Launched July 8, 2026 · Cursor co-training · Full benchmark tables · API pricing · TryAI hands-on tests · Switch decision matrix
Summary: On July 8, 2026, Elon Musk's SpaceXAI shipped Grok 4.5 — its first flagship model since going public. Musk called it "Opus-class intelligence at a fraction of the cost." After digging through every published benchmark, independent evaluation, and real-world coding test we could find, here's the unfiltered verdict — plus when it beats Claude on dollars, and when it doesn't on accuracy.
Grok 4.5 is SpaceXAI's frontier model built for:
The model was co-trained with Cursor. SpaceX acquired Cursor parent Anysphere in June 2026; training included trillions of tokens of real developer interaction data — how developers write, review, and debug inside an IDE, and how agents interact with live codebases.
| Spec | Detail |
|---|---|
| Architecture | Mixture of Experts (MoE) |
| Context window | 500,000 tokens |
| Reasoning modes | Low / Medium / High (default: High) |
| Speed | 80 TPS official, ~90 TPS measured |
| Training infra | Tens of thousands of NVIDIA GB300 GPUs (Memphis, TN) |
| Parameter count | Not disclosed |
Runaway agent bills — Claude Code and Codex costs compound at high volume
Best benchmark ≠ cheapest production — leaderboard winners can be the priciest daily driver
Cursor-native default — teams already in Cursor need a credible Opus alternative
Trust gaps — CursorBench was pulled over training-data contamination
| Model | Input (per 1M) | Output (per 1M) |
|---|---|---|
| Grok 4.5 | $2.00 | $6.00 |
| Grok 4.5 (cached input) | $0.50 | — |
| Grok 4.5 Fast | $4.00 | $18.00 |
| Claude Opus 4.7 | $5.00 | $25.00 |
| GPT-5.6 Sol | $5.00 | $30.00 |
| GPT-5.6 Luna | $1.00 | $6.00 |
| Model / platform | Avg tokens per task | Est. cost per task |
|---|---|---|
| Grok 4.5 / Grok Build | ~1.9M | $2.49 |
| GPT-5.5 / Codex | ~6.2M | $5.07 |
| Claude Fable 5 / Claude Code | ~7.2M | $11.80 |
On SWE-Bench Pro, Grok 4.5 averaged 15,954 output tokens per task. Claude Opus 4.8 used 67,020 — a 4.2× efficiency gap. At 500 tasks/day, that's roughly $1,245/day vs. $5,900/day.
prompt_cache_key or x-grok-conv-idus-east-1, us-west-2 (EU expected mid-July)| Benchmark | Grok 4.5 | Fable 5 | Opus 4.8 | GPT-5.5 |
|---|---|---|---|---|
| DeepSWE 1.0 (provider harness) | 62.0% | 66.1% | 55.75% | 64.31% |
| DeepSWE 1.1 (neutral harness) | 53% | 70% | 59% | 67% |
| Terminal Bench 2.1 | 83.3% | 84.3% | 78.9% | 83.4% |
| SWE-Bench Pro | 64.7% | 80.4% | 69.2% | 58.6% |
Caveat: CursorBench was pulled after Cursor codebase snapshots accidentally entered Grok 4.5 training data — a transparency problem with the launch.
| Benchmark | Grok 4.5 | Fable 5 | Opus 4.8 |
|---|---|---|---|
| AutomationBench-AA (657 workflows) | 51.4% | 48.6% | 48.5% |
| Snorkel GDPVal+ | 29% | — | 21% |
Grok 4.5 is the first model to complete more than half of enterprise workflow objectives without violating business constraints (Gmail, Slack, Salesforce, HubSpot, and 36 more simulated apps). Snorkel shows wide leads in legal (40% vs 27–28%), education (58% vs 35–42%), and healthcare (35% vs 23–25%).
Artificial Analysis Intelligence Index: 54/100 (4th), behind Fable 5 (60), Opus 4.8 (56), GPT-5.5 (55) — still a +16 jump over the previous Grok generation.
3D cube rendering (hardest test): Opus 4.8 and Fable 5 succeeded on the first try. Grok 4.5 rendered title and buttons but no cube on attempt one, then fixed on retry. GPT-5.5 failed.
Speed: Grok 4.5 delivered first token in under 500ms and streamed at ~110 tokens/second — roughly twice as fast as competitors.
Bottom line: High-volume, repetitive codegen favors Grok 4.5. Complex stateful UI that must be right the first time still favors Claude.
Create an API key at console.x.ai
Pick us-east-1 or us-west-2
Call Responses API with model: "grok-4.5"
Set prompt_cache_key for cache hits ($0.50/M input)
Enable Context Compaction on long agent loops
curl -s https://api.x.ai/v1/responses \
-H "Authorization: Bearer $XAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"grok-4.5","input":"Find and fix the bug: function median(a){a.sort();return a[a.length/2]}"}'| Scenario | Pick | Why |
|---|---|---|
| Hundreds–thousands of agent tasks/day | Grok 4.5 | ~$2.49 vs $11.80 per task |
| Terminal / tool-use heavy | Grok 4.5 | Leads or ties Terminal Bench & AutomationBench |
| Cursor-native teams | Grok 4.5 | Zero-friction integration |
| SWE-Bench Pro precision refactors | Claude Fable 5 | ~16-point lead |
| Hallucination-sensitive production | Claude + validation | Grok AA-Omniscience hallucination rate 54% |
| Mixed strategy | Grok subtasks + Claude architecture | Common enterprise pattern |
Grok 4.5 is not the most accurate coding model in mid-2026 — Claude Fable 5 holds that crown. What it delivers is the best intelligence-per-dollar ratio for agentic coding work available today. At $2.49 per real-world task versus $11.80 for Claude Code, the cost argument is arithmetic, not marketing.
Don't trust it blindly on your first production deploy. Validate outputs, watch hallucination rates, and keep a Claude model on standby for the hard stuff.
Data current as of July 10, 2026. Verify official docs before purchasing decisions.
Depends on the metric. Opus wins SWE-Bench Pro accuracy; Grok wins speed, token efficiency, and per-task cost — often by 4× — plus agentic workflow completion on independent benchmarks.
Limited free usage in Grok Build and Cursor for a limited time. API is $2/M input, $6/M output. Cursor plans include it in the pool.
All Cursor plans. Model picker → Grok 4.5. Doubled usage the first launch week.
500,000 tokens — enough for most large codebase tasks.
Cursor codebase snapshots contaminated training data for that benchmark. Results pulled; independent re-testing expected.
Yes — also Vercel, Cloudflare, Snowflake, and Databricks Mosaic.
Grok 4.5 makes Opus-class agent work affordable — especially inside Cursor. If your daily driver is Windows or Linux but you need a real macOS GUI to validate Cursor + Grok, handle Keychain prompts, or run iOS builds alongside agent loops, buying hardware is expensive and SSH alone won't click system dialogs. VNCMac rents physical Mac mini nodes by the hour with full VNC desktops — switch models, run agents, and sign off on the benchmarks in this guide, then stop renting when the sprint ends. See Mac rental plans.