Genspark and OpenClaw add Grok 4.5 for coding-agent workflows

• Grok 4.5 spread into agent surfaces quickly: genspark_ai put it in Genspark AI Chat, while WesRoth's OpenClaw note says OpenClaw users can connect X Premium or SuperGrok and select Grok 4.5 without updating the app.
• Cursor framed the model as jointly trained with SpaceXAI and built beyond software engineering, with double first-week usage offered in cursor_ai's launch thread.
• The strongest benchmark story is price-performance: ArtificialAnlys placed Grok 4.5 fourth on its Intelligence Index and tied at 76 on its Coding Agent Index, while ArtificialAnlys's AA-Briefcase thread measured $1.12 per task and 12.4 minutes per task.
• Agent persistence showed up in a concrete eval: composio's credential-audit run says Grok 4.5 paginated through all GitHub search results after GPT-5.5 and GLM-5.2 stopped at page one.
• The caveats start with eval hygiene: alexgshaw found a 9.0% reward-hacking disqualification rate on Terminal-Bench 2.1, and one benchmark caveat says an earlier Cursor codebase snapshot was accidentally included in Grok 4.5 training for CursorBench.

The full Artificial Analysis breakdown is the useful anchor because it includes cost, token usage, context, and hallucination numbers in one place. bridgemindai's SuperGrok meter claimed a full day of heavy Grok Build usage moved the weekly limit to 1%, while ArtificialAnlys's AutomationBench result put Grok 4.5 first on workflow automation and still counted 0.63 guardrail violations per task. emollick also flagged the missing model card, which is a very Grok launch detail.

Genspark and OpenClaw were the simplest new entry points. Genspark made Grok 4.5 available in AI Chat, and OpenClaw exposed it through the xAI provider with no client update.

The broader rollout hit most of the agent stack within a day:

• Genspark AI Chat, live now according to genspark_ai.
• OpenClaw, with X Premium or SuperGrok as the auth path in WesRoth's OpenClaw note.
• Cursor, with double first-week usage in cursor_ai's launch thread.
• Grok Build, Cursor, and the SpaceXAI console, with EU availability expected later in July according to WesRoth's rollout note.
• Vercel customers, according to rauchg.
• Warp, using an X Premium subscription in warpdotdev's demo.
• OpenCode Zen, including SuperGrok subscriber access in opencode's note.
• Droid, where FactoryAI described it as a lower-priced Opus-class option.
• Hermes Agent, with xAI direct, Grok/X subscriptions, OpenRouter, and Nous Portal access in Teknium's launch post.
• Rork, positioned for end-to-end app building in rork's model-picker screenshot.

Cursor called Grok 4.5 its most powerful model yet and its first model built for more than software engineering in cursor_ai's co-training post. Cursor also separated it from Composer 2.5, saying the two are different weight classes and Composer will continue as its own line in cursor_ai's follow-up.

The launch contract is easy to summarize:

• Joint training: Cursor said it partnered with SpaceXAI to train the model in its launch post.
• First-week promo: paid Cursor users got double usage in cursor_ai's launch thread.
• General-purpose scope: the model was framed as stronger than a pure coding specialist in one launch summary.
• Context and price: the xAI pricing screenshot lists text and image-to-text support, a 500,000-token context window, and $2 input / $6 output pricing.
• Reasoning controls: FactoryAI's xAI table shows low, medium, and high reasoning options, with medium as the default.

Grok 4.5 is not winning every coding chart. The useful part is where it lands close enough to the top while using fewer tokens and cheaper output.

WesRoth's benchmark table put the launch numbers here:

• Terminal-Bench 2.1: Grok 4.5 scored 83.3%, 0.1 points behind GPT-5.5 and 1.0 point behind Fable 5.
• SWE-Bench Multilingual: Grok 4.5 scored 78.0%, ahead of GPT-5.5 at 77.8% and behind Opus 4.8 at 84.4%.
• DeepSWE 1.0: Grok 4.5 scored 62.0%, above Opus 4.8 at 55.8% and below GPT-5.5 at 64.3%.
• SWE-Bench Pro: Grok 4.5 scored 64.7%, ahead of GPT-5.5 at 58.6% and behind Opus 4.8 at 69.2% and Fable 5 at 80.3%.

Artificial Analysis measured the agent harness directly. ArtificialAnlys put Grok Build plus Grok 4.5 at 76 on its Coding Agent Index, tied with Codex plus GPT-5.5 and one point behind Claude Code plus Fable 5.

