Grok 4.6 Is 60% Cheaper Than Opus 5. Until Token 200,000.

TL;DR

SpaceXAI released Grok 4.6 today, and it puts the company back on the frontier: 61 on the Artificial Analysis Intelligence Index, tied with OpenAI's GPT-5.6 Sol, two points behind Anthropic's Claude Opus 5, at $2 per million input tokens and $6 per million output. Against Opus 5 at $5/$25 that is 60% off input and 76% off output. The footnote lives in the pricing table: once a prompt reaches 200,000 tokens, every token in that request bills at $4/$12 instead. The model is sold for long-running agents and carries a 500,000-token context window, which means 60% of the advertised window sits on the far side of a price cliff.

Where it actually lands

The Intelligence Index is a composite of nine benchmarks, and the top of it is now genuinely crowded. Grok 4.6 scores 61. So does GPT-5.6 Sol. Claude Fable 5 is at 62 and Claude Opus 5 at 63, both in their maximum-reasoning configurations. Grok 4.5 sat at 56, so this is a five-point jump in one release.

Four models inside two index points is a statistical shrug, and you should read it that way. What separates them now is not raw capability but what you pay and how many tokens you burn getting there.

The vendor's own table admits second place

SpaceXAI's launch table is unusually candid about where Grok 4.6 does not win. On DeepSWE v1.1 it posts 65.9% against GPT-5.6 Sol Max at 73.0%. On CursorBench v3.2 it is 69.9% to Fable 5 Max's 70.5%. On FrontierCode v1.1 it is 61.3% to Fable 5 Max's 63.6%. On APEX-Agents it is 57.5% to Fable 5 Max's 59.2%.

Where it does lead is knowledge work. On GDPval-AA v2, which scores real-world professional tasks by Elo, Grok 4.6 lands at 1753 against Fable 5 Max's 1741, second only to Opus 5. Artificial Analysis also has it in the top two on the tau-cubed Banking agentic benchmark at 50.7%.

One caution before you screenshot any of this. Two Terminal-Bench numbers are circulating for the same model: 88.4% on v2.1, and 26.0% on v3.0. Both are real. They are different benchmark versions with different task sets, and v3.0 is brutal across the board. Benchmark numbers without version tags should be treated roughly like a stock chart with no y-axis.

The price, and the line running through it

Grok 4.6 is $2 input, $6 output, $0.50 for cache hits. That cache rate is a quiet increase: Grok 4.5 charged $0.30. Against Opus 5's $5/$25 and GPT-5.6 Sol's $5/$30, the sticker is the most aggressive on the frontier.

Then you read xAI's model docs. Requests whose prompt reaches 200,000 tokens are billed at the long-context tier: $4 input, $1 cached, $12 output. And the rule is not marginal. The higher rate applies to all tokens in that request, not just the ones past the threshold.

Work the arithmetic on a single call. A 199,000-token prompt costs about $0.40 in input. A 200,000-token prompt costs $0.80. You did not pay double for the extra thousand tokens. You paid double for all of them, retroactively.

The closest everyday version is a parking garage that charges $10 an hour up to four hours and $20 after, then rebills your entire stay at $20 the moment the fourth hour ticks over. You did not buy one expensive hour. You bought an expensive afternoon.

Above the line Grok 4.6 is still the cheaper model, to be clear. The gap against Opus 5 just narrows from 60% to 20% on input, and from 76% to 52% on output. A model marketed on staying with a task across many steps has a pricing page quietly hoping you take fewer of them.

Fewer turns is the actual argument

Which is why the efficiency numbers matter more than the index score. On Artificial Analysis's AA-Briefcase agentic evaluation, Grok 4.6 resolves tasks in roughly 53 turns and about 0.5 billion input tokens on average. Opus 5 at max configuration takes roughly 103 turns and about 2.0 billion input tokens for the same work.

Four times fewer input tokens at 40% of the input price compounds into something much larger than a two-point index gap. It also has a second effect that is easy to miss: a run that finishes in 53 turns is far less likely to drag a single request past 200,000 tokens in the first place. The efficiency and the pricing cliff are the same story told twice.

Where to get it

The model string is grok-4.6, live now on the xAI API through both the Responses API and Chat Completions, with function calling, web search, X search, and code execution as built-in tools. It also shipped day one in Cursor and Grok Build, plus OpenRouter, Vercel, and Cloudflare. There is a fast variant at twice the price, and SpaceXAI is running doubled included usage in Grok Build and Cursor for the first week.

Context stays at 500,000 tokens, unchanged from Grok 4.5. This release is an agent and coding upgrade, not a window expansion and not a price cut.

What to check before you switch

• Instrument prompt size, not just token spend. If your agent's context routinely crosses 200K, your effective rate is $4/$12 and your cost model is wrong by 2x.
• Watch the cache line. Cache hits went from $0.30 to $0.50 per million. Heavy prefix-caching workloads are the one place Grok 4.6 got more expensive than its predecessor outright.
• The coding table is vendor-reported. DeepSWE, CursorBench, FrontierCode, and APEX-Agents figures come from SpaceXAI's own launch post. The index, turn counts, and token counts are Artificial Analysis's independent runs.
• Second place on code is still second place. If your workload is pure software engineering, GPT-5.6 Sol Max and Fable 5 Max lead on every coding benchmark in the launch table.

Key Takeaways

• Grok 4.6 launched August 12, 2026, scoring 61 on the Artificial Analysis Intelligence Index, tied with GPT-5.6 Sol, behind Claude Opus 5 at 63 and Claude Fable 5 at 62, and five points above Grok 4.5.
• Base pricing is $2 input, $6 output, $0.50 cached per million tokens, against Opus 5 at $5/$25 and GPT-5.6 Sol at $5/$30.
• At 200,000 prompt tokens the long-context tier kicks in at $4/$12, applied to every token in the request, not just the overage. The context window is 500,000 tokens.
• Efficiency is the strongest claim: roughly 53 turns and 0.5B input tokens on AA-Briefcase versus roughly 103 turns and 2.0B for Opus 5 at max.
• SpaceXAI's own table concedes coding: 65.9% on DeepSWE v1.1 against GPT-5.6 Sol Max's 73.0%, and losses on CursorBench, FrontierCode, and APEX-Agents.
• Available now as grok-4.6 via the xAI API, Cursor, Grok Build, OpenRouter, Vercel, and Cloudflare, with a fast variant at double the price.

Sources: SpaceXAI, Introducing Grok 4.6, xAI model and pricing docs, xAI Grok 4.6 developer docs, Artificial Analysis benchmarks and analysis, OpenRouter model page, The Decoder, Kingy AI pricing breakdown