I opened my news feed this morning and had the peculiar feeling that I had missed a month of AI releases overnight. Grok 4.7 arrived on Monday. Tuesday brought Claude Opus 5.5, GPT-6 Sol and Luna, and Xiaomi's MiMo-V2.6 Pro and Flash. I had barely finished reading one launch post before the next tab was asking for my attention.
I love this pace, but I also find it exhausting. A model can be “new” for less than a day before a cheaper rival appears, and every company has a chart explaining why its own model sits on the frontier. The useful question is no longer simply which name is newest. It is what each release changes for someone actually paying for an API, running an agent, or picking a model for a product.
Here is my read of this unusually crowded week, as of September 23, 2026. The releases are real; the verdict on which model is best for a particular workload will take longer.
First, what actually launched?
The dates are easy to blur together. SpaceXAI announced Grok 4.7 on September 21. On September 22, Anthropic released Claude Opus 5.5, OpenAI released GPT-6 Sol and Luna, and Xiaomi released MiMo-V2.6 Pro and Flash.
One correction matters: GPT-6 Terra was not part of OpenAI's September 22 announcement. Sol and Luna have official release notes and API model IDs; I could not find an official GPT-6 Terra release or price as of September 23. Terra belonged to the earlier GPT-5.6 family, so it should not be treated as a sixth confirmed launch this week.
At a glance: the new models are aiming at different budgets
These are published standard API list prices in US dollars per million tokens, checked September 23. They do not include every cache rule, long-context tier, regional rate, tool charge, or the number of reasoning tokens a task actually consumes.
| Model | Published input / output | What caught my eye |
|---|---|---|
| Claude Opus 5.5 | $4 / $20 | Long-running coding and knowledge work; lower typical task cost than Opus 5, according to Anthropic. |
| GPT-6 Sol | $2 / $10 | OpenAI's stronger new everyday work model, with a lower API price than GPT-5.6 Sol. |
| GPT-6 Luna | $0.10 / $0.50 | Cheap enough to make high-volume model routing a serious design choice. |
| MiMo-V2.6 Pro | $0.435 / $0.87 | Open-weight flagship with very aggressive overseas API pricing. |
| MiMo-V2.6 Flash | $0.14 / $0.28 | Lower-cost member of Xiaomi's open-weight multimodal pair. |
| Grok 4.7 | $2 / $6 | 500K context and stronger coding-agent results, with a token-use caveat. |
Sources: Anthropic pricing, OpenAI API changelog, Xiaomi MiMo catalog, and SpaceXAI release notes. OpenAI's listed Sol and Luna rates apply to prompts up to 272K input tokens; Grok's standard rates apply below its 200K long-context threshold. Check the provider's current billing page before budgeting a deployment.
Claude Opus 5.5: efficiency is the surprising headline
When I see “Opus,” I expect a premium model and a premium bill. The interesting part of Opus 5.5 is that Anthropic is selling a more capable long-running agent model while saying it costs about 40% less per typical task than Opus 5. The token rates are $4 input and $20 output per million, down 20%; cache reads are $0.20 per million. Anthropic says Opus 5.5 also uses fewer tokens per task and produces output over 30% faster than Opus 5.
The company reports strong results on terminal coding, codebase migrations, computer use, and professional work. Those numbers are useful, but most of the head-to-head launch charts are company-run evaluations with different effort settings and agent harnesses. I would try Opus 5.5 on a difficult, messy repository or document workflow where finishing correctly matters more than the lowest rate on a pricing page.
Its 1 million-token context window is also notable, although a big context window is permission to send more material, not proof that every detail will be used well. Anthropic's platform notes confirm the model ID claude-opus-5-5 and the default 1M window.
GPT-6 Sol and Luna: one release, two very different jobs
OpenAI's launch is easier to understand if Sol and Luna are treated as two budget decisions. Sol is the model I would test for substantial coding and work tasks when I want strong capability without reaching for the most expensive tier. At $2 input and $10 output per million tokens, its standard API price is half the earlier GPT-5.6 Sol promotional rate.
Luna is the one that made me stop scrolling. At $0.10 input and $0.50 output per million, it invites a different question: which parts of an application genuinely need a flagship model? Classification, extraction, simple tool routing, and many first-pass drafts can be tested on a cheaper model before escalating the hard cases. That is a workflow choice, not a claim that Luna beats the larger models at everything.
OpenAI reports improvements over the GPT-5.6 versions in coding, factuality, computer use, and caching. Sol and Luna are listed for the API as gpt-6-sol and gpt-6-luna; the official changelog confirms both. Their appearance in ChatGPT Work and Codex is being rolled out, so product availability may lag the API for some users.