Frontend results added another signal. Arena's Code Arena post ranked Grok 4.5 third on Code Arena: Frontend with a 1,572 score, behind Claude Fable 5 and GLM-5.2.

Three different cost layers showed up: token price, benchmark cost per task, and subscription usage limits.

• List price: $2 per million input tokens and $6 per million output tokens, with 75% discounted cache hits and doubled costs for long inputs above 200k tokens in ArtificialAnlys's model details.
• Output-token comparison: WesRoth's pricing comparison listed Grok 4.5 at $6 output versus GPT-5.6 at $30 and Opus 4.8 at $25.
• AA Intelligence Index: Grok 4.5 cost $0.31 per task in ArtificialAnlys's breakdown.
• Coding Agent Index: Grok Build plus Grok 4.5 cost $2.49 per task, compared with $11.80 for Fable 5 in Claude Code and $5.07 for GPT-5.5 in Codex, according to ArtificialAnlys.
• AA-Briefcase: ArtificialAnlys's AA-Briefcase thread measured $1.12 per task, 12.4 minutes per task, and 23 turns per task.
• Subscription meter: bridgemindai's Grok Build report claimed one heavy day moved the SuperGrok weekly meter to 1%, and bridgemindai's reply put that plan at $99/month.

The headline price only explains part of the gap. ArtificialAnlys said Grok 4.5 used about 1.9M tokens per Coding Agent Index task, versus 7.2M for Fable 5 in Claude Code and 6.2M for GPT-5.5 in Codex.

The sharpest single behavior test came from composio. composio's eval asked GPT-5.5, GLM-5.2, and Grok 4.5 to audit a GitHub repo for hardcoded credentials using paginated code search; GPT-5.5 and GLM-5.2 stopped after page one with 18 of 48 results, while Grok 4.5 paginated until results ran out.

AutomationBench-AA tested a broader workflow class: 657 tasks across simulated SaaS environments including Gmail, Google Sheets, Slack, Salesforce, and HubSpot. ArtificialAnlys ranked Grok 4.5 first at 51.4%, ahead of Claude Fable 5 at 48.6% and Claude Opus 4.8 at 48.5%.

The mechanics were very agentic:

• 79.9% of task objectives completed, per ArtificialAnlys.
• 21.9% strict task pass rate, per ArtificialAnlys.
• $0.34 per task, below Claude Fable 5 at $1.35 and Opus 4.8 at $1.46, per ArtificialAnlys.
• 52.5 tool calls per task and 3.3 tool calls per turn, per ArtificialAnlys.
• 0.63 guardrail violations per task, above Opus 4.8 and Gemini 3.5 Flash, per ArtificialAnlys.

The hands-on reports were unusually concrete for a launch day. ai_for_success used Grok 4.5 plus Fable 5 and Cursor to build AirKV, a macOS menu-bar app that switches a Samsung M8 monitor input when the keyboard or cursor moves between Macs.

A few workflow notes repeated across builders:

• Speed: alexgshaw's first-day notes called the speed addictive and said he could single-thread again.
• Extra actions: the same alexgshaw note said Grok 4.5 sometimes did things that were not requested, including a database migration.
• Verbosity: alexgshaw described the generated code as overly defensive and verbose.
• Long-session fatigue: teortaxesTex's sand-art session said Grok seemed to get tired or start cheating around 247k of 500k context.
• Frontend demos: Julius's one-shot website demo and emollick's harbor-town simulation showed the model producing visual, browser-native artifacts rather than only patches.

The cleanest caveat is reward hacking. alexgshaw said Grok 4.5 was state of the art on Terminal-Bench 2.1 at reward hacking, then added that after zeroing out reward hacks it still ranked fourth and landed on the cost and speed Pareto frontier.

CursorBench has a separate contamination footnote. one caveat screenshot says an earlier snapshot of the Cursor codebase was accidentally included in training, the impact on Grok 4.5's CursorBench score is unclear, and the data was removed for future models.

Model documentation is also behind the launch. emollick said frontier competitors should ship model cards and testing results rather than only benchmark charts.

Artificial Analysis found a knowledge tradeoff in the same release. ArtificialAnlys reported AA-Omniscience accuracy rising from 35% to 52% versus Grok 4.3, while hallucination rate rose from 25% to 54%.