SuperGrok Heavy + Grok Build (swept 2026-07-14)

Trigger: user is taking the $99–100/mo SuperGrok Heavy promo (3 months). Question: what is it, quirks, strengths vs Claude Code / Codex, top-voice consensus, loop-engineering fit. Method: agent-reach (Exa + Twitter/OpenCLI + Reddit/OpenCLI + Jina), local ~/.grok forensics. All URLs verified live 2026-07-14.

The plan (facts)

• SuperGrok Heavy: normally $300/mo; current promo $99/mo for 3 months (67% off), started ~Jul 12 2026 → promo expiry ≈ 2026-10-12 for this account. (An earlier May promo was $99 for 6 months — do not confuse; the July offer is 3 months. Receipts: x.com/i/status/2076443434991120741 "only $99/month for the first 3 months"; x.com/i/status/2076277534912565495 "at least for the 3 months 67% off"; testingcatalog.net/xai-slashes-supergrok-heavy-price-by-67-to-promote-grok-build/ documents the May 6-month variant.)
• Claim flow is glitchy: "Claim Offer" pushes to the mobile Play Store; desktop shows normal pricing (x.com/i/status/2076321441079812416).
• What Heavy buys: 10× the usage of SuperGrok (screenshot receipt x.com/i/status/2076759599680491530 — who judged it "not worth it vs the ×20 from Ant/OpenAI"), Grok 4 Heavy multi-agent expert mode, highest rate limits, max Grok Build tier, priority everything.
• Context: base SuperGrok ($30) moved to a shared weekly pool in July and users hit 100% in days (x.com/i/status/2076701147910324457) — Heavy is the only escape besides credits.
• Community claims SuperGrok subs include a real xAI API credit allowance usable via OAuth proxies (reddit.com/r/LoveGrok/comments/1u6egj6/, unverified by xAI docs — verify before relying).

The model (Grok 4.5, current flagship)

• #1 on Long-Horizon Terminal-Bench as of Jul 14: 13/46 tasks vs Fable 5's 12 (Musk: x.com/i/status/2076909201524203973; analysis x.com/i/status/2076879157137686634 — tasks cost ~9.9M tokens/85min; Grok 4.5 "solves tasks in under half the number of steps" → input-replay efficiency compounds).
• API $2/$6 per MTok (felloai.com/grok-pricing/); Musk floated lowering cached-input price (x.com/i/status/2076411423563948142).
• Musk marketing: "ranks slightly above Fable on some software benchmarks" (x.com/i/status/2076434607562531049). Treat as marketing; the older grok-build-0.1 coding model scored 70.8% SWE-Bench Verified vs Opus 4.7's 87.6% / GPT-5.5's 88.7% (aitoolgrade.com, aicoderscope.com — May data, pre-4.5).
• Knowledge cutoff reported Nov 2024 for Grok 4 family (locoroo.net May report); xAI publishes no official coding benchmarks.

The harness (Grok Build CLI, beta since 2026-05-14)

• Native 8-way parallel subagents, each in an isolated git worktree — the category first; matches our write-set-lease doctrine structurally (aicoderscope.com, aitoolgrade.com).
• Plan mode default, clean diffs; Code/Plan/Ask modes; headless -p; lifecycle hooks; MCP day-one; ACP scripting; /dashboard backgrounds an agent and lets you start another; agents can spawn background subagents (xAI's milichab: x.com/milichab/status/2076388813581258818).
• Claude-compat layer: reads CLAUDE.md natively; [compat.claude] in ~/.grok/config.toml auto-discovers ~/.claude skills/rules/agents/MCPs/hooks/sessions. Local finding: this import pulled our python3 guard hooks + lavish-axi into Grok and they failed on Grok startup — that was the "startup failures" the user saw. Now clean-slated with all compat.claude flags = false (keep them false).
• Custom model routing to ANY provider — can drive Claude/OpenAI models from inside Grok Build (aimadetools.com); community grok-delegate plugin adds /delegate-claude, /delegate-codex (x.com/i/status/2076743478063583343).
• Weaknesses: 256K context with silent truncation on big repos (aicoderscope.com); no sandbox; local-process agents only — no cloud/background task infra, machine must stay awake (vs Codex cloud tasks); beta flakiness; Arena Mode announced not shipped.

⚠️ THE INCIDENT (2026-07-12→14) — trust status: probation

• Wire-level captures showed Grok Build uploading entire git repos — full history, unredacted .env secrets — to GCP bucket grok-code-session-traces (5.1GB from a 12GB repo when the task needed 192KB). The "Improve the model" toggle did NOT stop it. (IntCyberDigest x.com/i/status/2076689215258014069, 1.5M views; XBToshi x.com/i/status/2076521420017045618; mardehaym wire capture on 0.2.93 x.com/i/status/2076790621462044876.)
• xAI "fix" = a hidden server-side flag disable_codebase_upload: true; a researcher flipped it back via proxy and the client immediately retried the upload — the client code path still exists (WesEklund x.com/i/status/2076718841699164236; repro harness github.com/weklund/grok-network-monitor).
• xAI official: ZDR always respected incl. API-key use; /privacy opt-out disables retention and retroactively deletes synced data (x.com/i/status/2076692402442846289; milichab x.com/i/status/2076693464016994685). Independent retest confirms repo uploads stopped globally; /privacy now only gates an ~18KB trace channel (cereblab x.com/i/status/2076869830020575456).
• Top-voice verdicts: Gergely Orosz "Cannot see any sensible company use Grok CLI based on this terrible incident" (x.com/i/status/2076928626457518184). Kun Chen (@kunchenguid — author of no-mistakes, which we run): "genuinely a good model and harness… run /privacy asap… I almost bought the $300 heavy as my 3rd subscription, and this broke the deal" + hardened config.toml options (x.com/kunchenguid/status/2076810442358542551). aidenybai: "really like grok 4.5… massive fuckup, immediately lost all trust for grok build."
• Local forensics (this machine, 2026-07-14): ~/.grok/logs/unified.jsonl contains zero repo_state/upload events — user's sessions postdate the server-side kill and ran in non-repo ~. Still: run /privacy opt-out once inside Grok Build to trigger the retroactive delete guarantee.

