Grok 4.5: обзор coding-agent SpaceXAI vs Claude Opus vs GPT-5.5

~14 мин чтения · MACCOME · Обновлено: 11 июля 2026

Для кого: тимлиды, считающие cost of switch, подписчики Cursor и тех, кто смотрит на token burn и API bill. 8 июля 2026 SpaceXAI выкатил Grok 4.5 — первый flagship после IPO; Musk называет это «Opus-class за четверть цены». Внутри: specs, API/per-task pricing, coding benchmarks, agent scores, TryAI, platforms, cache tips, fit/caution matrix, 6-step runbook, 6 FAQ. Широкий сравнительный контекст: decision matrix AI coding assistants.

bolt

TL;DR — 30 секунд

  • Не #1 в benchmarks, но лучший price/perf в Opus-tier: SWE-Bench Pro #3 (64.7%), per-task ~$2.49 vs Claude Code ~$11.80.
  • Token efficiency — скрытый OP: 15 954 output tokens vs Opus 4.8 67 020 — 4.2×.
  • Agent workflows — home turf: AutomationBench-AA 51.4%, первый модель >50%; Snorkel доминирует в legal/edu/med.
  • Cursor co-training: первый deliverable после покупки Anysphere (июнь 2026); все планы Cursor.
  • Caveats: hallucination rate 54%, CursorBench снят из-за training contamination, EU API — середина июля.

Шесть blind spots: «Opus-class за 1/4 цены»

  1. Sticker price trap: $2/M in, $6/M out не отражают exponential token delta в high-frequency agent loops.
  2. Harness mismatch: DeepSWE 1.0 (vendor harness) — Grok #3; neutral 1.1 — gap 17 pp. Читайте footnotes.
  3. Coding vs agent: Fable 5 +16 pp на SWE-Bench Pro; Grok 4.5 #1 на AutomationBench-AA.
  4. CursorBench incident: snapshots Cursor repo попали в training set — scores retracted.
  5. First-shot vs cost curve: TryAI 3D cube — Opus/Fable one-shot; Grok retry; GPT-5.5 fail.
  6. Production hallucinations: AA-Omniscience 54% — finance/security code требует output validation.

Grok 4.5: specs

Release 8 июля 2026. Заточен под coding + code agents, agentic workflows, knowledge-heavy tasks (legal, med, edu, data). Co-trained с Cursor на триллионах tokens реальных dev interactions. SpaceX купил Anysphere в июне 2026.

ПараметрЗначение
ArchitectureMixture of Experts (MoE)
Context500 000 tokens
ReasoningLow / Medium / High (default: High)
SpeedOfficial 80 TPS, measured ~90 TPS; first token <0.5s, ~110 tok/s
Training HWTens of thousands NVIDIA GB300 (Memphis DC)
ParamsUndisclosed (MoE)

Pricing: sticker vs real per-task cost

API (per 1M tokens)

ModelInputOutput
Grok 4.5$2.00$6.00
Grok 4.5 (cache hit)$0.50—
Grok 4.5 Fast$4.00$18.00
Claude Opus 4.7$5.00$25.00
Claude Fable 5HigherHigher
GPT-5.6 Sol$5.00$30.00
GPT-5.6 Luna$1.00$6.00
Model / platformAvg tokensPer-task cost
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
info

4.2× token efficiency: SWE-Bench Pro output — Grok 15 954, Opus 4.8 67 020. 500 tasks/day: $1 245 vs $5 900.

Coding benchmarks

BenchmarkGrok 4.5Fable 5Opus 4.8GPT-5.5
DeepSWE 1.062.0%66.1%55.75%64.31%
DeepSWE 1.1 (neutral)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%
warning

CursorBench retracted: Cursor codebase snapshots в training data — data contamination. Не цитируйте до independent retest.

Agent benchmarks

BenchmarkGrok 4.5Fable 5Opus 4.8
AutomationBench-AA (657 workflows)51.4%48.6%48.5%
Snorkel GDPVal+29%—21%

Snorkel: legal 40% vs 27–28%; education 58% vs 35–42%; medical 35% vs 23–25%.

Artificial Analysis Intelligence Index

54 (4th place): Fable 5 (60), Opus 4.8 (56), GPT-5.5 (55). +16 vs previous gen.

TryAI four-model shootout

3D cube render: Opus/Fable one-shot OK; Grok retry; GPT-5.5 fail. Speed: first token <0.5s, ~110 tok/s (~2× peers), lowest run cost.

Platforms, API, cache

  • Grok Build, Cursor (all plans, 2× week one)
  • SpaceXAI Console API — us-east-1, us-west-2; 150 req/s, 50M tok/min
  • Office plugins, OpenRouter, Vercel, Cloudflare, Snowflake, Databricks Mosaic
  • EU API: mid-July 2026
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": "Найди и исправь баг: function median(a){a.sort();return a[a.length/2]}"
  }'

Cache tips: prompt_cache_key (Responses API) или x-grok-conv-id (Chat Completions) — cache hit: input $0.50/M. Context Compaction для long agent loops.

Fit / caution matrix

ScenarioVerdictWhy
High-frequency agentsПодходит — Grok 4.5$2.49 vs $11.80 per task
Terminal / tool callsПодходитTerminal Bench + AutomationBench top tier
Cursor-native teamsПодходитCo-training, zero-friction switch
Hybrid routingРекомендуетсяRoutine → Grok; arch → Fable 5
SWE-Bench Pro precisionОсторожноFable 5 +16 pp
Hallucination-sensitive prodОсторожноAA-Omniscience 54%
EU usersОсторожноUS regions only until mid-July
CursorBench claimsОсторожноContamination, scores pulled

6-step runbook

  1. Task profile 30d: calls, tokens, retry rate.
  2. Cursor pilot: model picker → Grok 4.5; terminal/bulk refactor first.
  3. API cache keys; measure hit rate + bill.
  4. SWE-Bench-style regression on real PRs.
  5. Hybrid route: lint/tests/docs → Grok; security → Fable 5.
  6. Output validation gates on hallucination-critical paths.

Три hard numbers

  • 4.2× — SWE-Bench Pro output tokens (15 954 vs 67 020)
  • 51.4% — AutomationBench-AA, first >50% with constraints
  • 500 000 — context window tokens

Verdict + MACCOME bridge

Grok 4.5 — не benchmark champion, а лучший price/perf Opus-tier coding agent. Local MacBook ломается на sleep, unified memory contention и отсутствии 7×24 gateway.

Для stable Grok 4.5 + Cursor hybrid / OpenClaw Gateway: MACCOME Mac cloud — real macOS, SSH handoff. Pricing: цены аренды Mac mini.

Sources: SpaceXAI, Cursor, API docs, TechCrunch, Awesome Agents, APIdog, Snorkel AI. Data as of 2026-07-10.

FAQ

Grok 4.5 лучше Claude Opus 4.8?

SWE-Bench Pro: Opus 4.8 (69.2% vs 64.7%). Grok 4.5 ~4× лучше по speed, token efficiency, cost; agent benchmarks чуть впереди.

Бесплатно?

Limited free credits в Grok Build и Cursor. Далее API $2/M in, $6/M out.

Grok 4.5 в Cursor?

Все планы. Model picker → Grok 4.5. IDE comparison: decision matrix.

Context window?

500 000 tokens.

Почему сняли CursorBench?

Training contamination от Cursor repo snapshots. Independent retest pending.

OpenRouter?

Да: OpenRouter, Vercel, Cloudflare, Snowflake, Mosaic. 7×24 agent node: MACCOME аренда.