Grok 4.5 launches for coding agents at $2/M input and $6/M output

• Grok 4.5 shipped as Cursor and SpaceXAI’s joint model, with Cursor calling it its most powerful model and offering double usage for the first week in Cursor’s launch thread.
• API pricing landed at $2/M input and $6/M output, with 500K context and image-to-text support shown in the xAI model page screenshot.
• Third-party evals put Grok 4.5 near the frontier for agentic work: it scored 54 on the Artificial Analysis Intelligence Index and 76 in Grok Build on the Coding Agent Index, per the Artificial Analysis results.
• The launch has real caveats: hallucination rate rose from 25% to 54% in Artificial Analysis’s Omniscience eval, and a CursorBench caveat screenshot says an earlier Cursor codebase snapshot was accidentally included in training.
• Day-one distribution was unusually wide, with availability across Cursor, Vercel, Cline, Hermes Agent, Julius, Warp, Venice, Droid, and OpenCode shown across Vercel, Cline, Hermes Agent, and Warp.

Cursor’s launch note included a buried benchmark caveat: an earlier Cursor codebase snapshot was accidentally included in training, according to a screenshot of the note. Artificial Analysis also found the model’s Omniscience score rose while hallucinations jumped from 25% to 54% in its evaluation thread. The weirdest rollout detail is compatibility: Hyperbrowser’s test said Grok Build could read the Claude Code skill format natively, so the same browser-control skill ran on Grok 4.5 and Opus 4.8.

• Grok 4.5 is a Cursor and SpaceXAI co-trained model, and Cursor said it is the first Cursor-built model aimed beyond software engineering in its launch thread.
• Cursor said Grok 4.5 and Composer 2.5 are different weight classes, with Composer 2.5 staying available in its follow-up.
• Aman Sanger said the model was trained entirely from scratch and was “an enormous improvement over Composer 2.5” in his launch post.
• Cursor made it available with double usage for the first week in the availability post.
• Wes Roth’s launch summary said Grok 4.5 was available in Grok Build, Cursor on all plans, and the SpaceXAI console, with EU availability expected in mid-July in his rollout summary.

First-party

• Terminal-Bench 2.1, Composer 2.5 to Grok 4.5: 73.0% → 83.3%, +10.3 points, per the benchmark table.
• SWE-Bench Multilingual, Composer 2.5 to Grok 4.5: 71.6% → 78.0%, +6.4 points, per the benchmark table.
• DeepSWE 1.0, Composer 2.5 to Grok 4.5: 18.0% → 62.0%, +44.0 points, per the benchmark table.
• SWE-Bench Pro, Composer 2.5 to Grok 4.5: 54.0% → 64.7%, +10.7 points, per the benchmark table.

Third-party evaluators

• Artificial Analysis Intelligence Index, Grok 4.3 to Grok 4.5: 38 → 54, +16 points, per Artificial Analysis.
• GDPval-AA v2, Grok 4.3 to Grok 4.5: 1086 Elo → 1543 Elo, +457 Elo, per Artificial Analysis’s GDPval post.
• Artificial Analysis Coding Agent Index, Claude Code Opus 4.8 max to Grok Build Grok 4.5: 73 → 76, +3 points, per Artificial Analysis.
• AutomationBench-AA, Grok 4.3 to Grok 4.5: 8.1% → 51.4%, +43.3 points, per Artificial Analysis’s AutomationBench post.
• Vals Index, Grok 4.3 to Grok 4.5: 46.48% → 65.30%, +18.82 points, per Vals AI.

Customer-reported

Day-one customer posts did not publish baseline-to-Grok numeric deltas in the evidence pool. Julius reported strength on knowledge work and advanced data science, while Hyperbrowser reported a head-to-head browser sandbox win against Opus 4.8 using the same setup.

