Grok 4.6 Review 2026: xAI's Agentic Flagship Ties the Frontier at Half the Price — AA 61, GDPVal-AA #1, $2/$6
Grok 4.6 Review 2026: xAI's Agentic Flagship Ties the Frontier at Half the Price — AA 61, GDPVal-AA #1, $2/$6
Verdict first: Grok 4.6 (August 12, 2026) is xAI's most complete agentic flagship yet — a release whose story is less "one huge leap" and more "a sharp model wrapped around a clearer agent-loop product." On xAI's vendor-reported numbers it now ties GPT-5.6 Sol Max at 61 on the Artificial Analysis Intelligence Index, takes #1 on GDPVal-AA v2 (Elo 1753) and AA-Briefcase, and reaches 69.9% on CursorBench v3.2 — all at a standard price of $2/$6 per million tokens, roughly half the sticker of comparable OpenAI/Anthropic models. The gains are concentrated exactly where xAI aimed: long-horizon agent work (DeepSWE 54→65.9%, APEX-Agents 47→57.5%). The honest caveats haven't moved: these are vendor-reported figures with no independent replication yet, Terminal-Bench still sits at a low absolute 26%, DeepSWE trails the leaders, and crossing the 200K-prompt price cliff bills your whole request at the higher tier. At half price it is currently the value pick of the frontier — but it is a value pick, not a clean sweep.
Grok 4.6 — xAI's long-horizon agent flagship, priced at half the frontier
What Is Grok 4.6?
Released August 12, 2026, Grok 4.6 is xAI's flagship upgrade to Grok 4.5 — the same 500K-context class, but trained with a longer supplemental run focused on long-running agents, agentic coding, and turning a broad product idea into a working first version (research → structure → implement → refine). It launched on the xAI API, Grok Build, Cursor, OpenRouter, Vercel and Cloudflare, plus Grok Bot on X. (Sources alternate between "xAI" and "SpaceXAI" following the reported SpaceX–xAI merger.) For family context see our Grok 4.5 review.
What Changed vs Grok 4.5
- Long-horizon agent gains (vendor-reported): DeepSWE v1.1 54% → 65.9%; APEX-Agents 47.1% → 57.5%; Terminal-Bench v3.0 ~15.7% → 26%; GDPVal-AA v2 Elo 1526 → 1753.
- Better first passes: more self-verification (runs its own tests, re-reads edited files) before continuing; stronger visual/interactive first attempts on dashboards and UI projects.
- Agent-loop plumbing:
prompt_cache_key(Responses) /x-grok-conv-id(Chat Completions) keep multi-turn loops on the same server — the difference between $0.50 and $2.00 cached input per agent turn — plus context compaction on long runs. - Cost-aware design: a 200K-prompt price cliff, a 2x Priority Processing latency lane (no separate fast model), and same opening list price as Grok 4.5.
Specifications at a Glance
| Item | Value |
|---|---|
| Vendor | xAI (SpaceXAI) |
| Release date | August 12, 2026 |
| Positioning | Long-running agents, agentic coding, interactive/visual first passes |
| Parameters | Not disclosed (one report cites ~1.5T on Grok 4.5's V9 foundation) |
| Context window | 500,000 tokens (same class as Grok 4.5) |
| Knowledge cutoff | February 1, 2026 |
| Modalities | Text + image input → text output |
| Reasoning | reasoning_effort low / medium / high (default) / xhigh |
| Tools | Function calling, web search, X search, code execution (billed separately) |
| Open weights | No (closed; API and hosted products only) |
| Availability | xAI API (grok-4.6), Cursor, Grok Build, OpenRouter, Vercel, Cloudflare |
Benchmarks: Tie at the Top, Not a Sweep
All figures below are launch-day, vendor-reported (xAI's launch post) and have not yet been independently verified. "High" refers to the reasoning_effort=high tier.
| Benchmark | Grok 4.5 High | Grok 4.6 High | Frontier context |
|---|---|---|---|
| AA Intelligence Index | 56 | 61 | Ties GPT-5.6 Sol Max (61); behind Claude Fable 5 (62) / Opus 5 (63) |
| GDPVal-AA v2 (Elo) | 1526 | 1753 | #1 — leads GPT-5.6 Sol Max (1728) and Claude Fable 5 Max (1741) |
| AA-Briefcase (Elo) | 1313 | 1577 | #1 on this eval |
| CursorBench v3.2 | 66.7% | 69.9% | Beats Sol Max (67.2%); trails Fable 5 Max (70.5%) |
| DeepSWE v1.1 | 54.0% | 65.9% | Trails GPT-5.6 Sol Max (73%) and Claude Fable 5 (70%) |
| APEX-Agents | 47.1% | 57.5% | Within 1.7 pts of Fable 5 Max; ahead of Sol Max (56.7%) |
| Terminal-Bench v3.0 | ~15.7% | 26% | Big jump, low absolute — trails Sol Max (34.6%) / Fable 5 (34.1%) |
| FrontierCode v1.1 Ext | 56.6% | 61.3% | — |
| Harvey LAB (Vals) | 12.9% | 15.8% | Still low in absolute terms |
Cost efficiency in practice: Artificial Analysis estimated roughly $0.84 average cost per composite eval task (~53 turns), and one developer reported a ~$0.41 overnight bill for building a Minecraft-like game.
