Grok 4.6 Ties GPT-5.6 Sol and Costs 60 Percent Less

The gap at the top of the model rankings has almost closed. The gap in what those models charge has not.

SpaceXAI released Grok 4.6 on August 12. On the Artificial Analysis intelligence index it scores 61 points, exactly level with OpenAI's GPT-5.6 Sol. It costs $2 per million input tokens and $6 per million output tokens. GPT-5.6 Sol costs $5 and $30.

Where it lands on the board

Sixty-one points puts Grok 4.6 five points ahead of its own predecessor, Grok 4.5, which is a substantial jump for a point release.

Anthropic still holds the top two slots, with Claude Opus 5 at 63 and Fable 5 at 62. That is the whole spread: two points separate the best model available from a model costing a fraction as much. Anyone choosing a provider on raw capability alone is now optimizing a variable that barely moves.

Step count is the number people miss

On GDPval-AA v2, which measures agentic work rather than single answers, Grok 4.6 places second with an Elo of 1,753. The more interesting figure sits next to it: the model finishes complex tasks in roughly 53 steps. Claude Opus 5 takes around 103.

In agent architectures every step is another model call, so step count multiplies against token price rather than sitting beside it. A model that is cheaper per token and finishes in half the steps compounds the difference twice over. This is the line item that surprises teams when the first month's agent bill arrives.

What the spread looks like on a real invoice

Take a mid-sized support assistant burning 10 million input and 2 million output tokens a month. The monthly cost comes out roughly like this:

• Grok 4.6: about $32
• Claude Opus 5: about $100
• GPT-5.6 Sol: about $110

At that volume nobody reorganizes a roadmap over $78. The picture changes when the same assistant runs across ten workflows, or when the work is document-heavy and token consumption climbs by an order of magnitude. Then the annual difference stops being a rounding error.

Access is available through the API, Cursor, Grok Build, and partners including OpenRouter, Vercel and Cloudflare.

Before you switch on price alone

Benchmark position and your position are different measurements, and confusing them is the most common expensive mistake we see.

An index score reflects performance on a fixed public test set. Your workload is not that test set. Domain vocabulary, non-English input, your own document formats and your particular failure tolerance can reorder the rankings completely. We have run comparisons where the model sitting third on the public board was clearly first on the client's actual support tickets.

There is a second question price tables never answer: where the data is processed, whether it is retained, and whether it can be used for training. A model that saves you seventy dollars a month while moving regulated data somewhere your compliance review has not approved is not a saving. It is an unpriced liability.

The method we recommend takes a day. Pull 50 to 100 real examples from your own system, run the identical prompt through three models, and score the outputs side by side against criteria you write down first. Do it before you look at pricing, so the scores stay honest. Only then does cost become a meaningful tiebreaker, and at that point you are comparing models on the only benchmark that bills you.

The durable takeaway from this release is not Grok 4.6 itself. It is that capability is converging while pricing diverges, and that combination rewards teams who kept their provider choice swappable. If switching models means rewriting business logic, this news is not yet good news for you.

Sources: The Decoder, Artificial Analysis