Grok 4.6 release date: реалистичен ли дедлайн xAI на 7 августа?

1.5T params · SFT/RL upgrade · roadmap Grok 4.7 · давление Kimi K3 · Musk time buffer

Grok 4.6 release timeline и xAI roadmap breakdown

28.07.2026 Elon Musk ответил на X CEO Vercel Guillermo Rauch: xAI планирует выкатить Grok 4.6 ~7 августа — 1.5T params, significantly improved SFT & RL — и через пару недель Grok 4.7 на 2.1T. Три frontier-модели за ~2 месяца после Grok 4.5, при этом benchmarks, pricing и model card отсутствуют. Если вы оцениваете agent stack или ждёте API pricing next-gen, здесь полный timeline, таблицы roadmap Grok, competitor comparison, контекст SFT/RL, flagged controversies и six-step pre-release runbook — каждый факт с source tier. Data as of 30.07.2026.

01

Confirmed vs unverified: что реально в data set

On the record — один X-post. Всё остальное (1.5T params, SFT/RL upgrade, Grok 4.7 2.1T) — из одного social post без model card, benchmark suite и pricing page как у Grok 4.5. Для build decision это критично.

Event chain:

DateEvent
08.07.2026xAI ships Grok 4.5 для coding/agentic work, co-trained с Cursor на real dev sessions. 500K context, $2/$6 per 1M tokens, model card с 15 benchmarks.
16–26.07.2026Moonshot Kimi K3: hosted preview → full open-weight (2.8T, 1M context), HF trending #1; Musk хвалит в comments.
28.07.2026Musk posts Grok 4.6/4.7 roadmap reply to Rauch. Same day: 1200+ employees публикуют «Pacing the Frontier».
~07.08.2026 (target)Grok 4.6, 1.5T, SFT/RL upgrade not raw scale-up.
Late Aug–early Sep 2026Grok 4.7, 2.1T — «better than 4.6 in every way except slightly slower to serve, better token efficiency».

Пять blind spots перед тем как строить roadmap на этот timeline:

  1. 01

    Single source = tweet: нет xAI blog, model card, product page — executive statement без SLA.

  2. 02

    Musk time track record: xAI/Tesla/SpaceX deadlines сдвигаются на days–weeks; «around Aug 7» = target window.

  3. 03

    Params ≠ capability: Musk акцент на SFT & RL; 4.7 bigger но «slightly slower to serve».

  4. 04

    Не игнорить Kimi K3: Grok 4.6 target ~10 days после K3 open-weight — главный driver release compression.

  5. 05

    Industry split on pacing: Grok announce day = «Pacing the Frontier» day — xAI absent из signatories.

02

Hard numbers: Grok roadmap + field comparison

Grok series roadmap table

ModelDateParamsFocusStatus
Grok 4.3 Beta17.04.2026UndisclosedBaselineShipped
Grok 4.508.07.2026Undisclosed (single SKU)Coding/agentic, Cursor co-trainShipped, benchmarked
Grok 4.6~07.08.20261.5TSFT/RL upgradeTweet announce
Grok 4.7~late Aug–early Sep2.1TUpgrade vs 4.6, token efficiencyTweet announce

All Grok 4.6/4.7 specs = unverified vendor claims из одного post — treat as directional.

Grok vs competition matrix

ModelVendorParamsContextPricing (in/out per 1M)Source
Grok 4.5xAIUndisclosed500K$2 / $6xAI official
Grok 4.6 (announced)xAI1.5TUndisclosedUndisclosedMusk X post (unverified)
Kimi K3Moonshot AI2.8T MoE (~16/896 active)1M$0.30 / $3 in, $15 outMoonshot + HF
Claude Fable 5.1 (rumored)AnthropicUndisclosedUndisclosedRumor: Fable 5 ($10/$50)36kr, WinCentral — unconfirmed
GPT-5.6 SolOpenAIUndisclosedUndisclosedUndisclosedOpenAI official

Pre-release rows — для timing, не head-to-head. См. Grok 4.5 review и Kimi K3 open-weight explainer.

«Significantly improved SFT & RL» — тот же post-training playbook что у Grok 4.5: ~15,954 output tokens per SWE-Bench Pro task vs 67,020 у Opus 4.8 (4.2× gap), credit to Cursor session data not sheer size.

03

Six-step runbook: prep before Grok 4.6 ships

For AI devs, tech leads и agent engineers — independent judgment framework:

  1. 01

    Verify source tier: Musk X post vs xAI blog/model card — label everything announced/unverified.

  2. 02

    Build Musk time buffer: «around Aug 7» = window not deadline; 1–2 week slack in budget/vendor pick.

