Grok 4.6 дата выхода: 1,5T SFT/RL pipeline · Grok 4.7 на 2,1T · zero public benchmarks
Для кого: Mac-разработчиков на Cursor/xAI API, которым нужен tech breakdown roadmap 4.6/4.7 после твита Маска 28 июля. Внутри: timeline table, specs matrix, SFT/RL stack analysis, competitor comparison vs Kimi K3 (2.8T) и Fable 5.1 rumor, CSAM lawsuit context, Pacing the Frontier letter, 5-step integration runbook, FAQ×5, destroy-after-acceptance Mac rental.
Оглавление
См. также: Grok 4.5 обзор (11.07) · Kimi K3 open weight (28.07) · OpenRouter июль 2026.
TL;DR
28 июля 2026 Маск ответил @rauchg: Grok 4.6 ~7 августа, 1.5T params, massive SFT/RL post-training upgrade. Через несколько недель — Grok 4.7 на 2.1T: better across the board except inference latency, higher token efficiency. xAI published zero model cards or benchmarks for 4.6. Single primary source: one X reply. Factor in Musk time (±days to 2 weeks slip).
Three weeks post Grok 4.5 GA (July 8: $2/$6 per M tokens, 500K context, Terminal-Bench 83.3%, SWE-Bench Pro 64.7%, co-trained with Cursor on real dev sessions), Musk signals a triple-frontier summer. Kimi K3 dropped 2.8T open weights July 27 with third-party verified scores. Grok 4.6 has none. This post separates vendor claims from reproducible metrics.
01 · Три tech pain points
- Single-source dependency — Release date, 1.5T count, SFT/RL focus: all from one X reply. Grok 4.5 shipped model card + 15 benchmarks. Writing 4.6 into production architecture docs today = betting on unverified vendor tweet.
- Benchmark void vs parameter PR — 1.5T/2.1T headline numbers without independent eval. Kimi K3: SWE-bench 93.4%, AA Index #3. Grok 4.5 proved token efficiency matters: ~15,954 output tokens/SWE-Bench-Pro task vs ~67,020 for Claude Opus 4.8 (~4.2× delta). Params ≠ $/task.
- Compressed eval window + Musk time — 4.5 → 4.6 → 4.7 in ~8 weeks. Musk release history: typical slip days to 2 weeks. Pinning xAI API keys to daily-driver Mac before 4.6 is measurable increases migration cost.
02 · Release timeline: Grok 4.5 → 4.6 → 4.7
| Date | Event | Status |
|---|---|---|
| 2026-07-08 | Grok 4.5 GA — coding/agent SKU, Cursor co-training, 500K ctx, $2/$6, Terminal-Bench 83.3%, SWE-Bench Pro 64.7% | Shipped |
| 2026-07-16 | Kimi K3 hosted launch (2.8T MoE, 1M context) | Shipped |
| 2026-07-27 | Kimi K3 full open weights (~1.56TB on HF) | Shipped |
| 2026-07-28 | Musk Grok 4.6/4.7 roadmap (@rauchg reply). Same day: «Pacing the Frontier» open letter (1,200+ signatories) | Announced |
| ~2026-08-07 | Grok 4.6: 1.5T, SFT/RL upgrade (target, unconfirmed) | Planned |
| ~late Aug 2026 | Grok 4.7: 2.1T, better but slower serving, higher token efficiency | Planned |
03 · Grok series specs matrix
| Model | Params | Context | API ($/M in/out) | Public benchmarks |
|---|---|---|---|---|
| Grok 4.5 | undisclosed | 500K | $2 / $6 | Terminal-Bench 83.3%; SWE-Bench Pro 64.7% |
| Grok 4.6 | 1.5T | TBD | TBD | None |
| Grok 4.7 | 2.1T | TBD | TBD | None |
Key metric from 4.5: post-training + real session data drove agent scores — not raw param count. 4.6 continues that axis.
04 · SFT/RL pipeline: что реально апгрейдится
4.1 SFT (Supervised Fine-Tuning)
Curated human demonstrations align output distribution. Grok 4.5 ingested Cursor dev session traces — measurable in agent benches, invisible in param table.
4.2 RL (Reinforcement Learning)
Reward signals optimize multi-step agent action sequences. Musk claims «massive» RL upgrade for 4.6 — zero published eval harness or reward model details.
