The AI industry spent the past week doing something it rarely does: getting cheaper and smaller. Within hours of each other, Anthropic and OpenAI released lower-cost models aimed at developers watching their inference bills. Meanwhile, China Telecom shipped an agentic large model designed to run on a single consumer-grade graphics card, and a startup put a 1-bit language model on Qualcomm-powered smart glasses. For Pakistan's growing community of freelance developers and AI startups, the takeaway is practical: capable AI is getting dramatically less expensive to run.
Anthropic's Opus 5.5 undercuts the flagship
Anthropic released Claude Opus 5.5 on September 22, just two months after its predecessor. The company says the new model matches Claude Fable 5.1's performance on most tasks while cutting API prices 20 percent, to 4 dollars per million input tokens and 20 dollars per million output tokens. Cache-read costs were cut 60 percent, a meaningful saving for long agentic workloads that reuse context, and Anthropic removed the five-hour usage cap for Pro, Max, Team and Enterprise accounts entirely.
The release arrives a few weeks before Anthropic's expected stock market debut, and the company credited outside evaluators including METR, Frontier Design and NIST's new AI standards center in its system card. Notably, both Anthropic and OpenAI released their cheaper models within days of their executives calling for a slowdown in AI development, a tension the industry will be watching closely.
OpenAI answers with GPT-6 Sol and Luna
OpenAI followed shortly after with GPT-6 Sol and GPT-6 Luna, described as cheaper and faster versions of GPT-6 Astra, the company's most powerful model launched earlier in September. OpenAI claimed the new models handle tasks substantially better than Anthropic's top models, setting up a direct duel over both capability and price.
For context, GPT-6 Astra launched at 10 dollars per million input tokens and 50 dollars per million output tokens, the same headline price as Anthropic's Claude Fable 5.1, so this week's discounts mark a genuine escalation of the price war between the two San Francisco labs.
China's Xing4.0-29B runs agentic AI on one GPU
On September 24, China Telecom AI officially released Xing4.0-29B-A4B, a next-generation lightweight agentic large model. It carries 29 billion total parameters but activates only 4 billion per request, built on a Mixture-of-Experts architecture with a 256K-token context window. The company says it can run locally on a personal consumer-grade graphics card or scale into enterprise production environments.
Unlike conventional chat models, Xing4.0-29B is built for execution: planning multi-step task paths, calling external tools and processing long-context inputs to deliver completed results. The trend is unmistakable. The competitive frontier is shifting from raw parameter counts to cost-efficient, agentic models that work on hardware developers actually own.
Tiny models head for your glasses
At Qualcomm's Snapdragon Summit on September 24, the startup PrismML showed off its 1-bit Bonsai LLM running locally on AI smart glasses built on the Snapdragon AR1 Gen 1 platform. The 2-billion-parameter model is tuned for vision and language, so a wearer could ask what they are looking at in real time without sending data to the cloud.
PrismML, founded by Caltech researchers, specializes in shrinking larger models dramatically, in this case by four times, while retaining nearly all of their performance on standard benchmarks. Its larger goal is open-weight AI that runs on devices people already carry, pitched as an alternative to depending on proprietary labs and their appetite for ever more compute.
Routing APIs and retail AI round out the week
Cost pressure also produced a new kind of product. On September 24, the AI agent company Liner launched its Liner Model API, which automatically routes each request to the most cost-efficient model capable of handling it, instead of sending everything to a frontier model. Liner says its own internal token expenses in August fell more than 50 percent compared with the first half of 2026 after deploying the routing system.
Separately, Amazon reported on September 23 that sellers now accept its Seller Assistant AI's recommendations more than 90 percent of the time. The assistant has reached more than 90 percent of Amazon's selling partners, keeps persistent memory of each seller's pricing and inventory patterns, and is gaining always-on workflows plus a plugin that will carry it into Anthropic's Claude in beta. For Pakistani sellers and exporters using Amazon, that AI is becoming a quiet co-pilot for their storefronts.
Taken together, the week's launches tell one story: the race is no longer only about who builds the smartest model. It is about who delivers usable intelligence at a price, on hardware and in workflows, that the rest of the world can actually afford.
