Обзор Grok 4.5

Уровень Opus · в 4 раза дешевле · бенчмарки · Cursor · API

Grok 4.5 xAI coding model review benchmarks and pricing

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

01

Что такое Grok 4.5? Контекст релиза и спецификации

Grok 4.5 is SpaceXAI's frontier model built for:

  • Coding and software engineering — bug fixes, large-scale refactors, end-to-end app building
  • Agentic tasks — multi-step automation across tools and enterprise apps
  • Knowledge-intensive work — legal, healthcare, education, data analysis

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.

SpecDetail
ArchitectureMixture of Experts (MoE)
Context window500,000 tokens
Reasoning modesLow / Medium / High (default: High)
Speed80 TPS official, ~90 TPS measured
Training infraTens of thousands of NVIDIA GB300 GPUs (Memphis, TN)
Parameter countNot disclosed

Teams evaluating a switch typically hit these pain points:

  1. 01

    Runaway API bills: Claude Fable 5 / Claude Code averages ~7.2M tokens per agentic task.

  2. 02

    Benchmark vs reality gap: Provider harnesses inflate scores; neutral harnesses tell a different story.

  3. 03

    First-shot accuracy vs retry cost: Precision work needs Opus reliability; high-volume loops need speed and unit economics.

  4. 04

    Training contamination: CursorBench was pulled after Cursor codebase snapshots leaked into training data.

  5. 05

    Regional limits: EU API expected mid-July; currently us-east-1 and us-west-2 only.

02

Цены: сколько реально можно сэкономить?

Sticker price is only half the story. Token efficiency compounds in agentic pipelines.

Token pricing vs the competition (per 1M tokens)

ModelInputOutput
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

Real-world cost per coding task

Model / platformAvg tokens per taskEst. 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.

03

Бенчмарки: кодинг, agentic, индекс интеллекта

Coding benchmarks

BenchmarkGrok 4.5Claude Fable 5Claude Opus 4.8GPT-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 2.183.3%84.3%78.9%83.4%
SWE-Bench Pro64.7%80.4%69.2%58.6%
  • Neutral DeepSWE: Grok 4.5 drops to 53%, trailing all three — Fable 5 leads by 17 points.
  • Terminal Bench 2.1: All four cluster within 5.4 points — cost and fit matter more.
  • SWE-Bench Pro: Grok 4.5 ranks third; the 15.7-point gap behind Fable 5 matters for complex multi-file work.

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.

Agentic benchmarks — where Grok 4.5 leads

BenchmarkGrok 4.5Claude Fable 5Claude 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%).

Overall intelligence ranking

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.

04

Реальные тесты, платформы и 6-шаговая настройка API

TryAI independent head-to-head

TestResult
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 costGrok 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.

Where you can use Grok 4.5 (EU expected mid-July)

  • Grok Build — native coding agent platform; Grok 4.5 is default
  • Cursor — all plans (desktop, web, iOS, CLI, SDK); doubled usage first week
  • SpaceXAI Console API — Chat Completions and Responses API; us-east-1 / us-west-2; 150 req/s, 50M tokens/min
  • Microsoft Office add-ins — default for Word, PowerPoint, Excel
  • Third-party gateways — OpenRouter, Vercel, Cloudflare, Snowflake, Databricks Mosaic
bash
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]}"
  }'

Six-step API runbook

  1. 01

    Create a SpaceXAI Console account at console.x.ai and generate an API key.

  2. 02

    Pick your path: direct API, Cursor model pool, or a gateway like OpenRouter.

  3. 03

    Enable caching: set prompt_cache_key (Responses API) or x-grok-conv-id header — cached input drops to $0.50/M.

  4. 04

    Turn on Context Compaction for long agent loops to cut token accumulation.

  5. 05

    Switch in Cursor: select Grok 4.5 and A/B test during the doubled first-week quota.

  6. 06

    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.

05

Стоит ли переходить? Матрица решений и цифры

When Grok 4.5 makes sense

  1. 01

    High-volume agentic pipelines — hundreds or thousands of coding tasks per day.

  2. 02

    Terminal and tool-use workflows — leads or ties on Terminal Bench and AutomationBench.

  3. 03

    Cursor-native teams — zero-friction integration.

  4. 04

    Cost-sensitive startups — ~4x less per task at comparable intelligence.

  5. 05

    Mixed-model strategies — Grok for routine work, Claude for the hard stuff.

When to think twice

ScenarioRiskMitigation
SWE-Bench Pro precision workFable 5 leads by 15+ pointsKeep Claude for finance/security-critical code
Hallucination-sensitive systems54% on AA-OmniscienceValidation layers and human review
EU teamsNo EU API yetWait for mid-July or use approved gateways
CursorBench-related claimsTraining contaminationWait for independent retests

Citable hard data

  • Context window: 500,000 tokens
  • Token efficiency: 15,954 vs 67,020 output tokens on SWE-Bench Pro (4.2x)
  • Per-task cost: ~$2.49 vs ~$11.80 for Claude Code
  • Intelligence index: 54/100, +16 vs prior Grok
  • Throughput: first token <500ms, ~110 tokens/sec (~2x peers)

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

FAQ

Частые вопросы

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.