Grok 4.5: характеристики флагманской модели SpaceXAI для кода
Grok 4.5 — первый major release SpaceXAI после IPO. Модель co-trained с Cursor после acquisition Anysphere SpaceX в июне 2026: trillions of tokens из реальных IDE-сессий — code review, debugging, agent-to-репозиторий interactions.
Оптимизирован под:
Coding agents: bugfixes, large refactors, end-to-end app builds.
Agentic automation: multi-step workflows across enterprise apps.
Knowledge work: legal, healthcare, education, data analysis.
High-volume cost pressure: сотни agent calls в день — API bill чувствуется сразу.
First-shot accuracy: complex stateful UI tasks наказывают модели с retry loops.
| Spec | Значение |
|---|---|
| Architecture | Mixture of Experts (MoE) |
| Context window | 500 000 tokens |
| Reasoning modes | Low / Medium / High (default: High) |
| Throughput | 80 TPS official, ~90 TPS measured |
| Training infra | Tens of thousands NVIDIA GB300 (Memphis, TN) |
| Parameters | Not disclosed |
Grok 4.5 pricing: сколько реально экономите
List price — полдела. В high-frequency agent scenarios драйвер cost — token efficiency, не $/M на бумаге.
| 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 |
| Claude Fable 5 | Выше | Выше |
| GPT-5.6 Sol (flagship) | $5.00 | $30.00 |
| GPT-5.6 Luna (economy) | $1.00 | $6.00 |
Cost per agent task (SWE-Bench Pro conversion):
| Platform | Avg tokens/task | 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 |
SWE-Bench Pro: Grok 4.5 averaged 15 954 output tokens/task vs 67 020 у Claude Opus 4.8 — 4.2× efficiency gap. At 500 tasks/day: ~$1 245/day vs ~$5 900/day.
Benchmark matrix: coding vs agentic workloads
3.1 Coding benchmarks
| Benchmark | Grok 4.5 | Fable 5 | Opus 4.8 | GPT-5.5 |
|---|---|---|---|---|
| DeepSWE 1.0 (vendor 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% |
Read: DeepSWE 1.0 с vendor harness — Grok 4.5 rank 3. Neutral harness 1.1 — Fable 5 +17pp. Terminal Bench 2.1 — все top-4 в пределах 5.4pp, почти tie. SWE-Bench Pro — hardest test, Grok 4.5 rank 3, ~16pp behind Fable 5.
Caveat: CursorBench pulled — snapshot репозитория Cursor попал в training data. Transparency problem launch-а; ждём independent retest.
3.2 Agent benchmarks — sweet spot Grok 4.5
| Agent benchmark | Grok 4.5 | Fable 5 | Opus 4.8 |
|---|---|---|---|
| AutomationBench-AA (657 tasks) | 51.4% | 48.6% | 48.5% |
| Snorkel GDPVal+ | 29% | — | 21% |
AutomationBench-AA: 40 simulated enterprise apps (Gmail, Slack, Salesforce, HubSpot). Grok 4.5 — первый model с >50% workflow objectives без violation business constraints. Snorkel: legal 40% vs 27–28%, education 58% vs 35–42%, healthcare 35% vs 23–25%.
3.3 Intelligence index: Artificial Analysis 54/100 (4th place), +16pp vs previous Grok. Order: Fable 5 (60), Opus 4.8 (56), GPT-5.5 (55).
3.4 TryAI real coding tests (independent)
3D cube render (hardest): Opus 4.8 и Fable 5 first try pass; Grok 4.5 retry; GPT-5.5 fail.
Latency: first token <500ms, ~110 tokens/sec — ~2× competitors.
Cost/run: Grok 4.5 lowest; Fable 5 slowest и priciest.
Bottom line: high-frequency repetitive coding — Grok 4.5 доминирует по speed/cost. Complex stateful UI one-shot — Claude series надёжнее.
Как подключить Grok 4.5: Cursor, API, six-step runbook
Live на: Grok Build (default model), все Cursor plans (desktop/web/iOS/CLI/SDK), SpaceXAI Console API, Office add-ins, gateways (OpenRouter, Vercel, Cloudflare, Snowflake, Databricks Mosaic). Regions: us-east-1, us-west-2. Rate limits: 150 req/s, 50M tokens/min. EU API — mid-July 2026.
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": "Найди и исправь баг: function median(a){a.sort();return a[a.length/2]}"
}'
API key: SpaceXAI Console, confirm US region access (us-east-1 / us-west-2).
Cursor model picker: select Grok 4.5; launch week — doubled usage.
Cache key: prompt_cache_key (Responses API) или header x-grok-conv-id — cached input $0.50/M.
Context Compaction: enable для long agent loops, cut token accumulation.
Hybrid routing: routine subtasks → Grok 4.5; architecture decisions → Claude Fable 5.
Production validation: output checks на finance/security-critical code; AA-Omniscience hallucination rate 54% — human review loop обязателен.
Best practice: always set conversation cache key. EU users — не деплойте production pipeline до EU API GA (mid-July).
Switch или нет: verdict и hard numbers
Grok 4.5 fit: high-volume agentic pipelines, terminal-heavy workflows, Cursor-native teams, cost-sensitive startups, mixed-model routing (Grok на repetitive subtasks, Fable 5 на architecture).
Think twice: SWE-Bench Pro-grade precision, hallucination-sensitive prod, EU deploy до mid-July, CursorBench claims без retest.
Per-task cost: ~$2.49 (Grok) vs ~$11.80 (Claude Code).
Output token efficiency: 15 954 vs 67 020 (4.2× gap vs Opus 4.8).
Throughput: 80 TPS official, 90–110 tokens/sec measured, first token <500ms.
Grok 4.5 — не самый accurate coding model (crown у Claude Fable 5). Но accuracy per benchmark ≠ value per dollar. API-only — ок для trials; local Mac under long agent loops ловит memory pressure и swap; VM жертвует Metal/Xcode fidelity. Для 7×24 Cursor agents, parallel dev servers, iOS CI/CD — MESHLAUNCH cloud Mac Mini bare-metal rental обычно лучший production fit: dedicated Apple Silicon, flexible daily/weekly/monthly terms.
Sources: x.ai · Cursor · API docs · TechCrunch · Snorkel AI
Depends on metric. Opus wins SWE-Bench Pro accuracy; Grok wins speed, token efficiency, per-task cost — often 4×. Budget hardware: цены аренды.
Limited free в Grok Build и Cursor at launch. Ongoing API: $2/M input, $6/M output.
Model selector → Grok 4.5 на любом Cursor plan. Deploy questions: центр помощи.
500 000 tokens — enough для большинства задач на крупном репозитории.
Training data contamination: snapshot репозитория Cursor invalidates numbers.
Да — также Vercel AI Gateway, Cloudflare, Snowflake, Databricks Mosaic.