Model prices crash 80% as Gemini surpasses one billion users

The language model market, in August, is moving along two axes at once: price and scale of use. On the first, the most aggressive decision dates back to July 30, when the cost of GPT-5.6 Luna was cut by 80%, from 1 to 0.20 dollars per million units of text processed. The direct consequence is that this model became ChatGPT’s free default: a capability that until a few weeks earlier was paid-for is now placed at the base of the free offering.

On the second axis, the number comes from Google: Gemini surpassed one billion monthly active users on August 11, 2026. This is the figure that, in this sector, measures distribution more than quality — and distribution, when models are roughly equivalent, is the competitive advantage that matters most.

Not everyone, however, is lowering their price lists. On August 14 DeepSeek raised the price of V4 Flash by 93%, bringing it from 0.14 to 0.27 dollars per million units processed. The new price list is thus above the 0.20 dollars of GPT-5.6 Luna, and the direction of the move is opposite to that of the rest of the market: it is the first recent case of a provider deciding to charge more instead of following the downward trend.

Operator	Move	Date
OpenAI	GPT-5.6 Luna from 1 to 0.20 dollars per million units (−80%)	July 30
Google	Gemini surpasses one billion monthly users	August 11
xAI	Grok 4.6, score of 61 on the Artificial Analysis Intelligence Index	August 12
DeepSeek	V4 Flash from 0.14 to 0.27 dollars per million units (+93%)	August 14

xAI’s launch on August 12 deserves a separate note: Grok 4.6 arrives with a score of 61 on the Artificial Analysis Intelligence Index and at the same price as its predecessor Grok 4.5. Superior performance at unchanged cost is, from the buyer’s point of view, an implicit price cut.

There is also the speed variable. OpenAI has announced a service tier called Ultrafast, based on Cerebras hardware, which runs GPT-5.6 Sol up to 14 times faster and generates up to 750 units of text per second. This is a figure that shifts competition from cost per unit processed to response time: for interactive applications, latency matters as much as price.

Above this shifting market sits an announcement of a different nature. On August 1, 2026 OpenAI presented the Astra family of models, stating that an internal version had solved ten open problems in mathematics and theoretical computer science. The formal proofs were published in Lean on GitHub — that is, in a language that allows automatic verification of logical steps, which makes the claim checkable by third parties rather than self-certified. Among the results are a construction proving the existence of non-sofic groups and new limits on sphere packing.

The most widely cited comment is that of Fields Medalist Timothy Gowers, who stated that he would unhesitatingly recommend one of those proofs for a top-tier journal. This is an authoritative assessment and should be taken seriously; it is also an assessment of a single result, not of the entire package.

Two limitations must be stated with the same clarity as the announcement. The first: the version of Astra used for those results has not been publicly released, so no one outside the company can reproduce the experiment under the same conditions. The second: industry analyses note that the results are concentrated in areas particularly suited to these models’ strengths and do not amount to general intelligence.

The picture that emerges is coherent. On the commercial front, prices are falling and distribution is concentrating; on the research front, the most striking announcements concern models that cannot be used. The two things are not contradictory: the product and the showcase, in this sector, have by now become separate objects.

Sources: AIToolsRecap; Releasebot — OpenAI release notes; Build Fast with AI; Kraviona.