Grok 4.6 Hits the Frontier at a Fraction of the Price

What just landed

xAI released Grok 4.6 on 12 August 2026, roughly five weeks after Grok 4.5 — a short gap between frontier releases, and not a cosmetic one.

• 61 on the Artificial Analysis Intelligence Index. Level with GPT-5.6 Sol, two points behind Claude Opus 5 at 63, one behind Claude Fable 5 at 62.
• A fast climb: five points up on Grok 4.5 and twenty-three up on Grok 4.3, on the same index.
• Pricing held flat at $2.00 per 1M input tokens and $6.00 per 1M output, unchanged from Grok 4.5 — against $5 and $30 for GPT-5.6 Sol.
• 500,000-token context window, also unchanged. Modality is text and image input, text output.
• An efficiency gap that flatters it further: roughly 53 turns per task against roughly 103 for Claude Opus 5 at maximum reasoning, and about a quarter of the input tokens across the same evaluation run.

Artificial Analysis is blunt about its own framing: the selling point is price-to-intelligence, not raw leadership. Grok 4.6 does not top the index. It arrives at the top table and charges considerably less than the models either side of it — which deserves a different kind of analysis than the usual leaderboard theatre.

What "61 on the index" does and does not tell you

A composite index is not your workload

The index is a composite: a basket of evaluations averaged into one number so models can be ranked on a single axis. That is useful for tracking the frontier and misleading as a forecast of how a model will behave inside your product. Composite scores compress variance — two models can reach the same number by being good at completely different things, and the sub-scores behind the average are far more informative than the average. The index tells you a model is in the conversation. It does not tell you it is right for you.

Strong on knowledge work, weak on terminal use

Grok 4.6's reported strengths are knowledge work and legal reasoning. Its reported weakness is terminal use. Those three words are the routing decision for most teams, and they cut in opposite directions.

If you are building a coding agent — something that plans, runs shell commands, reads the output, recovers from a failed build and tries again — terminal-use weakness is close to disqualifying for the autonomous tier. A model that fumbles command construction does not simply give a worse answer; it produces a longer, dearer loop that ends in a worse answer. That distance between benchmark scores and real agentic reliability is the territory the DeepSeek V4 Flash agent benchmarks made visible, and it has not narrowed.

If you are building a legal-research assistant, a contract-review tool or an internal knowledge assistant, terminal use is irrelevant to you and the reported strengths land exactly where your product lives. Same model, same 61, opposite verdicts. Which is why "which model is best" is a badly formed question, and "best for this task family, at this price, at this reliability threshold" is the one worth asking.

The price gap, and what we could actually verify

Here is the comparison, restricted to figures we verified for this piece. Where we could not confirm a number, the cell is an em dash rather than an estimate.

Model	AA Intelligence Index	Input $/1M	Output $/1M	Context
Grok 4.6	61	$2.00	$6.00	500,000
GPT-5.6 Sol	61	$5.00	$30.00	—
Claude Fable 5	62	—	—	—
Claude Opus 5	63	—	—	—

Only Grok 4.6's pricing and context window, and GPT-5.6 Sol's list pricing, were verified for this article; index scores for all four come from the Artificial Analysis Intelligence Index. The remaining cells are deliberately blank — a wrong number in a cost table is worse than no number.

On input, $2.00 against $5.00 is roughly sixty per cent cheaper. On output the gap is far wider: $6.00 against $30.00. For a chat product the input gap dominates; for anything generating long documents, code or structured reports, the output gap is where the money actually is. This is the same price-to-intelligence pressure Gemini 3.6 Flash brought to the subagent tier, arriving now at the frontier tier rather than beneath it.

The 200K billing cliff nobody put in the headline

Now the part that will catch teams out. Prompts of 200,000 tokens or more are billed at $4.00 input and $12.00 output per 1M — and critically, that repricing applies across the entire request, not only to the tokens above the threshold.

This is not how engineers instinctively expect tiered pricing to work. The mental model carried over from cloud billing is marginal: you pay the higher rate on the overage. Here it is retroactive to token one. And because $4/$12 is exactly double $2/$6, the arithmetic is unusually clean and unusually unforgiving — crossing the line costs precisely twice what the same request would have cost sitting just underneath it.

Three properties make this dangerous rather than merely annoying. It is invisible in code review, because no diff says "this doubles the bill". It is non-deterministic in RAG systems, because retrieved context length varies per query — you sit under the threshold for most requests and over it on the long tail. And agentic systems drift towards it naturally, because a long-horizon agent accumulates context as it works and crosses 200,000 tokens mid-run without anyone deciding that it should.

The mitigations are unglamorous. Instrument prompt length as a first-class metric and alert at 180,000 tokens, not 200,000. Hard-cap assembled context below the threshold and truncate deliberately rather than letting the retriever decide. And for genuinely long jobs, choose explicitly rather than by accident — our breakdown of long context windows versus retrieval is the right framework, and this pricing argues for tighter retrieval over lazily stuffing the window.

