This one is late, and it is late on purpose. Between June 9 and July 9, the entire frontier turned over — both flagships, a government intervention, a superapp, and the Siri deal finally going from rumor to beta. Ranking any of it in the first week would have been transcribing press releases. I wanted to actually live in these models first. I have now, and here is where everything landed.
Same format as always. Eight models. Four tiers. And for the second month running, a tie at the top — but the reasons underneath it are completely different from last month's.
The S Tier: Claude and ChatGPT — Still Tied, For Entirely New Reasons
Last month the tie was Opus 4.7 versus Codex 5.5, and the story was division of labor. This month both of those models are history.
Anthropic went first. On June 9 they shipped the Claude 5 family and a new tier that sits above Opus — Mythos-class. Claude Fable 5 is the generally available version; Claude Mythos 5 is the same underlying model with some safeguards lifted, offered only to a small set of approved organizations. I wrote about my first day and a half with Fable and made the claim I am still standing behind a month later: it doubled my pace, from roughly one month of production per day to two. It gets the first pass right more often, and — the part that still surprises me — it contributes ideas I didn't put there. Stripe reported it compressed a codebase-wide migration that would have taken a team two months into a single day. That matches the shape of my experience exactly.
Then it disappeared for 19 days. On June 12, the US Commerce Department issued an export-control order after Amazon researchers demonstrated a jailbreak that got Fable 5 to find software vulnerabilities and, in at least one case, produce working exploit code. Because the order covered "any foreign national" and nationality can't be verified in real time, Anthropic suspended the model globally rather than partially. They spent the outage training a new safety classifier that blocks the reported technique in over 99% of cases — requests that trip it get rerouted to Opus 4.8 instead of being served by Fable at all — and the Commerce Department's AI standards center called the result "extraordinarily strong." Controls lifted June 30. Fable came back July 1. I spent those 19 days on Opus 4.8 and, at the very end, the new Sonnet 5 (shipped June 30, near-Opus performance at $2 per million input tokens — quietly the best agent value on the market right now). The lesson I'm keeping: the frontier is now infrastructure, and infrastructure needs a fallback plan. Mine got tested involuntarily.
OpenAI went second, through a door I did not expect. GPT-5.6 arrived June 26 as a limited preview restricted to government-vetted partners — the White House asked OpenAI to control who got access, and OpenAI complied — then went fully public on July 8 once the administration approved a broad release. The family is three tiers with names that are apparently permanent now: Sol is the flagship, Terra is the workhorse, Luna is the fast cheap one. Sol is the one that matters here: long-horizon agentic work, coding, and reasoning at a level that is genuinely, honestly even with Fable 5 — at half the price ($5/$30 versus Fable's $10/$50). And on July 9 the superapp they teased in June actually shipped: ChatGPT and Codex are now one desktop app, with a new agent called ChatGPT Work that will stay with a project for hours and come back with finished output. Five-million-plus weekly Codex users woke up inside a rearranged product, and the rearrangement is good.
So: Fable 5 versus Sol. I have run both, hard, on the same real work, and I am not going to manufacture a winner. I like them equally, and I mean that literally — not as a diplomatic dodge. Fable is still the one I trust to hold a whole system in its head across days. Sol is relentless in exactly the way Codex was, with more brain behind it. On any given task the gap between them is inside my measurement error.
But the workflow changed, and this is the real headline of the month. Last month I divided the labor: Claude architected, Codex designed and bug-checked. This month I stopped dividing and started making them fight. I build with one, then hand the result to the other with explicit instructions to attack it — find the flaw in the auth path, argue the data model is wrong, break the migration, tell me why this will fall over at 10x load. Then I swap them and run it the other direction. Adversarial prompting, model versus model, against my own work. Each one finds holes the other missed — and holes I missed, which is the point. What survives both of them is what ships. The pair-of-specialists era lasted exactly one month; the sparring-partners era is better.
Why still a tie and not a decision? Capability is even, so it comes down to the trust ledger, and the ledger got muddier on both sides. OpenAI still runs ads for free users; Anthropic still doesn't. Anthropic proved it will take its flagship dark for 19 days rather than ship around a safety problem — which reads as integrity from one angle and as infrastructure risk from the other, and I felt both. And the biggest trust question of the month applies to both companies equally, so it can't break the tie: the US government now sits in the release pipeline. More on that below.