  3. 03

    Anchor on Grok 4.5: 15 benchmarks, $2/$6 — baseline for switch decision, not param count alone.

  4. 04

    Cross-check Kimi K3 + Fable 5.1 rumors: K3 tops Frontend Code Arena 1,679 pts, Artificial Analysis #3; Fable 5.1 leaks → August — three tracks, no head-to-head.

  5. 05

    SFT/RL operational meaning: SFT shapes behavior, RL optimizes agent chains. Watch metric: 15,954 vs 67,020 output tokens SWE-Bench Pro.

  6. 06

    Hybrid routing plan: Grok Build, xAI API, Cursor first — routine on proven models, architecture on verified frontier until pricing/benchmarks drop.

04

Post-training > raw scale: SFT/RL, competition, controversies

SFT/RL in plain geek terms

SFT = curated example outputs → behavior shaping. RL = reward signals → multi-step agent sequences that actually work. Grok 4.5 hit Terminal-Bench 2.1 83.3%, SWE-Bench Pro 64.7% при ~15,954 output tokens vs Opus 4.8 67,020 — post-training on real Cursor sessions, not size alone.

Scale vs speed — dual SKU strategy

4.6 → 1.5T jump; 4.7 → 2.1T but «slightly slower» — xAI segments like Anthropic Sonnet/Opus или OpenAI mini/full tiers.

Kimi K3 competitive pressure

Grok 4.6 target ~10 days post K3 open-weight shock: Frontend Code Arena 1,679 pts, first open-weight beating all closed models on that board, Artificial Analysis Intelligence Index #3. Chinese media market reaction estimates — analyst sentiment, not hard corp numbers.

!

Controversy 1: tweet-only source. No xAI blog, model card — executive public statement, zero obligation to hit date.

!

Controversy 2: blank benchmarks/pricing. Unlike Grok 4.5 launch — «1.5T» and «SFT/RL upgrade» remain unverified vendor claims.

Background: xAI safety drama. July 2026 xAI sued user for CSAM via Grok bypass — first lawsuit of kind. Common Sense Media Jan 2026 rated Grok among worst for child-safety. Vendor risk context, not pure tech eval.

Background: pacing split. Same day as Grok 4.6/4.7 announce → «Pacing the Frontier» 1200+ signatories — xAI absent, strategic divergence accelerate vs slow.

05

Hard facts, August bottleneck, VpsMesh fit

  • Grok 4.6: 1.5T params (Musk announce), zero independent benchmark.
  • Grok 4.7: 2.1T, «few weeks» after 4.6, better token efficiency, slightly slower serve.
  • 4.5 token efficiency ref: ~15,954 vs 67,020 output tokens SWE-Bench Pro — 4.2× gap, key metric for 4.6.
  • Kimi K3 coords: 2.8T MoE, 1M context, Frontend Code Arena 1,679, Artificial Analysis #3.
  • Fable 5.1 rumor window: August leaks, beat GPT-6 timing — Anthropic unconfirmed.
  • Triple-release cadence: 4.5 + 4.6 + 4.7 in ~2 months — flagship shelf life → weeks.

August 2026 = likely densest LLM release month — token efficiency + real task cost beat leaderboard rank alone as decision basis.

Multi-model API switching, agent orchestration pipelines или Grok/Cursor joint dev на laptop = unstable processes, no true 24/7, compile bottlenecks; self-hosted VPS без Apple Silicon = no Metal/Xcode chain. For production iOS CI/CD + AI agent automation, VpsMesh Mac Mini M4 cloud rental = usually better fit: unified memory for large-context agent orchestration, remote nodes 24/7 для multi-model routing tests без polluting local env. Pricing: цены аренды Mac Mini M4, setup: центр помощи.

FAQ

FAQ по Grok 4.6

Musk said «around August 7, 2026» в X-post — xAI официально не confirmed. Target с возможным shift. Follow xAI на X и x.ai.

Grok 4.6: 1.5T, SFT/RL post-training focus. Grok 4.7: 2.1T через few weeks, outperforms 4.6 except serving speed — trades latency for token efficiency. Dual SKU «faster» vs «stronger».

Too early. No published Grok 4.6 benchmarks; Kimi K3 has verified third-party scores; Fable 5.1 not confirmed. См. Kimi K3 open-weight explainer.

Unknown. Grok 4.5: $2/M input, $6/M output — reasonable reference. Agent test env: центр помощи.

Based on Grok 4.5 rollout: Grok Build, xAI API, console first, then third-party like Cursor — not confirmed for 4.6 yet.