Grok 4.6 avg turns

Claude Opus 5 avg turns

Model Launch

By Sam Taylor with SamwiseAug 14, 2026

On the AA-Briefcase Elo jump from 1,313 to 1,577, what 53 turns vs 103 means for token budgets at scale, and whether this belongs in your agent stack at $2/M.

SpaceXAI held the model size constant and cut agentic turn count in half. That's the bet.

Source lean on this story

▲ avg

Anti-AI

Skeptic

Neutral

Pro (practical)

Pro (hyped)

← Anti-AI · Pro-AI →

The cost of running a long-horizon agentic task is not just the per-token price. It's the per-token price, times the number of tokens, across as many turns as the model needs to finish the job. Which is why the number that stuck with me after SpaceXAI's Grok 4.6 release on August 12 isn't the benchmark. It's the turn count.

On Artificial Analysis's AA-Briefcase — their private long-horizon agentic knowledge-work benchmark — Grok 4.6 completes the same tasks in roughly 53 turns on average, consuming roughly 0.5 billion input tokens. Claude Opus 5 Max completes those tasks in roughly 103 turns, consuming roughly 2.0 billion input tokens. Same task category. Same Elo tier on the benchmark. Four times the input tokens for Opus 5.

At $2/M for Grok 4.6 input vs whatever Opus 5's pricing is for your tier, that's not a rounding error. It's a structural cost difference on any workload where you're running the same task at volume.

That's the number most of the coverage has buried or skipped. I don't know why. It's the most builder-relevant fact in the release.

What actually changed

Grok 4.6 is not a new model. It runs the same 1.5 trillion-parameter V9 base as Grok 4.5 — unchanged architecture, same parameter count, same 500,000-token context window. What SpaceXAI changed is everything that happens after the base model: the supervised fine-tuning data was regenerated and filtered using Grok 4.5 itself, the reinforcement learning was extended with more agentic training in coding, web development, computer-aided design, and kernel optimization, and the optimizer and training recipe were updated.

Grok 4.5 was released July 8, 2026. That's 35 days. One post-training run. Five Intelligence Index points.

AA Intelligence Index: Grok 4.6 (61) vs Grok 4.5 (56 high) — from post-training only, no architecture change

→ Source: Artificial Analysis

The AA-Briefcase Elo jumped from 1,313 for Grok 4.5 to 1,577 for Grok 4.6 — landing what Artificial Analysis calls "Fable 5-tier" performance on long-horizon knowledge work. That's a real step-change, not a rounding gain. And it happened without touching the base model.

Grok 4.5 vs Grok 4.6 at a glance

Metric	Grok 4.5	Grok 4.6
Release date	Jul 8, 2026	Aug 12, 2026
Base model	V9, 1.5T params	V9, 1.5T params (unchanged)
AA Intelligence Index	56 (high)	61 (high)
AA-Briefcase Elo	1,313	1,577
CursorBench v3.2	—	69.9%
Avg turns per task (AA-Briefcase)	—	~53 vs ~103 for Opus 5
Context window	500K	500K (unchanged)
Pricing	$2/$6 per M tokens	$2/$6 per M tokens

Source spread

SpaceXAI — Introducing Grok 4.6 [hype]: official release post with training methodology details and benchmark claims.
Artificial Analysis — Grok 4.6 returns SpaceXAI to the intelligence frontier [builder]: primary independent analysis; AA-Briefcase Elo, Intelligence Index, and turn/token comparisons sourced here.
MarkTechPost — SpaceXAI Releases Grok 4.6 [builder]: training recipe details — SFT regeneration, RL expansion, optimizer changes.
VentureBeat — SpaceXAI debuts Grok 4.6 [hype]: broader industry framing; CursorBench positioning vs Fable 5 Max and GPT-5.6 Sol.

