genαi's research.
- The launch has transparency gaps: no system card, no model card, no architecture paper, and no public safety-evaluation package as of July 10, 2026.
Grok 4.5 launch facts and gaps
The launch-window finding is a composite. Genαi Research Desk verified July 8 developer access and July 9 public availability from SpaceXAI's first-party page plus Reuters and TechCrunch coverage; each source covers a different part of the rollout.
Reuters framed the model as built for “coding, agentic tasks.” That matches the API surface: function calling, retrieval, code execution, Web/X search tools, and remote MCP-style orchestration appear in the xAI API overview.
The model string is simple: grok-4.5. Developers set reasoning effort per request instead of choosing separate reasoning and non-reasoning SKUs.
There is a real regression hiding inside the launch. Genαi's research records Grok 4.5 at 500K context, down from Grok 4.3's reported 1M context, and native video input was removed from the launch spec.
That matters for long-document review and multimodal triage. A cheaper coding model can still be the wrong model for a 900K-token contract corpus or a video-grounded support workflow.
The architecture claims deserve quarantine. V9, V9-Medium, and the widely repeated 1.5T-parameter number trace mainly to a Musk X post cited in the research, with no SpaceXAI architecture paper explaining expert routing, active parameters, training mix, or serving behavior.
If Grok 4.5 is mixture-of-experts, total parameters and active inference parameters are different operational facts. Active parameters are closer to serving cost and latency. The headline parameter count is marketing-grade until SpaceXAI publishes the technical report.
Is Grok 4.5 better than Claude Opus?
Grok 4.5 is better for workloads where price and terminal-task completion dominate. The supplied research reports 83.3% on Terminal-Bench 2.1 versus 78.9% for Claude Opus 4.8, while Opus leads on reported SWE-Bench Pro and neutral DeepSWE evidence as of July 2026.
Musk described Grok 4.5 as “Opus-class, but faster, more token-efficient and lower cost,” according to TechCrunch's launch report. The lower-cost part has first-party support. The capability claim is narrower.
Genαi's research, drawing in part on a Grok 4.5 benchmark review, reports Grok 4.5 at 64.7% on SWE-Bench Pro versus 69.2% for Claude Opus 4.8. It reports the reverse on Terminal-Bench 2.1: 83.3% for Grok 4.5 versus 78.9% for Opus.
The neutral-harness signal is the pressure point. Genαi's research records Grok 4.5 at 53% on DeepSWE 1.1 and says Opus leads there. That makes a broad Opus-class verdict too loose for repository repair.
Artificial Analysis cuts the same direction. Grok 4.5 ranked #4 of 168 at score 54, behind Fable 5, GPT-5.5, and Claude Opus 4.8, according to the research set.
Fable 5 is the shadow benchmark in this story. Genαi's research says xAI's chart context showed Fable 5 ahead on SWE-Bench Pro, Terminal-Bench 2.1, and DeepSWE 1.0, with 80.4%, 84.3%, and 66.1% respectively.
GPT-5.6 Sol is harder to use as a comparator. The supplied research includes a content-marketing report on GPT-5.6 claiming 91.9% on Terminal-Bench 2.1, but that is not a primary OpenAI benchmark in the research package.
How much does Grok 4.5 cost?
Grok 4.5 costs $2 per million input tokens and $6 per million output tokens on the Grok API, according to SpaceXAI's official model docs as of July 10, 2026. Cached input is listed at $0.50 per million tokens, and Web/X search calls cost $5 per 1,000.
This is the launch's cleanest operational advantage. Price is first-party, current, and directly testable on your invoice.
The Opus pricing signal in this research is weaker. A secondary Claude Opus 4.8 launch guide reports $5/M input and $25/M output, but teams should verify procurement terms before using that spread in a production budget.
