𝗙𝗢𝗥 𝟮𝟳 𝗬𝗘𝗔𝗥𝗦 𝗡𝗢𝗕𝗢𝗗𝗬 𝗞𝗡𝗘𝗪 𝗪𝗛𝗘𝗥𝗘 𝗧𝗛𝗘 𝗡𝗦𝗔 𝗚𝗢𝗧 𝗧𝗛𝗘 𝗡𝗨𝗠𝗕𝗘𝗥𝗦 𝗕𝗘𝗛𝗜𝗡𝗗 𝗬𝗢𝗨𝗥 𝗣𝗔𝗦𝗦𝗞𝗘𝗬𝗦 𝗖𝗟𝗔𝗨𝗗𝗘 𝗙𝗜𝗚𝗨𝗥𝗘𝗗 𝗜𝗧 𝗢𝗨𝗧 𝗜𝗡 𝗨𝗡𝗗𝗘𝗥 𝟯𝟬 𝗠𝗜𝗡𝗨𝗧𝗘𝗦 the curve P-256 protects most passkeys and a big chunk of HTTPS it was built from a “seed” the NSA handed over in the late 90s, with no explanation the rumor was it came from hashing a joke passphrase the man who made it, Jerry Solinas, died in 2023 without telling anyone cryptographers even put up a $36,864 bounty to find the answer then an Anthropic research model got the question, with no hints under half an hour later: the seeds weren’t hashes at all they came from lrand48, an old C random number generator seeded with the computer’s clock, between May 1997 and April 1998 same story for P-192, P-224, P-384 and P-521 [the part people will miss: this is not a crack. your passkeys are fine. it actually makes a hidden NSA backdoor much less likely, because now anyone can reproduce the seeds] what’s still open: why some valid seeds were skipped the B-163 seed is still unexplained most base points are still a mystery the bounty goes to ISRG, the nonprofit behind Let’s Encrypt 27 years of conspiracy theories one random number generator

𝗔𝗡𝗧𝗛𝗥𝗢𝗣𝗜𝗖’𝗦 𝗨𝗡𝗥𝗘𝗟𝗘𝗔𝗦𝗘𝗗 𝗖𝗟𝗔𝗨𝗗𝗘 𝗝𝗨𝗦𝗧 𝗕𝗘𝗔𝗧 𝗛𝗨𝗠𝗔𝗡 𝗤𝗨𝗔𝗡𝗧𝗨𝗠 𝗖𝗜𝗥𝗖𝗨𝗜𝗧𝗦 𝗕𝗬 𝟭𝟳𝗫 Claude Mythos 5 was given 8 building blocks of quantum algorithms and told to make them cheaper how it worked: Claude writes a circuit as Python code a separate program (not an AI) checks if it’s correct and scores the cost Claude gets the score and tries again no human between rounds the results: modular inversion: 17x fewer gates than the smallest previous design, with 12% more qubits a decoder for Google’s DQI algorithm: 12x fewer gates one run went for 53 hours straight the math was formally proved in Lean, not just tested [the part people will miss: modular inversion is a key step in quantum attacks on elliptic-curve crypto, the kind that protects Bitcoin wallets and HTTPS. cheaper circuits there means less hardware needed to break it] before anyone panics: nothing ran on a real quantum computer, these are resource estimates it’s a preprint, no one has reproduced it yet today’s machines are still far too small and noisy and you can’t use this model, Mythos 5 is only for vetted partners same week Google showed cleaner magic states on real hardware the hardware is getting better and now AI is shrinking the math

𝗡𝗩𝗜𝗗𝗜𝗔 𝗝𝗨𝗦𝗧 𝗙𝗜𝗧 𝗔 𝟳𝟰𝗕 𝗠𝗢𝗗𝗘𝗟 𝗜𝗡𝗧𝗢 𝗢𝗡𝗘 𝟮𝟰𝗚𝗕 𝗚𝗔𝗠𝗜𝗡𝗚 𝗖𝗔𝗥𝗗 Nemotron 3.5 Bolt: 74B parameters, 9B active the math on the weights: 16-bit: ~148GB 8-bit: ~74GB 4-bit: ~37GB 2-bit: ~18.5GB NVIDIA built it for 2-bit natively, so it fits on a single 24GB card like an RTX 4090 runs in Ollama and LM Studio, weights on Hugging Face for context, Meta’s Muse Spark Open this week needs two 512GB Mac Studios [the part people will miss: fitting is not the same as competing. on the Artificial Analysis index Bolt scores 21.4, Opus 5.5 sits at 58] where it actually makes sense: local agents that browse and search, it’s #2 among open models on BrowseComp private data you can’t send to an API the cheap repeat steps, while a frontier model handles the hard ones and a 4GB version (Bolt E4B) drops Oct 23 for voice and devices the gap isn’t closing at the top it’s closing on your desk