Comparative verdict (June–July press consensus)

X research — the actual moat (swept 2026-07-14; THE reason for this subscription)

What Grok has that nothing else does: first-party structured access to live X — posts seconds after publish with thread context, quote chains, reply structure, engagement metrics, verified/follower weighting, and the social graph (learn.techwithdarin.com/guides/grok/). Every other assistant sees X through a lagged web-crawl keyhole.

Programmatic surface (the unlock): server-side x_search on the Agent Tools API (docs.x.ai/developers/tools/x-search; the old Live Search API was deprecated end-2025):

• Four modes the model picks itself: x_keyword_search, x_semantic_search, x_user_search, x_thread_fetch.
• Filters: allowed_x_handles / excluded_x_handles (max 20), from_date/to_date (ISO8601), enable_image_understanding, enable_video_understanding (X-only — reads video content in posts).
• $5 per 1,000 successful calls (~$0.005/search; failed calls free) + model tokens; max_turns dials depth (1–2 quick, 3–5 balanced, 10+ deep). Answer-shaped output with citations in ONE API call — search + read + synthesize collapsed (hybrid-llm.com cost analysis: same per-read price as the official X API but you get the finished answer, not raw posts).
• Works from ANY OpenAI-compatible client (model="grok-4.5", tools=[x_search(...)]) → callable from a Claude Code skill; prior art: github.com/adaptationio/skrillz .claude/skills/xai-x-search. API-key use = ZDR per xAI's official statement — the trust-probation concern is Grok Build's client, not this API.
• ⚠️ CORRECTED 2026-07-14: SuperGrok Heavy does NOT include API access/credits — the API is separately billed prepaid (user has no key; the Reddit "$200 included credits via cliproxy" claim did not hold for this account and rides OAuth-behind-API shims besides). The sub-billed programmatic rail is headless Grok Build: ~/.grok/bin/grok -p "<research prompt>" --rules … --no-subagents --no-plan — VERIFIED LIVE 2026-07-14: a 20-handle X sweep with thread fetches returned same-day cited posts, $0 marginal. Run it in an empty non-git workdir (nothing for the client to upload → probation-compliant). This is how ~/.claude/skills/xalpha-loops/ runs; a Keychain xai-api-key auto-upgrades it to the true x_search API.
• NOT a firehose: batch queries against the live index, minutes-fresh; for as-they-appear streaming you'd still need X's filtered stream. Polling on a loop covers the alpha use case.

Interactive surface: DeepSearch/DeeperSearch on grok.com — agentic X+web research (20–50 sources, minutes); Heavy = effectively uncapped DeepSearch + Grok 4 Heavy multi-agent research ("aimed at researchers, traders, journalists"). SuperGrok base is capped and the July weekly-pool change made caps bite in days.

vs agent-reach (our OpenCLI twitter backend): browser-session scraping of the search UI — latest/top only, no semantic search, no thread-fetch at scale, no handle-scoped 20-account sweep in one call, no engagement weighting, no media understanding, suspension risk, and Claude burns tokens reading raw YAML. x_search replaces the analysis layer, not just fetch. agent-reach stays for: platforms beyond X, and when we need raw post objects rather than answers.

Caveats (field-tested by others): ~1 in 10 DeepSearch citations materially wrong under time pressure — open sources before relying (techwithdarin); sarcasm/bot blindness; hallucination rate flagged in some API configs (byteiota) → alpha digests should carry post URLs so claims are one click from verification.

Hackathon-alpha workflows this enables:

1. Top-voices sweep as a cron loop: grok-4.5 + x_search(allowed_x_handles=[≤20 ledger voices], from_date=yesterday) → daily digest. Cost/run ≈ a few cents (5–10 searches + tokens). Feeds the RESEARCH_top_voices ledger automatically.
2. Sponsor/deadline watch during hackathons: handle-scoped sweep of sponsor orgs + judges; semantic query "teams shipping for — what's getting traction?"
3. Live-event readout: DeepSearch on grok.com (uncapped on Heavy): "consensus on X about in the last hour, weighted toward verified, 3 representative posts per opinion cluster."
4. Never the orchestrator. No self-wake/scheduled-resume primitive found; parallel agents are local-process (machine must stay online). Same seating as CX: function call in, terminal artifact out — no async-wait supervision.
5. Executor niche it actually wins: cleanly-splitting parallel work (test backfills, N-endpoint migrations, K-hypothesis exploration) — its native worktree isolation IS our write-set-lease rule, enforced by the harness.
6. Reviewer seat: a third non-author provider for the peer pattern (Claude/OpenAI/xAI 3-way on irreversible/security tier) — decorrelation gain, but weigh the model gap.
7. Security quarantine (hook-shaped, not prose): never point Grok Build at a repo containing secrets until N clean weeks post-incident; /privacy opt-out + disable_codebase_upload: true in config.toml; keep [compat.claude] all-false so it never re-ingests ~/.claude; prefer API-key mode (ZDR) for anything sensitive.
8. Route tolerance applies: a Grok-authored artifact gets reviewed by Claude/Codex, not iterated on mid-run by a different model.
9. Promo economics: at $99–100/mo it's an off-Claude-quota, off-OpenAI-quota third pool. Decision gate ~2026-10-05 (a week before promo end): earned its keep as parallel executor / third reviewer? If not, cancel before the $300 revert.