{"slug": "ai-model-costs-could-fall-3-4x-in-six-months-perplexity-ceo-predicts", "title": "AI model costs could fall 3-4x in six months, Perplexity CEO predicts", "summary": "Perplexity CEO Aravind Srinivas predicted on July 11, 2026 that top-tier AI model costs could fall 3-4x within six months and that a locally-runnable Opus-grade model would arrive within 12 months. Y Combinator CEO Garry Tan amplified the forecast, signaling to founders that cheaper frontier inference could enable new product architectures. The prediction, while lacking specific benchmarks or vendor details, serves as a market signal for startups to plan for cost compression and local execution.", "body_md": "[Aravind Srinivas (@aravsrinivas)](https://x.com/aravsrinivas), the co-founder and CEO of [Perplexity](https://www.perplexity.ai/es/hub/blog/perplexity-raises-series-b-funding-round), predicted on July 11th, 2026 that top-tier AI model capability is about to get cheaper and closer to the device, giving founders a concrete timeline to plan against rather than another vague claim about model progress.\n\n[https://twitter.com/aravsrinivas/status/2075831774450770243](https://twitter.com/aravsrinivas/status/2075831774450770243)\n\nIn a [post on X](https://twitter.com/aravsrinivas/status/2075831774450770243), Srinivas said there is a greater than 50% chance that a model with \"fable 5 quality\" will be 3 to 4 times cheaper within six months. He also predicted that an \"Opus 4.8 grade\" model capable of running locally will arrive within 12 months. The six-month window points to roughly January 2027. The 12-month local-device forecast points to roughly July 2027.\n\nYC president and CEO [Garry Tan (@garrytan)](https://x.com/garrytan) pushed the claim into the startup conversation minutes later, writing in a [post on X](https://x.com/garrytan/status/2075841038829400546): \"We are all just getting started.\" Tan's amplification matters because [Y Combinator](https://www.ycombinator.com/people/garry-tan) is one of the main distribution channels through which model-cost assumptions get turned into company formation. If founders believe frontier-class inference will get materially cheaper within two fundraising cycles, they can underwrite products that would have looked too expensive to serve at scale in mid-2026.\n\nThe forecast is still only a forecast. Srinivas did not publish a benchmark, identify the vendor behind the cheaper model, define the evaluation that would make something \"fable 5 quality,\" or show the hardware target for a local \"Opus 4.8 grade\" system. Those omissions matter because AI cost claims often collapse several separate variables into one headline number: list price, output length, caching, latency, batch pricing, tool-use overhead and the number of failed attempts before a task is accepted.\n\nThat is why the post is useful as a market signal rather than as procurement guidance. Srinivas is not a detached commentator. Perplexity sells an AI answer engine, and the company said in January 2024 that it had reached 10 million monthly active users and handled more than half a billion queries in 2023. At that scale, the difference between paying premium rates for each generated answer and routing work to cheaper models is a product and margin question, not a lab argument.\n\nThe local-device portion of the prediction carries the sharper strategic implication. If an Opus-class system can run locally, builders get a different cost structure, lower dependency on hosted APIs and a path to products that keep more data on the user's machine. The tradeoff is that local models have to fit inside consumer hardware constraints while still being reliable enough for work that today is often sent to cloud-hosted systems. Srinivas's post gives no evidence that this tradeoff has been solved. It says he thinks the odds are better than even that it will be solved within a year.\n\nFor startups, the practical takeaway is to separate product architecture from model identity. A founder building on the assumption that today's highest-priced model will remain the only acceptable option is making a bet against cost compression. A founder hard-coding one model provider into every workflow is making the opposite mistake. The more durable design is routing: use stronger models when the task requires them, cheaper models when the risk is low, and local execution when privacy, latency or unit economics justify it.\n\nTan's reaction also shows how quickly model forecasts become founder permission. A short prediction from Srinivas turned into a rallying line from the head of YC because the startup economy has been waiting for this part of the AI cycle. The first wave rewarded wrappers that could prove users wanted AI inside existing workflows. The next wave depends on whether those workflows can be served at a cost that leaves room for real gross margin.\n\nSrinivas's July 11th post does not prove that frontier-class intelligence will be cheap by January 2027, or that local Opus-grade models will be practical by July 2027. It does set a clear expectation from one of the founders operating closest to consumer AI demand: model quality is still rising, and price is the constraint he expects to break next.", "url": "https://wpnews.pro/news/ai-model-costs-could-fall-3-4x-in-six-months-perplexity-ceo-predicts", "canonical_source": "https://runtimewire.com/article/ai-model-costs-3-4x-cheaper-six-months-perplexity-ceo-prediction", "published_at": "2026-07-11 17:17:07+00:00", "updated_at": "2026-07-11 17:17:23.096728+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-startups", "ai-infrastructure", "ai-policy"], "entities": ["Perplexity", "Aravind Srinivas", "Garry Tan", "Y Combinator"], "alternates": {"html": "https://wpnews.pro/news/ai-model-costs-could-fall-3-4x-in-six-months-perplexity-ceo-predicts", "markdown": "https://wpnews.pro/news/ai-model-costs-could-fall-3-4x-in-six-months-perplexity-ceo-predicts.md", "text": "https://wpnews.pro/news/ai-model-costs-could-fall-3-4x-in-six-months-perplexity-ceo-predicts.txt", "jsonld": "https://wpnews.pro/news/ai-model-costs-could-fall-3-4x-in-six-months-perplexity-ceo-predicts.jsonld"}}