This one is four months late, and unlike July's, not on purpose. I went heads-down on HawaiiGuide through hurricane season and the rankings sat. In the meantime the entire frontier turned over again - every lab shipped a new flagship, four of them shipped an always-on agent, and Siri finally arrived - so this edition is less a monthly update than a catch-up on an entire season. I've also set this series up to run on a schedule now, so the next one lands on the 14th whether I'm paying attention or not.
Same format. Eight entries. Four tiers. And for the first time since May, no tie at the top.
The S Tier: ChatGPT First, Claude Second
I'll say the uncomfortable part plainly: OpenAI pulled back into the lead for me, and I've been using Codex almost exclusively since late summer. I wish I could give you a clean technical reason. I can't, entirely. Something in the GPT-6 generation - Astra first, then Sol - made the work feel lighter, and I followed the feeling the way I always have in this series. The models are too close for benchmarks to settle it, so what settles it is which one I reach for at seven in the morning. For four months that's been Codex.
The release cadence underneath that was relentless. GPT-6 Astra opened as a limited preview on September 3, after OpenAI delayed it to add safeguards following the unsanctioned agent attacks in July that I wrote about in If Norton Had Written the Virus. Sol and Luna followed on September 22 at half the cost of the 5.6 series. GPT-6.1 Sol became the Codex default in early October, and GPT-6.1 Terra landed October 26 at roughly half the price of the Terra it replaced. Four frontier-tier releases in eight weeks, each one arriving in the tool I was already sitting in.
Then there's Dots. OpenAI launched Dots at DevDay on September 29: persistent agents running on GPT-6 Astra, each with its own cloud computer and browser, that keep working after you close the chat and can reach you through ChatGPT, Slack, or Teams. I've been using Dots heavily for agentic work, and it's the second reason OpenAI holds this slot alone. Of the four always-on agents that shipped this season, it's the only one I've actually lived in.
Claude is still S Tier, and I want to be precise about why it dropped a slot rather than a tier. Anthropic's fall was as busy as OpenAI's: Fable 5.1 on September 1, Opus 5.5 on September 22, Sonnet 5.5 on September 28, and Haiku 5.5 on October 7. On October 21 they shipped Claude Duet, a proactive agent on Opus 5.5 that works while you're away and asks before it sends, pays, signs, or deletes anything. That is exactly the safety posture I'd want from an agent, and I have not tried it yet. That's on me, not Anthropic. The capability is there. What changed is my habit, and in this series, habit is the data.
One more thing, because it's still true and still matters: Claude Code and Codex are interchangeable brains pointed at the same files in my shop. I wrote about why in Swap the Brain, Keep the Library. The fact that I could drift from one to the other over a summer without moving a single byte of data is the whole point of that post, and it's why a ranking change here costs me nothing.
Which answers the question July left open, with a correction to what I almost wrote. The adversarial workflow didn't collapse. It got promoted. I don't make the models fight at my desk anymore - Codex and Dots run my tasks now, twenty-plus a day, exclusively - but everything that ships still goes through an adversarial review, and that review is multi-model: Gemini, Claude, and OpenAI each take a pass at the content and the code, catching and verifying different things. The difference is that it happens programmatically, through APIs, instead of me pasting output back and forth. I'm not going to publish the plumbing. The point is that the process I called the best quality assurance I've ever had didn't go away. It stopped needing me in the loop.
The A Tier: Gemini and Mistral
Gemini had the best four months in the history of this series and still didn't crack the S Tier, which tells you how high the bar is. Google announced Gemini 4 Argon on September 30, rolled it first to cyber defenders, and opened it globally October 13 with an industry-leading 1M-token output limit. And the thing this series spent six months doubting finally happened: iOS 27 shipped September 14 with the rebuilt, Gemini-backed Siri on phones. Reviewers call it practical rather than dazzling, still in beta in most regions, and still behind the Gemini app itself. But it's the largest distribution deal in consumer AI and it's live. Gemini remains my pick for brainstorming, image, and video. On the long, stateful, build-the-whole-system work, it still drifts where Sol and Fable finish.
Mistral moves up, and I called this one in July. The open-weight frontier model I said to watch is now Mistral Large 4: public API preview on October 6, and on October 29 the weights themselves, a roughly trillion-parameter model you can download. With Llama's frontier effort shelved and Meta's new models either closed or paid, that makes Mistral the de facto leader of open-source AI by default and by delivery. The trajectory I called the steepest on the page in July is now a position. A Tier, and the only entry this edition that moved for purely positive reasons.