Turns and tokens: the number that beats the sticker price

Why turn count multiplies token cost

The efficiency data is the most interesting thing in this release and is barely mentioned in the coverage. Grok 4.6 resolves tasks in roughly 53 turns on average, against roughly 103 for Claude Opus 5 at maximum reasoning. Across the evaluation run, Artificial Analysis logs about 0.5B input tokens for Grok 4.6 against about 2.0B for Opus 5 — a run-level total, not a per-task figure.

Note that the token gap is wider than the turn gap: about twice the turns, four times the input tokens. That is not a coincidence. In a long-horizon agent loop each turn re-sends the accumulated conversation — system prompt, tool definitions, every prior tool call and every prior result. Turn twenty pays for turns one to nineteen; turn fifty pays for one to forty-nine. Input consumption therefore grows roughly with the square of turn count rather than linearly: double the turns an agent needs and you quadruple the input bill. Prompt caching blunts this where the prefix is stable, but the tail of a trajectory is by definition the part that keeps changing, so caching softens the curve rather than flattening it.

What 53 versus 103 turns actually buys

This is why efficiency matters more than per-token price for agentic work, and why the two compound rather than merely add: a model that is cheaper per token and needs fewer turns wins on two multiplied axes at once. A four-fold difference in token consumption deserves more attention than a three-dollar difference in list price.

The caveat is real: these are averages across an evaluation suite, and fewer turns is only a virtue if the work gets finished. A model that gives up early looks wonderfully efficient until you measure success rate. Turn count belongs in a table beside accuracy and escalation rate, never on its own.

Rupee revenue, sterling runway

The India arithmetic

For an Indian startup earning in rupees and paying for inference in dollars, model pricing is a gross-margin question before it is an engineering one. A product priced for the Indian market cannot absorb frontier-tier token costs the way a US enterprise tool on dollar contracts can, so a frontier-class model at $2 input rather than $5 is often the difference between a feature that ships and one that stays in the backlog. The cliff bites hardest here too: thin margins leave the least slack to absorb a surprise doubling, and small teams are the least likely to have anyone watching for it.

The UK arithmetic

For a UK seed-stage team the pressure is shaped differently and points the same way. The constraint is runway rather than margin: an eighteen-month cash position, a sterling burn rate, and investors who now expect AI cost of goods sold broken out in the board pack rather than buried in hosting. The trap in both markets is identical — a headline price cut tempts teams to route more traffic to the frontier tier because it suddenly feels affordable, and total spend rises rather than falls. Cheaper tokens are only a saving if volume holds constant, which it rarely does.

What a builder should actually do

This is a routing-tier decision, not a migration. A two-point index gap does not justify rewriting a working system, and a sixty-per-cent price cut does not justify moving traffic you have not measured.

1. Add it as a tier, not a replacement. Put Grok 4.6 into an existing cascade alongside what you run today. If you have no cascade, that is the real gap — our guide to model routing cascades covers the pattern and how to set escalation thresholds that do not silently send everything to the expensive tier.
2. Measure on your evaluations, not the index. Route a slice of genuine production traffic and compare accuracy, escalation rate and cost per completed task. If you cannot answer "did quality drop" with a number, you are not ready to route anything anywhere.
3. Split by task family. Knowledge work, document reasoning and legal-style analysis first. Terminal-heavy coding agents last, and only with evidence.
4. Instrument prompt length before routing a single request. Alert below 180,000 tokens. The cliff is the likeliest way this release costs you money instead of saving it.
5. Keep the expensive tier wired in. The point of a cascade is that the top tier is there when the cheap tier fails, not deleted because month one looked good.

Remember how fast this ground moves. When Grok 4.5 reached general availability the framing was that xAI had a credible frontier entrant; five weeks later it has one at a fraction of a competitor's input price. Build the routing layer so that next week's release is a config change rather than a project.

The bottom line

Grok 4.6 does not lead the index and does not claim to. It arrives at 61 — level with GPT-5.6 Sol, within two points of Claude Opus 5 — while charging $2 and $6 where Sol charges $5 and $30, and while finishing tasks in roughly half the turns of Opus 5 at maximum reasoning, and consuming about a quarter of the input tokens over the same evaluation run. For a large class of workloads that combination matters more than the two points it gives away at the top.

The caveats are specific. Terminal use is a reported weakness, which rules it out of the autonomous tier for many coding agents, and the 200,000-token repricing is a live hazard for RAG and long-context systems. Add it to the cascade, point your evaluation suite at it, put an alert on prompt length — then let your own numbers decide, in Chennai or in Cambridge, which tier deserves the traffic.

Benchmark, pricing and efficiency data come from Artificial Analysis' Grok 4.6 model page and its Grok 4.6 benchmarks and analysis. Figures for models other than Grok 4.6 and GPT-5.6 Sol are limited to index scores.