| Model or API | Price signal as of July 10, 2026 | Benchmark signal in research | Access or license evidence | Limit to watch | Best-fit use case |
|---|---|---|---|---|---|
| Grok 4.5 | $2/M input, $6/M output; $0.50/M cached input | 83.3% Terminal-Bench 2.1; 64.7% SWE-Bench Pro | API, Grok Build, Cursor, grok.com, X, iOS/Android in research | 500K context; no model card or system card | Cost-sensitive coding agents and terminal automation |
| Claude Opus 4.8 | Reported $5/M input, $25/M output via secondary source | 69.2% SWE-Bench Pro; stronger neutral DeepSWE evidence in research | Hosted Claude model access is inferred from product context; exact access terms absent from report | Pricing source is secondary in this research | High-trust repository repair bake-offs |
| GPT-5.6 Sol | Pricing not disclosed in supplied research | 91.9% Terminal-Bench 2.1 via content-marketing source | Public launch reported; exact API terms absent from report | No primary OpenAI benchmark or pricing source in report | Terminal-heavy trials after pricing clears |
| Gemini 3.5 Pro | Enterprise only; no public price in supplied research | No comparable score in report | Enterprise access noted in research | Public pricing gap | Enterprise evaluation tracks |
The output-price gap is large if the secondary Opus figure holds. Grok's $6/M output price is about 4.2x lower than the reported $25/M Opus output price.
That still doesn't make Grok cheaper in production. Retries, failed patches, CI failures, long review cycles, and factual cleanup can erase a token-price win.
Consumer pricing belongs in a separate lane. A Grok pricing roundup reports SuperGrok tiers from roughly $10 to $300 per month, but serious migration decisions should use API pricing, rate limits, and measured task acceptance.
What should you benchmark before switching?
Benchmark Grok 4.5 on accepted-task cost before any migration. Your test should count pass rate, retries, tool failures, CI failures, latency, review time, and human edits across repository repair, terminal automation, long-context recall, factuality, and tool-calling tasks using identical prompts.
Start with shadow traffic. Send the same representative tasks to Grok 4.5, your incumbent model, and one frontier alternative.
Use tasks recent enough that memorization is unlikely. Private repositories, current bugs, internal runbooks, and clean containers are better than public demo prompts.
Your benchmark should have five lanes:
- Repository repair: choose real issues, require tests to pass, and reject patches that need manual correction.
- Terminal automation: replay shell-heavy tasks and score command safety, recovery, and final state.
- Token efficiency: measure output tokens on identical tasks and verify the reported 4.2x efficiency claim yourself.
- Factuality: ask dated, source-verifiable questions and count unsupported claims, stale answers, and invented citations.
- Long context: feed 400K-token materials and test recall from the middle and end.
Set thresholds before reading the results. For coding-agent migration, a reasonable first pass is 50 to 200 tasks, at least 65% accepted fixes on your neutral repair set, terminal completion near the reported 80% range, and lower accepted-task cost than your incumbent.
Factual workloads need a harder gate. Genαi's research records a reported 54% hallucination rate for Grok 4.5 on AA-Omniscience versus 25% for its predecessor. If your measured rate stays above 30%, keep it behind retrieval, citations, and human review.
Cursor-specific results need special handling. The research says Cursor disclosed an earlier snapshot of its own codebase was accidentally included in Grok 4.5 training, and an xAI engineer described the data as supplemental training. Treat Cursor wins as suspect until they reproduce on your own code.
Safety documentation is also a procurement issue. Genαi's research records that SpaceXAI safety adviser Dan Hendrycks said dangerous-capability evaluations were performed but not publicly shared. As of July 10, 2026, SpaceXAI had not published a system card, model card, or full safety-evaluation report for Grok 4.5.
What this means for you
Grok 4.5 is worth testing if you spend real money on coding agents, CLI automation, long multi-turn implementation sessions, or Cursor workflows. Its API price is low enough that a second-place model can win the production bake-off.
Start where failure is recoverable: code suggestions, batch refactors behind CI, terminal copilots in disposable environments, and internal build tools with existing review loops.
Keep your current top model in place for regulated work, high-stakes factual answers, and repository repair where neutral benchmark strength matters more than the token bill. The missing system card is a governance gap, not a footnote.
Use this migration checklist:
- Build a 50 to 200 task eval set from current work, not public benchmark prompts.
- Run Grok 4.5, Claude Opus 4.8, and one other frontier model with identical tool access.
- Score accepted-task cost, including retries, human edits, and failed CI.
- Separate Cursor tasks from non-Cursor tasks.
- Require factual workflows to pass a hallucination and citation audit.
- Move one low-risk workflow first, then expand only after a week of production logs.
The decision rule is simple enough to write into your eval harness: ship Grok 4.5 where it lowers accepted-task cost without increasing worst-case failure review. Keep it in shadow mode everywhere else.