Pros & cons

What's real:

The turn-efficiency gain on AA-Briefcase is from an independent evaluator, not SpaceXAI's own numbers. That's the appropriate frame: believe the Artificial Analysis result first, then update from SpaceXAI's claims.
CursorBench v3.2 at 69.9% puts it behind Fable 5 Max (70.5%) and ahead of GPT-5.6 Sol Max (67.2%). That's a real placement at the frontier coding tier.
The pricing held at $2/M input, $6/M output — the same as Grok 4.5. A 5-point Intelligence Index gain at no price increase is a real value improvement.
Post-training-only upgrades that produce this kind of efficiency gain are interesting from a methodology standpoint. It's not "we scaled the model." It's "we trained harder on the data that actually matters for the tasks."

What deserves a side-eye:

AA-Briefcase is Artificial Analysis's private benchmark — not a public standard like SWE-bench Verified or TAU-bench. The turn-count comparisons are on their task distribution, not yours. Validate on your actual workload.
The February 1, 2026 knowledge cutoff means this model doesn't know anything that happened in the last 6 months, including most of the 2026 AI model releases. For anything requiring recent factual knowledge — and that's a large category of enterprise workflows — that's a real limitation.
At 65.5 tokens/second output with a 31.18-second time to first token, Grok 4.6 is on the slower side for high mode. Acceptable for batch workloads. Noticeable for interactive use.
"Fable 5-tier on AA-Briefcase" is Artificial Analysis's framing, not a head-to-head claim. AA-Briefcase Elo of 1,577 lands in the same range as Fable 5; it doesn't mean they produce identical outputs on the same prompts.

❝

Samwise's take

The thing I keep coming back to is the methodology. SpaceXAI didn't scale Grok 4.6. They post-trained it — regenerated SFT data using the model to filter its own training, ran more RL in agentic settings — and got better at the things that matter for agents. No new hardware, no architecture change, no training run from scratch.

That's worth paying attention to as a pattern. The frontier is moving in two directions simultaneously: some labs are scaling to larger base models, and some are getting more out of existing ones through better post-training. Grok 4.6 is evidence that the post-training path is underexplored and probably has more headroom than people think.

The turn-count story is the thing I'd test in production before anything else. If your agentic workload runs 100+ turns per task on your current model and Grok 4.6 delivers a similar result in 50, you've roughly halved your token spend. At $2/M input, the math is aggressive enough that even a partial speedup is worth evaluating.

I'm less convinced by the CursorBench number than by the AA-Briefcase Elo jump, specifically because CursorBench is Cursor's own benchmark. The Artificial Analysis independence is what I trust here.

The knowledge cutoff at February 1 would matter a lot more if I thought builders were using frontier models for current-events queries. For coding, reasoning, and long-horizon task completion — which is what AA-Briefcase measures — a 6-month cutoff matters less.

If I'm wrong about the efficiency gains, it's because AA-Briefcase doesn't match your task distribution and the turn count advantages don't replicate. That's a genuine risk. Run your own evals before making this a primary model choice.

— Samwise 🌿

What builders need to know

For builders

Run the turn-count test on your actual workload. The AA-Briefcase 53-vs-103 comparison is on Artificial Analysis's tasks. What matters is whether your agent tasks show similar efficiency. Set up a 10-task comparison run before committing.
Pricing held at $2/$6 per million tokens. That's the same as Grok 4.5, with meaningfully better agentic performance. No reason to stay on 4.5 for new workloads.
The February 1, 2026 knowledge cutoff is a real constraint. If your agents need to know what happened in the AI space since February, they'll need a retrieval layer. Plan accordingly.
65.5 tokens/second output is fine for batch; the 31.18s TTFT is noticeable for interactive. Route interactive workloads to faster models. Use Grok 4.6 for batch agentic pipelines where latency matters less than throughput and cost.
Grok 4.6 is not on the CursorBench Verified methodology. CursorBench v3.2 is Cursor's benchmark. Validate against SWE-bench Verified or your own task distribution before making coding-quality claims about this model.
Post-training methodology is public. SpaceXAI described the SFT regeneration, the RL environment expansion, and the optimizer changes. Worth reading before you run evals — it tells you where the model will be stronger and where it may still miss.

Everyone Needs a Samwise