AI model prices collapse as attention turns to Gemini 3.5

The headline that AI model prices are collapsing is true for one side of the market and false for the other. On the cutting side, DeepSeek made a 75% price reduction to its flagship V4-Pro permanent, landing it near $0.44 per million input tokens and $0.87 output - numbers that undercut GPT-5, Opus 4.7, and Gemini Flash by a wide margin ^[1]. Anthropic followed the gravity, dropping Opus from $15 per million to $5 per million, a roughly 67% cut ^[2]. OpenAI, per Wall Street Journal reporting, is weighing steep cuts of its own ahead of a confidential IPO filing ^[3].

But the model everyone was waiting on moved the opposite way. Google's Gemini 3.5 Flash launched at $1.50 per million input and $9 per million output - roughly a 3x increase over the Gemini 3 Flash tier it replaced ^[1]. One report frames this as the era of subsidized Western AI pricing coming to an end, arguing those low prices were never sustainable ^[1]. So the accurate picture is a two-sided market: open-weight labs are collapsing the floor while frontier tiers quietly lift the ceiling. Calling it a uniform collapse misses the actual competitive geometry - the cheap tokens and the expensive tokens are moving apart at the same time.

The most under-appreciated fact in this cycle is that the price of a token and the size of your bill are decoupling. Inference cost for a fixed capability has fallen about 95% over two years - GPT-4-quality output dropped from around $30 per million input tokens in 2023 to under $0.50 per million in 2026 ^[4]. Epoch AI puts the broader trend at a median of roughly 50x cheaper per year, rising to 200x for benchmarks measured after January 2024 ^[5]. By any per-unit measure, intelligence has never been cheaper.

And yet the bills are ballooning. Average enterprise AI budgets reportedly grew from about $1.2M per year in 2024 to $7M per year in 2026 as agentic workloads consume far more tokens per task ^[6]. That is the paradox: each token is nearly free, but a single agentic 'answer' now spends thousands of them chaining tool calls, retries, and long-context reasoning. The result is a documented shift CNBC reported from 'tokenmaxxing' - throwing tokens at every problem - toward hard efficiency, with buyers metering usage and routing cheaper workloads to the lowest-cost provider ^[7]. For OpenAI and Anthropic, that behavioral turn is a revenue risk even while their sticker prices drop ^[7].

Three forces converged to make mid-2026 the flashpoint. The first is open-weight competition: Chinese labs are shipping frontier-comparable models at a fraction of the cost - DeepSeek roughly one-thirteenth the cost of closed alternatives - which forces Western providers to cut or cede share ^[3]. The second is the pending liquidity event. Both OpenAI and Anthropic have filed confidentially for IPOs, and the numbers underneath are stark: OpenAI's Q1 2026 adjusted operating margin was reported at -122%, and its share of global generative-AI web traffic fell from 77.6% in May 2025 to 53.7% by April 2026 ^[3]. Defending market share before going public is now a strategic imperative.

The third force is anticipation itself. As the industry waits on Google's next Gemini release, labs are aggressively pricing new models and floating cuts to lock in developers before the launch reshuffles the board ^[3]. Underneath the pricing theater, the cost curve is being pushed by real engineering - Epoch AI attributes the declines to hardware moves like H100 to Blackwell and algorithmic efficiency such as quantization, speculative decoding, and mixture-of-experts routing ^[5]. The competitive trigger is commercial; the ammunition is technical.

Away from the benchmark charts, the practitioner community is reading this less as a price collapse and more as a fight nobody is yet winning. The dominant view among developers is that headline per-token prices are still set below true cost, which is why the same providers face pressure to raise limits or quietly cut usage caps - a 'collapse' that partly masks the fact that no one is clearly profitable. Delphi Ventures' Tommy Shaughnessy captured the underlying economics bluntly, arguing the $20-per-month flat fee was always priced below what heavy usage actually costs ^[3].

The more strategic community read is that the weapon is not the token price at all - it is bundling and total cost of ownership. The recurring argument is that Google can pair a strong model with Workspace, storage, and other consumer products at a single subscription price, pulling users away from standalone AI subscriptions on economics no pure-play lab can match. Product-operator voices frame Google's edge as owning the full stack - chips, data centers, and search cash flow - which lets it compete at both the cheap and premium ends simultaneously. The tension running through the discussion is genuine enthusiasm that competition is finally forcing lower prices, paired with hard skepticism that today's numbers reflect sustainable economics rather than a subsidized land grab.