Pricing: The Half-Price Economics
| Tier | Input / 1M | Cached / 1M | Output / 1M |
|---|---|---|---|
| Standard (<200K prompt) | $2.00 | $0.50 | $6.00 |
| Long-context (≥200K prompt) | $4.00 | $1.00 | $12.00 |
| Priority Processing | 2x standard rates (faster lane via service_tier) | ||
That is roughly half the sticker price of comparable frontier models (e.g., Claude Opus 5 at $5/$25 or the Fable line at $10/$50). Three gotchas matter for real agent workloads:
- The 200K cliff: once a prompt crosses 200K tokens, the entire request bills at the higher tier.
- Cache keys are the real cost lever: without a sticky key, multi-turn loops often pay full input on a cache-cold box — the difference between $0.50 and $2.00 per agent turn.
- No separate fast model: the only way to buy latency is the 2x Priority lane. Server-side tool calls (web/X search, code execution) add per-invocation fees.
The Honest Take
The real headline is value, not dominance. Grok 4.6 is the first xAI flagship that genuinely ties the frontier on agentic knowledge-work evals (GDPVal-AA #1, AA index 61) while staying at $2/$6. For anyone running long agent loops, the combination of frontier-adjacent scores at half price is compelling — and the platform changes (cache keys, compaction) are aimed exactly at the thing that makes frontier agents expensive: token burn over many turns.
The asterisks are the usual ones plus one new one. Every number here is xAI's own, unreplicated. Terminal-Bench at 26% and Harvey LAB at 15.8% are honest reminders that "knowledge-work Elo #1" does not equal "wins every software-engineering suite" — on DeepSWE and Terminal-Bench it still trails GPT-5.6 Sol Max and Claude Fable 5 by wide margins. And the 200K billing cliff means long-context users must read the pricing table carefully before they trust the "half price" story.
Bottom line: if you want the absolute software-engineering ceiling, Grok 4.6 is not your model. If you want a frontier-class agentic workhorse at roughly half the list price of its rivals — and can manage cache keys and the 200K cliff — it is currently the strongest value argument in the frontier tier.
Pros & Cons
Pros
- ✅ Ties the frontier on the AA Intelligence Index at 61; GDPVal-AA v2 #1 (Elo 1753) and AA-Briefcase #1 (vendor-reported)
- ✅ Real long-horizon gains: DeepSWE +11.9, APEX-Agents +10.4, Terminal-Bench ~+10 (vendor-reported)
- ✅ Genuinely cheap: $2/$6 standard — about half the list price of OpenAI/Anthropic frontier models
- ✅ Agent-first product: prompt cache keys, context compaction, self-verification on long runs
- ✅ Broad availability day one: API, Cursor, Grok Build, OpenRouter, Vercel, Cloudflare; 2x included usage promo first week
Cons
- ❌ Vendor-reported benchmarks with no independent replication yet
- ❌ Not a clean sweep: DeepSWE (65.9%) and Terminal-Bench (26%) trail GPT-5.6 Sol Max and Claude Fable 5
- ❌ 200K-prompt price cliff bills the whole request at the higher tier
- ❌ Cache-key dependence: un-cached multi-turn loops can erase the cost advantage
- ❌ No separate fast model (latency only via 2x Priority); server-side tools billed per invocation
- ❌ Closed weights, undisclosed parameters; Harvey LAB and some enterprise evals still low
Who Should Use It (And Who Shouldn't)
Great fit
- Developers running long-horizon agents where token cost across many turns dominates
- Teams that want frontier-adjacent agentic performance on a budget
- Cursor / Grok Build users (day-one integration, launch promo)
- Knowledge-work and product-first-pass workloads (GDPVal-AA, AA-Briefcase strengths)
Poor fit
- Users who need the absolute software-engineering ceiling (DeepSWE / Terminal-Bench leaders win there)
- Long-context workloads that routinely cross 200K prompt tokens (watch the billing cliff)
- Strict buyers who only move on independently replicated benchmark numbers
- Teams that can't manage cache keys and risk paying full input per agent turn
Frequently Asked Questions
When was Grok 4.6 released?
August 12, 2026, by xAI — a direct upgrade to Grok 4.5 at the same 500K context class, available on the xAI API, Grok Build, Cursor, OpenRouter, Vercel and Cloudflare.
How much does it cost?
$2/$0.50 cached/$6 per 1M tokens under 200K prompt; $4/$1/$12 above 200K (whole request). Priority Processing is 2x. Roughly half the list price of comparable frontier models.
Is it better than GPT-5.6 Sol?
Ties on AA Intelligence Index (61), leads on GDPVal-AA, AA-Briefcase and CursorBench; trails on DeepSWE and Terminal-Bench. Vendor-reported, not independently verified.
What is the context window?
500K tokens, knowledge cutoff February 1, 2026. Text + image in, text out; function calling, web/X search, code execution.
What's new vs Grok 4.5?
DeepSWE 54→65.9%, APEX-Agents 47.1→57.5%, Terminal-Bench ~15.7→26%, GDPVal-AA Elo 1526→1753; plus cache keys, context compaction, and self-verification for long agents.
Is Grok 4.6 open source?
No — closed weights, API/hosted only, parameters undisclosed (~1.5T reported).
Sources
- LLM Stats — Grok 4.6 Release, Benchmarks and Agent Loops
- LLM Stats — Grok 4.6 API Pricing, Context Window & Benchmarks
- Apidog — What Is Grok 4.6? Features, Benchmarks, Pricing, API Access
- BenchLM — Grok 4.6 Benchmarks & Pricing (Aug 2026)
- Edgen — Grok 4.6 matches frontier rivals at half the price
- 华尔街见闻 — Grok 4.6 智能体测试力压 GPT-5.6 / Fable 5,定价仅竞品一半
Disclosure: benchmark figures are launch-day, vendor-reported by xAI and not yet independently verified; pricing and availability are subject to change — confirm on xAI's official pages.
评论
发表评论