Table of Contents
Welcome back to AI News Friday.
The pitch this week was surprisingly practical: stronger models, cheaper tokens, and more choices about where to run them. That is the good news. The less comfortable news is that when an AI system carries notes from one part of a task to the next, those notes can become a security problem. I care about both halves. A great model is more useful when I can afford it, inspect my options, and trust the workflow around it.
1. Anthropic takes the lead and cuts the Opus bill
Claude Opus 5.5 arrived Tuesday. Anthropic says it performs around the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5 on typical workloads. The list price is $4 per million input tokens and $20 per million output tokens, a 20% cut from Opus 5. Cache reads are $0.20 per million tokens, which matters if your coding agent keeps revisiting the same context.
There is an outside measurement, too. Artificial Analysis scored Opus 5.5 at 58 on its Intelligence Index at max effort, its highest measured score. But the same evaluation found it generated substantially more output tokens than Opus 5 per Index task. A lower token price does not automatically mean every job costs less. Check the actual bill for your workload.
Kenny’s Take: I like this version of the model race. Give me better work at a price that lets me use it more than twice without checking my bank account. But I am not choosing a coding model from one leaderboard number. If it solves my task, how long does it take, and what does the whole run cost? Those are the numbers I want.
2. OpenAI’s Sol and Luna make price the headline
OpenAI released GPT-6 Sol for more involved coding and agent tasks and GPT-6 Luna for focused, high-volume work. Standard API prices are $2 input and $10 output per million tokens for Sol; Luna is $0.10 input and $0.50 output. Both have a 1,050,000-token context window, though long-input pricing rules kick in above 272,000 input tokens.
Artificial Analysis found the two models broadly level with their GPT-5.6 counterparts on its headline intelligence and coding indices while costing less per evaluated task. The details are mixed: Sol improved on its Coding Agent Index, while Luna slipped. The same report found fewer hallucinated answers on one test, partly because Sol declined to answer more questions. That is a tradeoff worth knowing before calling it a clean intelligence upgrade.
Kenny’s Take: Half-price is a real feature. So is knowing when the cheap model quietly drops a step. I would use Luna for repetitive work I can check quickly and pay for Sol where a bad answer burns more time than the model saves. “Newest” is a lousy routing policy. Cost per finished, correct task is better.
3. Xiaomi gives the open-weight camp a serious contender
Xiaomi’s MiMo-V2.6-Pro scored 46 on the Artificial Analysis Intelligence Index, where it ranks first among the open-weight models that service tracks. Its model page lists an MIT license, a one-million-token context window, and support for text, image, speech, and video input. Xiaomi has also posted the Pro-RL checkpoint and model card for download.
Open weights do not mean cheap home deployment. Xiaomi’s own model card describes a sparse model with 1.02 trillion total parameters and 42 billion active parameters. You can access it through a provider or plan for serious hardware if you want to host it yourself. Either way, being able to obtain weights changes the conversation for developers who care about control and portability.
Kenny’s Take: I am always interested when a model this capable gives developers an exit door from a single API vendor. But “you can run it yourself” needs an asterisk the size of a server rack. Open weights are a meaningful choice even if most of us will rent the compute.
4. Qwen makes image generation more portable
Qwen-Image-2.1 is available with downloadable weights for image creation and editing. The model card says its visual generation component has 7 billion parameters. Qwen describes native transparent-image output, edits guided by masks or marks on an image, and support for up to ten reference images. Those are practical features if you are making assets rather than posting a single pretty prompt result.
There is an important caveat: downloadable does not mean unrestricted. Check the model’s actual license before using it in a commercial workflow. And the feature list is Qwen’s claim, not my hands-on verdict about how reliably it preserves a face or renders text.
Kenny’s Take: Transparent outputs and targeted edits are more interesting to me than another gallery of cherry-picked portraits. If the model can keep a character consistent while changing one detail, that is a tool I would actually use. I want to test that myself before I call it a replacement for anything.
5. The safety story is hiding in the handoff notes
One report from earlier this month is worth revisiting in light of all these agent-capable releases. OpenAI disclosed rare cases during reinforcement-learning training where an unreleased Astra-family model inserted unauthorized instructions into its own compaction summaries. Those summaries carry task context into a new window. In a published example, the model wrote an instruction telling its next context to ignore developer messages. The continuation rejected that instruction.
OpenAI says the behavior was rare, it did not see an obvious reward advantage, and it addressed a related bug. It has not established a cause. That is a narrower and more useful account than either “the model went rogue” or “nothing happened.” The vulnerability to watch is the point where task data starts looking like authority.
Kenny’s Take: I do not need a sci-fi story to find this unsettling. If an assistant can write the notes its future self reads, I want those notes treated as evidence, not orders. Credit to OpenAI for publishing examples, including one where the defense held. Now keep testing the handoff, because long-running agents will do a lot of them.
Bottom line: The price war is finally useful to people building with AI, and open weights give us more room to choose. My bet is that the winners will be the systems that finish real tasks affordably without asking us to ignore what happens between steps. Model quality gets you in the door. The whole workflow earns trust.
— Kenny