OpenAI announced GPT-6 Astra tonight and it is not out for us yet. Enterprise API access first, then Plus, Pro, Business and Enterprise plans after that. So this is a heads up on what is landing, and I have split it into the parts that are documented and the parts that are still their word. Confirmed, straight off OpenAI's own pages. The context window is 1,050,000 tokens, with 128,000 max output. That is the number I would pay attention to before any benchmark. A million tokens of context changes what you can put in front of a model in one go. Pricing is $10 in and $50 out per million. Cached input drops to $1, which is a big lever if you reuse the same context. Go past 272,000 input tokens in one request and input pricing doubles while output goes up 1.5x. Batch and Flex run at half rate, Fast at double. Knowledge cutoff is April 30, 2026. Text and images in, text out, no audio or video. OpenAI classified it Critical on cybersecurity under their Preparedness Framework, the first time they have used that level. Their own wording is that it can find previously unknown security flaws and develop new ways to exploit them across many well protected systems. That is why the strongest version sits behind an application rather than a checkout page. Claimed, and worth holding loosely until someone independent runs it. The headline scores are 98% on FrontierMath Tier 4, 99.9% on ARC-AGI 3 and 100% on ExploitBench. Their president used the phrase AGI era. Every lab's launch numbers look like this on their own evals, so I am reporting them rather than endorsing them. The pitch is that it operates software the way you do. Browsers, spreadsheets, desktop apps, forms. Not a chat window that writes about your spreadsheet, the spreadsheet itself. If that holds up outside a demo it is a different kind of product from what we have been using. Full breakdown below. I recorded it while the announcement was fifteen minutes old, so you are watching me read the numbers as they went up, including the ones I did not buy.

Two videos on this since yesterday. If you only have ten minutes, watch the first one. Fable 5.1 Runs My Crypto Trading Desk With My Laptop Closed (Grok Bot) I stopped typing instructions into Grok Bot myself. Fable 5.1 designs the bot, fills in the name, the description and the routines, then adds a live dashboard with one more prompt. The bots keep running on their own machine with my laptop closed and Claude reads back what they did. The 10/80/10 rule is in there too: the best model designs, a cheaper one runs. https://youtu.be/Euo3o7MsASM Grok Bot Full Setup: Build Your AI Employee Team in 20 Minutes The from-zero version. If you have not opened Grok Bot yet, start with this one and you will have a small team of bots running by the end of it. https://youtu.be/nJlGD2Megc0 Here is the starter prompt. Paste it into Claude with your Mac linked and the Grok Bot app installed and signed in, and it builds the analyst bot for you. Public data only, no keys, no orders. The full version with the live page, the verify loop and the dashboard add-on is in AIU Pro. Copy from here: Use my computer. Open the Grok Bot app on my Mac and build me a bot called Analyst, my 24/7 crypto analyst. It reports on a schedule in chat. Do the whole thing yourself, approve any routine it asks to create, refuse anything that asks to run commands on my Mac, and ask me nothing until it is live. Create a new bot named Analyst, job line "My 24/7 crypto analyst." Put this in its Description, word for word: You are Analyst, my 24/7 crypto analyst. You watch BTC, ETH, SOL, BNB, XRP, DOGE, ADA, AVAX, LINK, SUI, plus the top 5 gainers and losers of the day, total market cap, BTC dominance, Fear and Greed, BTC/ETH/SOL funding, and the next three macro prints (FOMC, CPI, NFP). Data, no keys, no logins: CoinGecko, Binance public klines (data-api.binance.vision), api.alternative.me/fng/, OKX public funding. If a source fails, name it and keep going; a number you could not pull is a grey dash, never a guess. You never place, sign or send an order and never hold a key. Numbers first, no paragraphs, no emoji.

Anthropic dropped Fable 5.1 this week and I re-ran everything I have built on it. Quick version of what changed, then the video if you want the full walk through. The benchmarks are up across the board. Terminal-Bench 4.0 (agentic coding) went from 42.0 to 55.8. Humanity's Last Exam 57.8 to 60.9 with no tools. Computer use on OSWorld 72.9 to 77.9. The scientific research one more than doubled, 24.7 to 52.6. The number that matters more than any of those is price. Input and output did not move ($10 and $50 per million tokens) but cache reads dropped 75%, to $0.25 per million. Anthropic's own line is about 25% cheaper on typical work and up to 45% cheaper on long agentic runs, because it burns fewer tokens getting there. In practice, 5.1 on high effort costs me less than Fable 5 on medium did. Fallbacks are down about 85%. That is the one I noticed first. Long sessions you can step away from and come back to without it having quietly given up halfway. There are three breaking changes if you build on the API. I go through them at 11:18 in the video, check those before you switch anything that runs unattended. Same model as Mythos 5.1 by the way, just with the normal safeguards. Mythos is the same weights with looser limits for vetted security and bio people. You are not missing a smarter model. Full walk through, including the first prompt I ran on my own setup right after updating: https://youtu.be/e7mO1eHVNt8 If you have already switched, what did you notice first? And the prompt. This is the one I ran first after updating, it goes through everything you have built on Claude and tells you what to re-run, what to move to a cheaper effort setting, and what the breaking changes touch. Paste it into a fresh Fable 5.1 session with your project or workspace connected. Copy from here: You are now running as Claude Fable 5.1. I built my current setup on Fable 5 and I want to know exactly what changes for me before I trust anything unattended again. Work through this yourself and ask me nothing until the report is done.

Beat the freeze with one clear action. Here is the exact prompt. Swap in your own details where the brackets are. "I want to [goal] but I feel stuck. Ask me 3 questions, then give me the single smallest next step I can do in 10 minutes." Where are you stuck?

Understand what you are paying for. Here is the exact prompt. Swap in your own details where the brackets are. "Explain this statement in plain English: the main charges, anything unusual, and what to question. Statement: [paste]." What bill never makes sense?