Tokens got cheaper this week. Agents got a meter.

Two frontier labs cut model prices on the same Tuesday, and two platform companies spent the rest of the week putting agent work on a meter. Read together, the week's announcements say something simple: the unit price of a model call keeps falling, but the number of calls an agent makes, and the way they are billed, is where the real cost now sits.

Two price cuts, 90 minutes apart

On 22 September Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output, down from $5 and $25 for Opus 5, with cache reads falling from $0.50 to $0.20. Anthropic says the model runs about 40% cheaper than Opus 5 on typical workloads, generates output more than 30% faster and performs at the level of the larger Fable 5.1 on most work. Those are Anthropic's own measurements. The upgrade is not free of friction: thinking can no longer be switched off on Opus 5.5.

OpenAI followed the same day with GPT-6 Sol and GPT-6 Luna, two reasoning models that sit below GPT-6 Astra. Its API changelog lists Sol at $2 input, $0.20 cached input and $10 output per million tokens, and Luna at $0.10, $0.01 and $0.50, for prompts up to 272K tokens. TechCrunch reports that is half the price of the GPT-5.6 versions and that Anthropic's release landed about 90 minutes earlier. OpenAI says Sol makes about half as many mistakes as its predecessor and outperforms Anthropic's top models; no independent evaluation backs either claim yet. On 25 September OpenAI also posted a fix for an image-encoding defect that had degraded both models' vision performance, and asked image-heavy users to retest.

Then the meters arrived

On 25 September Microsoft reorganized Copilot around Home, Code and Autopilot, the persistent agent previously called Scout. The pricing split matters more than the names. Everyday chat and Office work stays on the per-user license. Cowork, Code, Autopilot and frontier models such as Astra and Fable run on usage-based billing, measured in Copilot Credits, with admin spending policies that by default apply automatically to new services.

Google's Live Avatar for Gemini 3.8 Live, available in Gemini Enterprise from 24 September, shows the same pattern at the API level. Its pricing page converts avatar video output at 6,192 tokens per second, billed at $1.00 per million, which works out to roughly $0.37 per minute of a speaking avatar for the video alone. Google does not bill idle listening time, but it does re-process and re-bill earlier turns of a session on every new turn, up to the context limit.

The compute underneath

Akamai announced on 24 September that Anthropic has committed about $11.6 billion over seven years to its cloud, aimed at CPU workloads rather than GPUs, with an option to expand by up to $9 billion more. Akamai issued Anthropic a warrant for up to about 5% of its stock, vesting as spending grows, and the commitment depends on Akamai meeting delivery and availability requirements. TechCrunch ties the CPU focus to agents, which spend much of their time running code and browsing rather than doing inference. Earlier in the week ABB launched Infinitus, a direct-current power portfolio for AI data centers. The efficiency figure it leads with, more than 5%, comes from a report ABB co-produced with BCG, and the release names no customers or delivery dates.

A ruling that cuts the other way

On 25 September the D.C. Circuit upheld, 2-1, the Department of War's exclusion of Claude from its supply chain under the Federal Acquisition Supply Chain Security Act. Judge Katsas wrote that the department had ample support for treating Claude as a supply-chain risk after Anthropic refused to relax contract prohibitions on lethal autonomous warfare and domestic surveillance, pointing to restrictions built into the model that had stopped it from completing government users' requests. Judge Henderson dissented. The practical reading: a vendor's usage policy can itself count as the risk. Anthropic had not posted a response on its newsroom when we checked.

Also this week

• Anthropic reported that roughly 950 Claude agents flagged a new CRISPR-like enzyme system, ART, in 21 hours. Human scientists did the lab work, and the system's function is still unknown.
• OpenAI formed an unpaid mathematics advisory group hosted at the Institute for Advanced Study. TechCrunch notes its charter excludes advising on research pace, and OpenAI's claim of 100+ newly resolved open problems has not been independently verified.
• Hugging Face shipped GGUF loading in transformers with weights kept quantized, for now only on Apple Silicon and starting with Qwen3.5, and still recommends llama.cpp for efficient local inference.

What it means if you build with these models

Per-token list prices are no longer the number to budget on. A cheaper model that runs more turns, re-bills its accumulated context or streams video can cost more than the expensive one it replaced. Before switching, measure cost per completed task on your own workloads, check whether caching applies (Opus 5.5 made it cheaper; the Live API lists no cached-input rate), and read the migration notes: mandatory thinking on Opus 5.5 changes latency and output-token spend.