Model prices fall 80 percent as OpenAI and Anthropic confront open weight pressure in a rapid inference war

The AI market has seen a rapid decline in inference costs in recent months, as leading language model developers cut API prices by up to 80 percent in some categories.The shift was not a fleeting publicity stunt but a direct pricing battle driven by operational efficiency and free-competition pressures.The wave began in late June 2026 with OpenAI’s launch of the GPT-5.6 Sole model at $5 per million input tokens and $30 per million output tokens, plus $0.50 for temporary memory reads, coinciding with the lightweight Luna model priced at $1 per million input and $6 per million output. Anthropic responded quickly with the launch of Claude Sonnet 5 at a promotional rate that later settled at $2 per million input and $10 per million output.

Reductions accelerated to a unprecedented pace when OpenAI slashed Luna’s price by 80 percent at the end of July, bringing it down to $0.20 per million input, $1.20 per million output, and $0.02 for temporary storage reads to prevent developers from switching to smaller models. Anthropic followed in early September with the announcement of Vibe 5.1, achieving a 75 percent reduction in temporary memory read costs compared with the June version.The core significance of this reduction lies in the fact that temporary storage reads account for between 96 and 99 percent of total token consumption in software agent applications.Since the entire conversation is re-loaded at each inference step, and stored-data retrieval rates are practically between 90 and 100 percent, lowering read costs translates directly into an overall model-running cost decrease of roughly 70 to 75 percent.

Competition peaked on 22 September 2026 when OpenAI released GPT-6 Sole and GPT-6 Luna at prices equal to half those of the prior versions, and Anthropic announced the same day that Claude Opus 5.5 cost 40-60 percent less overall, explicitly stating it requires less compute to serve requests. When the new Sole version proved under-performing, OpenAI corrected course a week later with GPT-6.1 Sole and halved temporary memory read costs to $0.10 per million tokens, matching Anthropic’s five-percent rate.

Behind this heated race, structural economic pressures imposed by open-weight models and independent inference providers are emerging. Data from the OpenRouter platform showed that the updated DeepSeek Flash v4.1 model scored 45 points onBehind this heated race, structural economic pressures imposed by open-weight models and independent inference providers are emerging. Data from the OpenRouter platform showed that the updated DeepSeek Flash v4.1 model scored 45 points on the Artificial Analysis Index, priced from fifteen cents per million input tokens, and its hosting was spread across 29 independent providers such as Cloudflare and DeepInfra, giving developers the freedom to switch instantly based on the lowest price without any technical barrier. MonShot AI also entered the competition with its Kemi K 3.1 Pro model, whose performance approaches that of the Opus 5.5 model at an input cost of only two dollars, and its weights are expected to be released in early November and distributed through more than twenty service providers.

These facts confirm the convergence of the hypothesis that promoted the possibility of future token price spikes on the grounds of monopoly, as data from the “Virsel” portal shows that token cost has fallen to 43.3 percent of its level recorded in April 2026, while competition among dozens of providers pushes prices toward the marginal operating cost.Conversely, corporate moves are not without complex calculations linked to the impending public offering of their shares.Especially with the leaked billion-dollar research and development losses and OpenAI’s reduction of paid-package usage shares on the CodaX platform by 50 percent to offset profit margins.

For technology companies and development teams in Saudi Arabia, the United Arab Emirates, Egypt and Jordan, this shift carries strategic implications that go beyond the immediate budget surplus. The cost of invoking models is no longer a barrier that prevents the deployment of automated software agents for processing banking transactions, indexing massive records, or managing continuous technical support. The practical step that now imposes itself on technology leaders in the region is to renegotiate, immediately, the fixed annual contracts with cloud service providers and to move toward flexible software integration layers that differentiate between closed-source models and open alternatives based on real-time cost, while recognizing that building products on pre-packaged model interfaces without local added value has become a losing gamble given the convergence of capabilities and the erosion of protection margins.