Xiaomi MiMo-V2.6: the open-weight release I would not ignore
Xiaomi's release could get lost behind the louder Claude and GPT headlines, but I think it deserves close attention. MiMo-V2.6 Pro and Flash are open-weight, multimodal models, and Xiaomi has published models and training materials. That gives developers another path besides relying entirely on a hosted frontier API.
Overseas API list rates on Xiaomi's site are $0.435 input and $0.87 output per million for Pro, and $0.14 input and $0.28 output for Flash. Xiaomi says MiMo-V2.6 Pro scores 46 on the Artificial Analysis Intelligence Index and that its large-scale reinforcement-learning run was accompanied by a live training dashboard. Those are interesting signals; a single index score does not settle whether it will be good at your own agent, language, or content task.
The appeal here is choice. A team can test Xiaomi's hosted API, inspect the released weights, or explore self-hosting if it has the hardware and expertise. For a site like ours that tracks many kinds of AI tools, that matters: the model underneath an app increasingly affects its cost, speed, privacy options, and how much control its builder has.
Grok 4.7: a strong launch with a visible cost trade-off
We covered Grok 4.7 in detail yesterday. The short version: SpaceXAI kept its standard $2 input and $6 output rates, gave it a 500K context window, and reported substantial coding and agent gains over Grok 4.6.
The independent result is more nuanced. Artificial Analysis measured better agentic performance, but Grok 4.7 at xhigh effort used about 81,000 output tokens per Intelligence Index task versus 38,000 for Grok 4.6 at xhigh. If an agent keeps working for minutes, the output meter keeps running too. This is why I want cost per successful task beside price per million tokens whenever someone declares a model a bargain.
So which one would I choose?
I would not crown a universal winner after roughly 48 hours. My first tests would be practical:
- Complex, high-stakes coding or analysis: compare Opus 5.5, GPT-6 Sol, and Grok 4.7 on the same real task and measure corrections needed.
- High-volume, routine work: start with GPT-6 Luna or MiMo-V2.6 Flash, then route difficult cases upward.
- Open-weight control: examine MiMo-V2.6 Pro and Flash, while accounting for hosting and operational costs.
- Long agent runs: record total input, cached input, output, wall time, retries, and completion quality. The cheapest token can still belong to an expensive task.
Benchmarks help narrow the field, but their harnesses, effort settings, and scoring rubrics differ. A model that wins a coding test in its first-party agent may not lead in a bare API call or a character chat app. The same caution applies to the tools we list on NSFWAITool: compare the experience the site delivers, not just the model name in its marketing. Our SillyTavern guide shows how much prompts, presets, and providers can change the result even when the base model stays the same.
What this week says about the next year
The part that sticks with me is how quickly “frontier” and “affordable” are starting to overlap. Anthropic is reducing the cost of its premium tier. OpenAI is making its smaller model much cheaper. Xiaomi is pushing open weights and low API rates. SpaceXAI is competing on agent performance while holding its token price steady. These are different strategies, but all four companies are asking us to think beyond a single chatbot answer.
My guess is that the next advantage will come from the whole system around a model: caching, tools, memory, evaluation, safe recovery, and sensible routing between expensive and cheap models. The model still matters enormously. So does the amount of work it can finish before a human has to clean up after it.
I am excited by that, even if the pace makes my browser feel permanently out of date. In a few weeks, I expect some of these launch-week rankings to look different. The durable lesson is simpler: keep a small set of real tasks, rerun them when a model changes, and judge the final work alongside the bill.
Frequently asked questions
Did Claude Opus 5.5 launch on September 23?
No. Anthropic dated its Claude Opus 5.5 announcement September 22, 2026.
Did OpenAI release GPT-6 Terra with Sol and Luna?
No official GPT-6 Terra launch was listed in OpenAI's September 22 announcement or API changelog as of September 23. The confirmed releases were GPT-6 Sol and GPT-6 Luna.
Are MiMo-V2.6 Pro and Flash open-weight?
Yes. Xiaomi describes the V2.6 series as released and open-sourced and links to the model collection. Running a large model yourself still requires suitable hardware and setup.
Which new model is cheapest by published API token price?
Among these listed standard overseas rates, MiMo-V2.6 Flash has the lowest output-token price at $0.28 per million; GPT-6 Luna has the lowest input-token price at $0.10 per million. These figures do not predict the total cost of completing a task.
Can the launch benchmarks be compared directly?
Only with care. Scores can depend on the agent harness, tools, reasoning effort, data set, and who ran the test. Compare models using the same workload and count the cost of unsuccessful attempts.
Which model should I try first for coding agents?
Use a representative issue from your own repository. Test GPT-6 Sol, Claude Opus 5.5, and Grok 4.7 under comparable conditions; add MiMo-V2.6 Pro if open weights or its API economics matter to you. Record correctness, elapsed time, tokens, and rework.