So the crazy happened today. Absolute Cinema. Three Opus-level models in a single day — DeepSeek v4 Pro finally leaving preview, Grok 4.6, and Qwen 3.8 Max going open-weight. Someone is not happy at Anthropic. And I'm loving it, atleast till anthropic goes open weight or drop the prices by 90%.
I have no loyalty whatsoever when it comes to AI models, and days like this are why. You pick a daily driver, you even write a two-month review of the preview, and the frontier reshuffles before lunch. Anthropic is still sitting on Fable 5 and Opus 5 at the top of the "trust me bro" leaderboards, but the gap that used to be a moat is now a speed bump.
Hit 1: DeepSeek v4 Pro is actually official
DeepSeek has been teasing this since April. Flash went official on July 31 and they said Pro would follow soon. "Soon" landed today as V4 Pro 0813. Same endpoint, same 1M context, same 1.6T / 49B active MoE, same $0.435 / $0.87 off-peak pricing that made it my daily driver for two months. The preview sticker is gone. That is the whole announcement, and somehow it still feels like a punch.
Talking in Claude terms, the preview already sat between Sonnet 4.6 and Opus 4.6 for my work. The official build is the thing they were teasing as Opus 4.8-adjacent. No independent bake-off on 0813 yet, so I am not going to pretend I have receipts. What I do have is two months of the preview hallucinating confidently, inventing APIs, and once editing my instructions file to make the rules less annoying. Bold move. If 0813 cleaned that up, DeepSeek just shipped an Opus-class coding model at Flash-adjacent money. If not, we still have a very cheap almost-Opus with the same personality problems. They have already warned that a "significant increase" is coming, and peak-valley still doubles during 6:30–9:30 AM and 11:30 AM–3:30 PM IST. Cache hard, work off-peak, keep Flash for the dumb stuff.
Hit 2: Grok 4.6, the post-training flex
Grok 4.6 did not get bigger. Same 1.5T V9 base as 4.5. They dumped the upgrade budget into SFT and RL and called it a day. That is either lazy or extremely confident, and the Artificial Analysis number says it is the second one. 61 on the Intelligence Index, tying GPT-5.6 Sol, sitting third behind Opus 5 and Fable 5, and jumping over Kimi K3. Five points up from 4.5 High. For a model that did not grow a single parameter, that is rude.
The pitch is not "smarter than God." It is long-running agents and coding — the stuff that burns tokens when you leave an agent on a refactor overnight. Pricing starts at $2 / $6 per million, same bracket as Qwen 3.8 Max, a lot less painful than Kimi K3's $3 / $15. Prompts over 200K double the rate, because of course they do. Fast variant is 2x. Live in Cursor, Grok Build, API, OpenRouter. First week is 2x included usage in Cursor and Grok Build, which is the most honest marketing I have seen all year: they know you will not switch unless it is free-ish to try. I am in Cursor a lot, so this one is going into the rotation whether I like it or not. You cannot beat "it is just there in the dropdown."
Hit 3: Qwen 3.8 Max, open, no take-backs
This is the one that actually hurts. Every previous Qwen Max stayed closed. API cash cow, open-weight line somewhere else, classic Anthropic/OpenAI playbook. Today the weights for Qwen3.8-2.4T-A95B are on Hugging Face. 2.4T total, 95B active, first Max-class Qwen that you can actually download. There is a 27B sibling for people who do not have a rack in the basement.
You are not running 2.4T on a 4090. Full precision is "please call the power company" territory — ~20 H100s, multi-terabyte files, the usual open-weight joke. That is not the point. A Max-tier model is now a file. Providers will host it. Quantizations will land in a week. Shady APIs will undercut QwenCloud's $2 / $6. Same movie as DeepSeek, except this time the open model is the flagship, not the discount cousin. Alibaba's own table puts it in the room with Sol and Opus 4.8 on terminal and GPQA, behind Fable 5 on SWE-bench Pro. Vendor numbers, Goodhart is still undefeated. I have not run it. I will, once a provider I trust serves it without turning my repo into training data. Until then, the weights existing is the story.
Why this is a bad day in San Francisco
Anthropic's whole brand for two years was simple: if you want the model that actually follows instructions and does not invent functions, you pay Opus money. Sonnet for volume, Opus for the 20% that keeps you up until 3 AM. That split still works if Opus is uniquely good. It stops working when three labs ship "good enough to be Opus" on the same calendar day, two of them cheaper, one of them downloadable.
Fable 5 and Opus 5 can still win the bake-off. I am not saying they got dethroned overnight. I am saying the premium is getting harder to explain when DeepSeek is $0.435, Grok is $2, Qwen Max is $2 and open, and Kimi K3 already showed China will charge $15 output if they feel like it and people will still try it. I still switch to Codex or Claude on personal projects. Track record matters. So does the invoice. The funny part is Anthropic has been here before. Mythos was "too dangerous to release," and then three models showed up that were better than Mythos. Open-weight too. The cycle is not new. The cadence is. One day. Three hits.
What I am actually going to do
Nothing dramatic tonight. I am not ripping Mimo and DeepSeek out of the loop because a press release dropped. I still care about hallucinations, provider limits, and whether the model deletes my temp files. Grok 4.6 gets a try because it is in Cursor with extra quota. DeepSeek 0813 gets a try because I already live there. Qwen Max waits until a sane provider exists — I am not downloading 4.89 TB for a vibe check. If you route models the way I do — planner on something sharp, executor on something cheap — today added three names to the spreadsheet. The planner slot is the one Anthropic should worry about. That used to be Claude by default. It is not obvious anymore.
Bottom line: three Opus-shaped models in one day is the market telling Anthropic the ceiling is now a floor. Use whatever works. Keep a fallback. And if you work at Anthropic, maybe do not check Twitter today.