Anthropic cuts its Opus price by 20 percent and says the new model costs 40 percent less to run
The AI stack on 22 September 2026. 9 moves, 6 with a primary source.
Anthropic cuts its Opus price by 20 percent and says the new model costs 40 percent less to run
· L4 Models · input $5 to $4, output $25 to $20 per 1M tokens; cache reads -60% · Confirmed
Anthropic released Claude Opus 5.5 on 22 September and priced it below the model it follows, Opus 5. Text sent to the model now costs $4 per million tokens instead of $5, and text it writes back costs $20 instead of $25, a 20 percent cut; a token is a small piece of text, a word or part of one. Cache reads, the rate charged when the model reuses text it has already processed, such as a long document an AI agent keeps going back to, fall from 50 cents to 20 cents per million tokens, 60 percent less. Anthropic says cache reads make up most of the cost of agent and coding work, and that because the new model also uses fewer tokens per task, it costs 40 percent less than Opus 5 to run on typical workloads. It also says Opus 5.5 performs at the level of its Claude Fable 5.1 model on most work. Anyone paying for an app built on Claude may see some of that saving passed on, though that is the app maker's choice.
OpenAI's new GPT-6 Sol and Luna models cost about half as much as the GPT-5.6 versions before them
· L4 Models · Sol $4 to $2 in, $20 to $10 out; Luna $0.20 to $0.10 in, $1.20 to $0.50 out · Confirmed
OpenAI released GPT-6 Sol and GPT-6 Luna, two models for work that does not need the full depth of its top model, GPT-6 Astra, and priced them well under the GPT-5.6 versions they follow. For developers, Sol now costs $2 for every million tokens sent in and $10 for every million sent back, down from $4 and $20, and Luna costs 10 cents in and 50 cents out, down from 20 cents and $1.20; a token is a small piece of text, a word or part of one. OpenAI says the older prices it compares against were promotional ones, and that improvements in caching and in how it runs the models let it serve them at lower cost. Both are available from launch in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. For anyone building a product on these models, a major cost has roughly halved, and some of that can reach the people using those products as lower prices.
A Texas company that powers AI data centres tells bond buyers it will more than triple its power fleet by 2029
· L1 Power and land · 3,300+ MW by end-2029, from ~950 MW; $1,000m of notes offered · Confirmed
Solaris Energy Infrastructure, a Houston company that runs power plants on data-centre sites and supplies them directly rather than through the public grid, set out its plans for investors in a $1,000 million bond offering. It says it has firm orders with equipment makers for enough generators to reach more than 3,300 megawatts by the end of 2029, up from about 950 megawatts it ran on average in the three months to 30 June. It says three long-term contracts with investment-grade technology customers already cover about 2,200 megawatts, and that the money from the notes will help pay for contracts it has signed and ones it expects to sign. It is one more power company borrowing to build generation for data centres.
Alibaba Group's chief executive, Eddie Wu, told the company's annual Apsara Conference in Hangzhou on 22 September that its target is for the data centres Alibaba Cloud runs around the world to pass 20 gigawatts of capacity by 2032. A gigawatt is roughly what one large power station produces. This is a stated target, not a signed contract, a budget or a building plan, and the speech does not put a cost on it. It tells you Alibaba means to keep building AI computing for years to come; nothing changes for users of its cloud or AI products now.
More Americans say AI is hurting society: 39 percent, the highest since a Reuters/Ipsos poll began asking
· Band Rules and society · 39% negative (from 36%); 11% positive; 73% say AI firms not doing enough · Reported
A Reuters/Ipsos poll of 1,277 US adults, run online over four days to 20 September, found 39 percent think AI is having a negative effect on society, up from 36 percent a month earlier and the highest share since Reuters and Ipsos began asking in March. Only 11 percent called its effect positive; the rest were unsure or did not answer. Some 73 percent worry AI companies have not done enough to stop AI causing serious harm to society, and 55 percent said slowing AI development would be a good thing, against 13 percent who said it would be bad. The poll's stated margin of error is 3 percentage points, as large as the one-month rise. Public opinion like this is part of the backdrop as lawmakers decide how to regulate AI.
Cognex, which makes machine-vision systems for industry, has agreed to buy RealSense for approximately 500 million dollars, paid from its existing cash and investments. RealSense makes depth-sensing cameras that let robots judge distance; it was founded by Intel in 2014 and spun out in 2025. Cognex expects RealSense to bring in 80 to 90 million dollars of revenue in 2026, more than 50 percent growth on last year, and puts the robotic-perception market at 600 million dollars today, growing to about 1.6 billion dollars by 2030. It also plans a three-year cash retention programme of 56.5 million dollars for RealSense staff, plus about 50 million dollars in restricted stock. For most people nothing changes yet: the deal is expected to close in the last quarter of 2026, and the cameras it buys are used in mobile robots, robot arms and humanoids.
Taiwan's export orders reached 102.96 billion dollars in August, up 71.4 percent on a year earlier and the first month ever above 100 billion dollars, according to the economy ministry's figures as reported by the Taiwanese financial news site cnyes.com. The report credits demand for AI, high-performance computing and cloud services. Orders for electronic products rose 83.9 percent to 45.75 billion dollars and for information and communication products 100.5 percent to 34.21 billion; the United States was the biggest source of orders at 42.13 billion dollars, up 88.9 percent. This desk could not read the ministry's own release, so the item runs as Reported. For most people nothing changes directly, and the ministry expects orders to keep growing in September.
Construction starts on a 640-megawatt AI data centre in West Java, with its first 120 megawatts due in early 2027
· L1 Power and land · 640MW campus; 120MW first phase, early 2027 · Reported
BDx Data Centers, a Singapore-based company, started construction on 22 September on a 640-megawatt AI data centre campus called CGK4 in Jatiluhur, West Java, Indonesia, according to The Jakarta Post. Its first 120-megawatt building is expected to enter service in early 2027. BDx says it has secured 845 megavolt-amperes of grid capacity from PLN, Indonesia's state electricity company, and the first building will use liquid cooling piped directly to the chips, for power densities of up to 500 kilowatts. For most people nothing changes yet; the first building is not due to run until early 2027.
China Telecom's AI unit announced on 22 September that it has released Xing4.0-29B-A4B, an open model with 29 billion parameters, of which only 4 billion are used for each token it produces. The company says that with low-bit quantization, a way of compressing a model, it needs 15 gigabytes of graphics memory, enough to run on a single consumer graphics card, and that it can plan multi-step tasks, call outside tools and handle long documents. It has published the model on GitHub and Hugging Face and says it already runs in China Telecom's own customer-service platform. For developers, it is an agent model that the company says can run entirely on their own computer, with no data leaving it.