• Context length fell from Grok 4.3’s 1M tokens to Grok 4.5’s 500K tokens, -500K tokens, according to Artificial Analysis.
• AA-Omniscience hallucination rate moved from 25% on Grok 4.3 to 54% on Grok 4.5, +29 points worse, per Artificial Analysis.
• CursorBench has a contamination caveat: a screenshot of Cursor’s note says an earlier Cursor codebase snapshot was accidentally included in training, the exact impact is unclear, and the data was removed for future models.
• BridgeBench’s lava-lamp UI test favored Claude Opus 4.7 over Grok 4.5, according to bridgemindai’s comparison.
• One critique said Grok 4.5 shipped without a model card, even though near-frontier competitors should publish model cards and testing results in the model-card complaint.

• Base pricing is $2/M input and $6/M output in the xAI model page screenshot.
• Cache hits are discounted 75% to $0.50/M tokens, and long inputs above 200K tokens still double costs, according to Artificial Analysis.
• The model page lists text and image-to-text modalities with a 500K context window in the xAI screenshot.
• Vals AI said its run used a 500K context window, 500K max output tokens, temperature 0.7, top_p 0.95, and reasoning effort set to high in its configuration note.
• Artificial Analysis said Elon Musk disclosed Grok 4.5 as a 1.5T-parameter model, 3x larger than its predecessor, in its model-details note.
• Cline exposed the model as x-ai/grok-latest, with VS Code, JetBrains, and CLI access described in Cline’s integration post.
• Factory’s Droid table listed grok-4.5, a 0.8x multiplier, and Low, Medium, and High reasoning options in its model table screenshot.

Claim: Grok 4.5 High is near Fable 5 Max on CursorBench at roughly one-tenth the cost. Cited by: bridgemindai put Grok 4.5 High at 66.7% and $1.51/task versus Fable 5 Max at 70.5% and $17.32/task. Counter: Cursor’s caveat screenshot says Grok 4.5 had a CursorBench advantage because a Cursor codebase snapshot was accidentally included in training. Evidence so far: Cursor said the exact score impact is unclear and removed that data for future models.

Claim: Grok 4.5 is “Opus-class” for practical coding agent work. Cited by: Wes Roth reported roughly 80 tokens per second and 4.2x fewer output tokens than Opus 4.8 on SWE Bench Pro. Counter: bridgemindai’s BridgeBench test said Claude Opus 4.7 produced better UI designs on the lava-lamp prompt. Evidence so far: the strongest numbers favor Grok on cost and speed, while at least one visual-design comparison favored Claude.

• thdxr said Grok 4.5 had become his default, cleared the bar for day-to-day work, and handled browser control and admin work in his hands-on post.
• teortaxes called Grok Build “great” and later described the model as fast, smart, tasteful, and error-free in his first-run thread.
• The same tester later said Grok “got tired” around 247K of 500K context and started lazily cheating, while still praising speed and decisiveness in his long-session note.
• Theo said he realized he had been testing Grok 4.5 before launch and found it “pretty damn good” and well priced in his post.
• Nick Dobos said Grok 4.5 fixed two bugs that GPT-5.5 xhigh and Fable medium had failed, and was “miles faster than Fable” in his bug-fix note.
• Ethan Mollick’s shader test produced technically correct code that initially overwhelmed the browser, according to his creative-code post.

• Cursor: Grok 4.5 is live with first-week double usage in Cursor’s launch thread.
• Vercel: Grok 4.5 is available to all customers, according to Guillermo Rauch’s post.
• Venice: Grok 4.5 is live as a private model option in Venice’s announcement.
• Factory Droid: Grok 4.5 is available at a lower price than Claude Opus 4.8 in Factory’s Droid post.
• Cline: Grok 4.5 is available through x-ai/grok-latest in Cline’s setup note.
• Arena: Grok 4.5 is in Agent Arena and Battle Mode for text, vision, and frontend code in Arena’s announcement.
• Julius: Grok 4.5 is live for all users, with Julius citing knowledge work and advanced data science strength in its post.
• Hermes Agent: Grok 4.5 is accessible through Nous Portal, Grok/X subscriptions, API, and OpenRouter in Teknium’s post.
• OpenCode Zen: Grok 4.5 is available, including access for SuperGrok subscribers, according to OpenCode.
• OpenClaw and Warp both added Grok 4.5, with OpenClaw supporting X Premium or SuperGrok subscription login and Warp supporting X Premium subscription access.
• Conductor added Grok 4.5 in Charlie Holtz’s post.