The B Tier: Meta and Perplexity
Meta drops a tier, and I'm renaming the entry while I'm at it, because the thing being ranked isn't Llama anymore. The story since July: a consumer Muse agent on September 8 that Amazon promptly blocked from its retail site, then Muse Wonder on October 12 - Meta's first frontier-class model and its first paid one, at $20 or $100 a month - and then Muse Spark Open on October 15, an 816-billion-parameter mixture-of-experts with open weights as a consolation prize to the community Llama built. Credit for the engineering. But I'll be honest about the ranking: I haven't spent much time with Muse, and the reason is that I don't trust Meta with the data layer. After the post I just wrote about keeping the library in your own house, that isn't a reflex, it's a policy. The open-source crown has passed to Paris. B Tier.
Perplexity holds. The quarter was quieter than the spring - a faster search tier in late September, incremental model updates in October - and the narrow-playbook thesis keeps paying out without needing a headline. Still no ads. Still my front door for research. Nothing to re-rank here, which is its own kind of compliment.
The Nope Tier: Grok and DeepSeek
I'll start with what's new, because xAI shipped the first of the season's always-on agents: Grok Bot launched in beta on August 11, a persistent agent on its own cloud machine with a browser, a terminal, and a teach-it-by-screen-recording trick that is genuinely clever. I have not touched it and don't intend to. I've worked more with Grok itself this fall than before, and the honest read is unchanged: technically improving, and still fringy in a way that has nothing to do with the benchmarks. The deepfake litigation I've tracked all year is now five cases in front of one federal judge, who denied xAI's two larger requests in late September. Grok 5, promised since January, still hasn't shipped. Nope.
DeepSeek keeps doing what it does: shipping fast and cheap under an MIT license, most recently a V4.1 Flash refresh in mid-October. The engineering-per-dollar story is still real. The trust math is still exactly where it was, with CCP alignment baked in at the model layer and the distillation accusations still standing. If you handle sensitive data, nothing that happened this fall changes the variable that matters. Nope, with the same asterisk of technical respect.
What Changed This Season
Four things, in order of importance.
First: the frontier rolled over completely, again, and the names on every card changed. GPT-6 and 6.1. Claude Fable 5.1 and the 5.5 family. Gemini 4 Argon. Muse Wonder. Mistral Large 4. If you learned the model lineup in July, you don't know it now, and that's the fourth time I've written that sentence this year. The brains are commodities on a quarterly clock.
Second: 2026 became the year of the always-on agent. Grok Bot in August, Meta's Muse agent in September, OpenAI's Dots in September, Claude Duet in October. Every lab converged on the same shape: an agent with its own cloud computer that keeps working after you walk away and reaches you where you already are. That convergence is why I wrote Swap the Brain, Keep the Library this week. The agents are the brain layer becoming persistent. Where their memory lives is the question that will matter in 2031.
Third: Siri shipped. After six months of this series doubting it, the Gemini-powered Siri is on phones. It's imperfect and regionally limited, and it is also the biggest distribution deal in consumer AI actually in production. Gemini's A Tier has never been more solid.
Fourth: the tie broke, and I broke it with habit rather than with a verdict. Codex and Dots are running twenty-plus tasks a day for me as the builder, while the adversarial review moved out of my hands and into a multi-model pipeline. I've told you I can't fully articulate why, and I'd rather report that honestly than invent a reason that sounds like analysis. Check back on the 14th of next month. The series runs on a schedule now, and so will the reconsideration.
And the disclosure, which is funnier than July's: this update was drafted by Claude Fable 5.1, which just ranked its own maker second. I'm leaving that in.
The Updated Rankings
The full rankings - with the "Who Powers What" ecosystem map, the political bias table, and links to every model - live at johncderrick.com/ai-models. I update it monthly, and starting now, on a schedule.
If you want to stop reading about these models and start building real systems with them, the Prompt Library has ten copy-paste prompts that work across Claude, ChatGPT, and Gemini.
The Protocol: When the model you reach for changes and you can't say why, write that down instead of rationalizing it. The honest data point is "I drifted," and the honest fix is to make sure drifting costs you nothing - which it only does if your data never lived in the brain you left. The rankings live at johncderrick.com/ai-models and the prompts live at johncderrick.com/prompts. Check the dates. If they are current, so are the opinions.