4.3 Dual-SKU inference tradeoff
4.7: «better in almost everything except serving speed, more token efficient» — classic fast/capable tier split (Sonnet/Opus pattern). Don't expect one model to win latency and capability simultaneously.
4.4 Kimi K3 pressure
4.6 target lands ~10 days after K3 open weights. K3: Frontend Code Arena 1679, AA Index #3. Musk commented «impressive» on K3 benchmark posts — competitive signal, not benchmark data for 4.6.
05 · Competitor comparison table (2026-07-30)
| Model | Vendor | Params | Context | Verification |
|---|---|---|---|---|
| Grok 4.5 | xAI | n/a | 500K | xAI official |
| Grok 4.6 | xAI | 1.5T | n/a | Musk tweet only |
| Kimi K3 | Moonshot | 2.8T MoE | 1M | Official + third-party |
| Claude Fable 5.1 | Anthropic | n/a | n/a | Aug rumor (unconfirmed) |
| GPT-5.6 Sol | OpenAI | n/a | n/a | OpenAI official |
06 · Risk factors
- CSAM lawsuit (July 2026) — xAI sued user for bypassing safety filters to generate CSAM. First deepfake/CSAM litigation for the company. Compliance risk orthogonal to benchmark scores.
- Pacing the Frontier (July 28) — 1,200+ employees from OpenAI, Anthropic, Google DeepMind, Meta urge frontier AI slowdown. xAI not a signatory while accelerating 4.6/4.7.
- Zero 4.6 benchmarks — Unlike 4.5's model card drop, no independent eval exists. «1.5T + SFT/RL» = vendor claim until reproduced.
07 · August 2026: peak release density
Grok 4.6 (~Aug 7), Grok 4.7 (late Aug), Fable 5.1 rumor (Aug), Kimi K3 aftermath — highest LLM release density month of 2026. Decision metric: $/task and output token count, not leaderboard rank. Budget Musk time.
08 · 5-step integration runbook (Mac devs)
- Lock Grok 4.5 baseline — xAI API or Cursor: long-doc summary, bugfix, multi-file refactor ×1 each. Log output tokens and $/task.
- Watch official channels — x.ai/news, @xai. On X vs blog divergence, trust blog.
- Run Kimi K3 A/B now — Same repo, same prompts — bridge 4.6 benchmark void.
- Isolated rental Mac — xAI key off daily-driver Keychain; validate Grok Build / Cursor routing.
- Wiki with «unverified» tag — Source: Musk Jul 28 @rauchg. Update SLA: 24h post xAI GA.
09 · FAQ
Q: Когда выйдет Grok 4.6?
A: Маск назвал ~7 августа 2026 на X — не официальная дата xAI. Заложите Musk time (± дни до 2 недель).
Q: Отличие 4.6 vs 4.7?
A: 4.6 = 1.5T + SFT/RL focus. 4.7 = 2.1T, лучше кроме serving latency, выше token efficiency. Оба не выпущены.
Q: Сильнее Kimi K3 / Fable 5.1?
A: Неизвестно. K3 имеет верифицированные бенчмарки; Fable 5.1 — слух; у 4.6 нет публичных скоров.
Q: Цена 4.6?
A: Не объявлена. Референс 4.5: $2/$6 per M tokens.
Q: Где использовать?
A: Нигде официально. Путь 4.5 (Grok Build, API, Cursor) вероятен, но не подтверждён.
10 · Изолированный Mac rental: Grok API без Keychain pollution
1.5T local inference on MacBook: not happening. Path = xAI API / Cursor routing. But pinning xAI keys on daily driver + long-context burn tests + 4.5→4.6 switch rehearsal = Keychain leak, shell history exposure, runaway billing risk.
Dedicated rental Apple Silicon node: lock 4.5 workflow today, A/B 4.6 post-GA, destroy instance after acceptance. Windows/Linux can hit grok.com but can't validate macOS-native Xcode/Cursor/Keychain agent pipelines. Pricing: тарифы вычислений серии M.
11 · Sources
- Primary: Musk X post Jul 28 2026 (@rauchg reply)
- Official: x.ai/news Grok 4.5, media.x.ai model card
- Competition: Moonshot Kimi K3, Artificial Analysis, Vals AI
- Context: Pacing the Frontier (The Verge), xAI CSAM suit (Ars Technica, TechCrunch)
- Fable 5.1 rumor: 36kr, WinCentral
Дата: 30 июля 2026. Grok 4.6/4.7 — provisional until xAI GA.