𝗜𝗡 𝗠𝗔𝗬, 𝗠𝗘𝗧𝗔 𝗦𝗔𝗜𝗗 𝗠𝗨𝗦𝗘 𝗦𝗣𝗔𝗥𝗞 𝗪𝗔𝗦 𝗡𝗢𝗧 𝗦𝗨𝗜𝗧𝗔𝗕𝗟𝗘 𝗙𝗢𝗥 𝗢𝗣𝗘𝗡 𝗦𝗢𝗨𝗥𝗖𝗘 𝗬𝗘𝗦𝗧𝗘𝗥𝗗𝗔𝗬 𝗧𝗛𝗘𝗬 𝗢𝗣𝗘𝗡𝗘𝗗 𝗜𝗧 the timeline: Apr 8: Muse Spark ships, closed May 13: Meta’s AI chief calls it not suitable for open sourcing Aug 10: Zuckerberg promises the weights in the coming weeks Oct 15: Muse Spark Open lands on Hugging Face 816B parameters, 42B active, 1M context where it beats the best open model: MCP-Atlas (tool use): 89.2 vs 84.5 HLE-Diamond: 26.1 vs 22.2 JobBench: 61.8 vs 61.4 where it loses: Terminal-Bench: 76.4 vs 85.0 for Kimi K3 DeepSWE: 60 vs 68 for Kimi K3 so Meta says it’s built for coding, but Kimi K3 still codes better where it really shines is using tools [the part people will miss: “runs on a Mac Studio” means roughly 400GB of weights at 4-bit. that’s the top config, not the one on your desk] before you download: these scores are from Meta’s own table check the license, it’s not the same as Glimmer’s Apache 2.0 for local use, the 30B Glimmer is still the realistic pick open weights are back at Meta open to everyone’s desk is another story

𝗦𝗣𝗔𝗖𝗘𝗫’𝗦 𝗚𝗥𝗢𝗞 𝗩𝗢𝗜𝗖𝗘 𝟯.𝟬 𝗜𝗦 #𝟭 𝗢𝗡 𝗧𝗛𝗘 𝗧𝗘𝗦𝗧𝗦 𝗣𝗘𝗢𝗣𝗟𝗘 𝗦𝗧𝗜𝗟𝗟 𝗣𝗜𝗖𝗞 𝗚𝗢𝗢𝗚𝗟𝗘’𝗦 𝗩𝗢𝗜𝗖𝗘 𝗔𝗜 Artificial Analysis Speech-to-Speech Index: Grok Voice 3.0: 86.9 GPT-Live-1 Astra: 83.2 Gemini 3.8 Live: 82.6 clear win, highest score on the board now look at the Speech Agent Arena, where real people pick the voice they prefer: Gemini 3.1 Flash Live: 1100 Gemini 3.8 Live: 1088 Grok Voice 3.0: 1070 same model, third place [the part people will miss: xAI’s own launch table shows Grok on top for human preference, because the two Gemini models above it aren’t in the table] what Grok actually does best: finishes the task, 94.8% success rate, best on the board cheap at $0.08 a minute fast, about 0.81s to first audio, but that’s a bit slower than Grok Voice 2.0 if you’re building support calls, Grok is a strong pick if your users will just talk to it, test Gemini too tests measure the task people measure how it feels

𝗚𝗢𝗢𝗚𝗟𝗘’𝗦 𝗡𝗘𝗪 𝗚𝗘𝗠𝗜𝗡𝗜 𝗕𝗘𝗔𝗧𝗦 𝗖𝗟𝗔𝗨𝗗𝗘 𝟲 𝗧𝗢 𝟰 𝗥𝗘𝗔𝗗 𝗧𝗛𝗘 𝗙𝗜𝗡𝗘 𝗣𝗥𝗜𝗡𝗧 𝗙𝗜𝗥𝗦𝗧 Gemini 4 Argon vs Claude Opus 5.5, Google’s own table: Argon wins: Vals Index 68.9 vs 67.0 AutomationBench 51.3 vs 42.5 DeepSWE 77.9 vs 74.2 GraphWalks at 1M tokens 84.2 vs 66.8 long video, lab tasks Opus wins: FrontierSWE 62.3 vs 55.0 Terminal-Bench 66.4 vs 57.4 PostTrainBench and Terminal science tasks so Argon is great at huge context and long workflows Opus still leads when an agent lives in the terminal [the part people will miss: $2/$10 is an intro price. Google says it goes to $4/$20 later and hasn’t said when] before you switch: almost every number above comes from Google, no one has reproduced them yet early testers say frontend and design are its weak spot run it on 5 of your real tasks, not on a leaderboard 6 to 4 is a headline your tasks are the real benchmark