The A Tier: Gemini and Llama
Gemini finally got the win this series has spent five months doubting. At WWDC on June 8, Apple unveiled the rebuilt Siri, and its cloud intelligence runs on a custom Gemini model — reportedly around 1.2 trillion parameters, under a multi-year deal worth roughly a billion dollars a year, confirmed in an actual joint statement from both companies. It's in the iOS 27 developer beta right now, public beta this month, shipping September. After watching every previous window slip, I'll hold my applause until it's on phones — but this is no longer vaporware, it's in beta, and it is the biggest distribution win in consumer AI. Between that, Gemini 3.5 Flash still being the best speed-to-intelligence ratio on the market, and 3.5 Pro landing out of testing, Gemini's A Tier has never been more solid. The gap to the S Tier is unchanged though: on the long, stateful, build-the-whole-system work, it still drifts where Fable and Sol finish — and it's not in my sparring rotation yet.
Llama drops in my estimation this month, even though it holds the rank. Behemoth is now effectively shelved — never released, never formally cancelled, just quietly not talked about anymore. Meanwhile the tell is Muse Spark, the closed-weight, API-only frontier model Meta shipped this spring: Meta's frontier effort has gone proprietary, and Llama is becoming the legacy open line rather than the main event. Llama 4 still has the widest deployment of any open model family on the planet, which is worth an A Tier slot by itself. But for the first time since I started this series, the open-source crown is genuinely contested — and the contender is one tier down this page.
The B Tier: Mistral and Perplexity
Mistral is having its most interesting month ever. A new open-weight frontier model — a "fat but sparse" mixture-of-experts — entered early access this month with research, government, and industry partners, broader release later this summer. If it lands anywhere near the frontier, it will be the first credible open-weight bid for the top since Llama stalled, and it would make Mistral the de facto leader of open-source AI. They also shipped Robostral Navigate — a robotics navigation model that runs on any robot fleet from a single camera and plain-language prompts — and the valuation is closing in on $23 billion. Watch this one. B Tier today, but the trajectory is the steepest on the page.
Perplexity keeps compounding. The new thing is Personal Computer — an always-on agent living on a dedicated Mac mini that monitors triggers and carries work forward around the clock, including local files and local browsing. Comet went enterprise with full MDM deployment, Voice Mode landed, and Max subscribers can now pick the model driving the browser agent. Still no ads, another quarter running. Still my front door for research, and the narrow-playbook-executed-well thesis keeps paying out.
The Nope Tier: Grok and DeepSeek
The Grok situation got worse, not better. On July 7 the plaintiffs filed an amended complaint adding two new victims — both minors when the images were made — and a new defendant, Stability AI. The detail I can't get past: the National Center for Missing & Exploited Children found that 90% of xAI's CyberTipline reports were unusable by law enforcement because xAI declined to include user information. Not a capability problem. A choice. The model keeps improving on the technical axis and it keeps not mattering. Nope.
DeepSeek gets the closest thing to good news I've written about it all year: V4 is actually going official this month, after months of preview limbo — a 1-million-token context window across the lineup, MIT license, and (a first for any major lab) peak and off-peak API pricing, with peak hours charged at double the rate. Credit where due: it shipped, and the engineering-per-dollar story is real again. The trust math is exactly where it was: CCP censorship baked in at the model layer, the distillation accusations from OpenAI and Anthropic still standing. If you handle sensitive data, the pricing innovation is not the variable that matters. Nope, with an asterisk of technical respect.
What Changed This Month
Three things, in order of importance.
First: the government is now in the release pipeline, and I think it's permanent. Fable 5 spent 19 days dark under an export-control order. GPT-5.6 launched through White House vetting and needed administration approval to go public. Whatever your politics, the mechanics changed this month: frontier releases now have a regulatory stage gate, the same way drugs and aircraft do. Every future entry in this series gets written in that world.
Second: my workflow went from "use both, in order" to "use both, against each other." The division-of-labor combo I described in June assumed the models had different strengths. Fable 5 and Sol are too evenly matched for that to hold, so the leverage moved: now they battle-test what I've developed, each one attacking the other's output and mine. It is the single best quality-assurance process I have ever had, human teams included, and it costs me a prompt.
Third: the open-source torch is visibly passing from Meta to Mistral. Behemoth shelved and Muse Spark closed on one side; an open-weight frontier model in early access on the other. Next month's entry will probably have more to say.
And the disclosure, because this month it is funnier than usual: Fable 5 is the model writing this update, ranking itself into a tie. I had Sol review the draft and told it to argue for a different verdict. It found two factual nits and let the tie stand. Make of that whatever you like — I've already told you what I make of it.
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.
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 adversarial workflow above starts with exactly one sentence: "Attack this."
The Protocol: When two models are genuinely even, stop trying to rank them and start making them fight. Build with one, hand the output to the other with orders to break it, then swap. What survives both is what ships. 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.
