Уровень Opus · в 4 раза дешевле · бенчмарки · Cursor · API
8 июля 2026 года SpaceXAI выпустила Grok 4.5 — первую флагманскую модель после IPO. Илон Маск назвал её «интеллектом уровня Opus за долю стоимости». Обзор включает все бенчмарки, таблицы цен, тесты TryAI и 6-шаговый API runbook. Данные на 10 июля 2026
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 a real 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 |
Teams evaluating a switch typically hit these pain points:
Runaway API bills: Claude Fable 5 / Claude Code averages ~7.2M tokens per agentic task.
Benchmark vs reality gap: Provider harnesses inflate scores; neutral harnesses tell a different story.
First-shot accuracy vs retry cost: Precision work needs Opus reliability; high-volume loops need speed and unit economics.
Training contamination: CursorBench was pulled after Cursor codebase snapshots leaked into training data.
Regional limits: EU API expected mid-July; currently us-east-1 and us-west-2 only.
Sticker price is only half the story. Token efficiency compounds in agentic pipelines.
| Model | Input | Output |
|---|---|---|
| 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 4.2x efficiency gap. At 500 tasks/day, that is roughly $1,245/day vs $5,900/day.
| Benchmark | Grok 4.5 | Claude Fable 5 | Claude 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% |
CursorBench pulled: A snapshot of Cursor's own codebase was accidentally included in Grok 4.5 training data — a clear contamination issue. Launch materials were revised accordingly.
| Benchmark | Grok 4.5 | Claude Fable 5 | Claude Opus 4.8 |
|---|---|---|---|
| AutomationBench-AA (657 enterprise workflows) | 51.4% | 48.6% | 48.5% |
| Snorkel GDPVal+ | 29% | — | 21% |
AutomationBench-AA covers 40 simulated enterprise apps including Gmail, Slack, Salesforce, and HubSpot. Grok 4.5 is the first model to complete more than half of all workflow objectives without violating business constraints. 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 — fourth overall, behind Fable 5 (60), Opus 4.8 (56), and GPT-5.5 (55). Still a 16-point jump over the previous Grok generation.
| Test | Result |
|---|---|
| 3D cube rendering (hardest) | Opus 4.8 and Fable 5 correct first try; Grok 4.5 missed cube on attempt one, fixed on retry; GPT-5.5 failed |
| Speed and cost | Grok 4.5 first token under 500ms, ~110 tokens/sec (~2x competitors); cheapest per run |
Bottom line: one-shot complex UI favors Claude. High-volume repetitive codegen favors Grok 4.5. See also our AI coding assistants comparison.
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]}"
}'
Create a SpaceXAI Console account at console.x.ai and generate an API key.
Pick your path: direct API, Cursor model pool, or a gateway like OpenRouter.
Enable caching: set prompt_cache_key (Responses API) or x-grok-conv-id header — cached input drops to $0.50/M.
Turn on Context Compaction for long agent loops to cut token accumulation.
Switch in Cursor: select Grok 4.5 and A/B test during the doubled first-week quota.
Build a mixed-model router: routine subtasks to Grok 4.5; architecture and precision refactors to Claude Fable 5.
Production note: Independent evaluators report Grok 4.5 hallucination rate on the AA-Omniscience Index at 54% — significantly higher than prior models. Build robust output validation.
High-volume agentic pipelines — hundreds or thousands of coding tasks per day.
Terminal and tool-use workflows — leads or ties on Terminal Bench and AutomationBench.
Cursor-native teams — zero-friction integration.
Cost-sensitive startups — ~4x less per task at comparable intelligence.
Mixed-model strategies — Grok for routine work, Claude for the hard stuff.
| Scenario | Risk | Mitigation |
|---|---|---|
| SWE-Bench Pro precision work | Fable 5 leads by 15+ points | Keep Claude for finance/security-critical code |
| Hallucination-sensitive systems | 54% on AA-Omniscience | Validation layers and human review |
| EU teams | No EU API yet | Wait for mid-July or use approved gateways |
| CursorBench-related claims | Training contamination | Wait for independent retests |
Grok 4.5 delivers 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.
Облачные API снижают стоимость инференса, но цепочки подписи iOS, сборки Xcode, Metal и CI 24/7 требуют реального macOS-железа. Для продакшн iOS CI/CD и автоматизации агентов аренда Mac Mini в облаке VpsMesh — обычно лучший выбор.
Sources: SpaceXAI announcement · Cursor launch post · API docs · Snorkel AI
Depends on the metric. Opus 4.8 wins SWE-Bench Pro accuracy (69.2% vs 64.7%). Grok 4.5 wins speed, token efficiency, and per-task cost — often by 4x. On agentic workflow completion it edges Opus on independent benchmarks.
Limited free usage in Grok Build and Cursor for a limited time. After that: $2/M input, $6/M output via API. Cursor plans include it in the model pool.
All Cursor plans. Open model picker and select Grok 4.5. First-week usage was doubled. For 24/7 build nodes, see Mac Mini M4 rental pricing.
500,000 tokens (500K) — enough for most large codebase tasks.
Cursor codebase snapshots were accidentally included in training data, contaminating that benchmark. Results were pulled; independent retesting is expected.
Да — OpenRouter, Vercel, Cloudflare, Snowflake, Databricks Mosaic. Подробности в центре помощи и ценах аренды Mac Mini M4.