Grok 4.6 in line with top-tier models at 60% lower prices: the race shifts to computing cost

The comparison between large language models is shifting ground. For two years the question has been which system answers better; in summer 2026 the question is how much it costs to get that answer. The clearest case is Grok 4.6: in measurements by Artificial Analysis picked up by the MarketingProfs roundup, the model scores 61 points on the intelligence index, the same level as GPT-5.6 Sol, with performance judged solid even on tasks requiring the autonomous execution of action sequences.

The difference lies in the pricing. The stated prices are $2 per million input text units processed and $6 per million output units: over 60% below comparable offerings from Anthropic and OpenAI. This figure should be read for what it is — a list price announced by the manufacturer, not an industrial cost verifiable from outside — while the score comes from a third-party evaluator. The distinction matters: on the first number the provider has full discretion, and nothing guarantees the gap will remain as such once computing capacity becomes more expensive or scarcer.

The context is a market that is crowding quickly. The release log compiled by LLM Gateway counts twelve new models distributed by seven providers in August 2026 alone; the latest recorded is GLM-5.2 Turbo from the Chinese company Z.AI, released on August 17. With so many systems placing within a few points of each other on evaluation scales, the remaining competitive lever is price per processed volume: those buying capacity to integrate it into a product choose based on how much they spend per million units, not on half a point of index.

In the same month, however, came a reminder about limits. The comparative test called “Reconstruction,” reported by AI Weekly, asks models to trace back the ideas contained in a scientific article starting only from the bibliography. Leading systems used individually succeed in a share ranging between 3 and 15% of cases; an architecture that has multiple agents work in sequence reaches 42%. These are two figures to be handled with caution: a test measures what it was designed to measure, and reconstructing a hypothesis from cited sources is an artificial task, not the activity of a researcher. What remains is that, on the type of generative reasoning most often invoked in commercial promises, the gap between an isolated model and an organized structure is an order of magnitude.

On the revenue side, the same roundup records the second leg of the strategy: OpenAI will bring advertising into ChatGPT in 31 European markets starting August 24, 2026, extending an activity launched in the United States six months earlier. Initially, ad space is purchased through major agency groups, with direct sales planned at a later stage. The company states it has adapted its privacy rules to the European data protection regulation and maintains that conversations are not shared with advertisers; it also states that about 20% of user requests have a direct commercial purpose. These are statements from the interested party, not independent verifications, and should be considered as such until an authority or a third party verifies them.

The two news items describe the same pressure from opposite sides: the price of access to models is falling, and the search for revenue is shifting toward the content shown to users. The calendar is set: on August 24, advertising appears in 31 European countries.

Sources: MarketingProfs, AI Weekly, LLM Gateway.