𝗔𝗡𝗧𝗛𝗥𝗢𝗣𝗜𝗖’𝗦 𝗔𝗜 𝗪𝗥𝗜𝗧𝗘𝗦 𝗦𝗢 𝗠𝗨𝗖𝗛 𝗖𝗢𝗗𝗘 𝗧𝗛𝗔𝗧 𝗚𝗜𝗧 𝗖𝗢𝗨𝗟𝗗𝗡’𝗧 𝗞𝗘𝗘𝗣 𝗨𝗣 May: ~2,500 merged changes a day September: 30,000+ that’s 12x in 4 months, mostly from Claude and CI, not people git was built for humans who pull the whole repo and push once in a while agents do the opposite, all day so they built their own source control, Scion: 7 engineers, 5 months, with Claude writing a lot of it about 200x less CPU and memory than their old git mirrors checkouts went from minutes to seconds 25% more CPU freed up for CI [the part people will miss: Scion isn’t open source and Anthropic says it’s not a drop-in git replacement. what they shared is the design] 2 ideas you can steal anyway: do commits and merges on the server, not on every machine fetch files on demand instead of cloning the whole repo fun detail: the weird “Antique change requests” on their chart is their new code review tool the model wasn’t the bottleneck the pipes were

𝗠𝗘𝗧𝗔’𝗦 𝗡𝗘𝗪 𝗠𝗢𝗗𝗘𝗟 𝗜𝗦 𝗔𝗟𝗠𝗢𝗦𝗧 𝗔𝗦 𝗚𝗢𝗢𝗗 𝗔𝗦 𝗖𝗟𝗔𝗨𝗗𝗘 𝗜𝗧𝗦 𝗖𝗢𝗗𝗜𝗡𝗚 𝗔𝗣𝗣 𝗧𝗥𝗔𝗜𝗡𝗦 𝗢𝗡 𝗬𝗢𝗨𝗥 𝗖𝗢𝗗𝗘 𝗕𝗬 𝗗𝗘𝗙𝗔𝗨𝗟𝗧 Meta just dropped Muse Wonder and Muse Wonder Full the numbers are real: Artificial Analysis index: Opus 5.5 at 58, Wonder Full at 56, base Wonder at 52 API price: $4 in / $20 out, same as Opus cost to run the full AA test: $4.52 for Wonder vs $5.98 for Opus long coding runs (SWE-Marathon): Full 57.5 vs Opus 46.3 the catch: only base Wonder is on the API Full lives inside Muse Code, $50 or $200 a month, no pay-as-you-go and it’s not open weights, only the small Glimmer model is [the part people will miss: Muse Code ships with “contributor mode” ON, your submissions help train Meta’s models, and turning it off only covers what you send after] 3 rules before you try it: flip contributor mode off before your first prompt no client or company code in it, Meta’s own terms say nothing confidential in contributor mode benchmark it on your repo, not on a chart cheap and close to the top is great just read the settings first

𝗠𝗜𝗖𝗥𝗢𝗦𝗢𝗙𝗧 𝗝𝗨𝗦𝗧 𝗦𝗛𝗜𝗣𝗣𝗘𝗗 𝗔𝗡 𝗔𝗜 𝗠𝗢𝗗𝗘𝗟 𝗕𝗨𝗜𝗟𝗧 𝗢𝗡 𝗔𝗟𝗜𝗕𝗔𝗕𝗔'𝗦 𝗤𝗪𝗘𝗡. 𝗜𝗧 𝗖𝗔𝗡𝗡𝗢𝗧 𝗪𝗥𝗜𝗧𝗘 𝗔 𝗦𝗜𝗡𝗚𝗟𝗘 𝗦𝗘𝗡𝗧𝗘𝗡𝗖𝗘, 𝗔𝗡𝗗 𝗧𝗛𝗔𝗧 𝗜𝗦 𝗧𝗛𝗘 𝗪𝗛𝗢𝗟𝗘 𝗣𝗢𝗜𝗡𝗧 it's called Microsoft-Decision-1 you give it a task and a fixed list of answers. it gives back a probability for each one. no text, no reasoning, one pass refund 0.97 bug report 0.02 sales 0.01 the numbers, from Microsoft: - $0.042 per million input tokens, output is free - about 35x faster than GPT-6 Sol - 36 benchmarks, about 150,000 questions it never trained on - post-trained from Alibaba's open Qwen3.5-9B [the part people will miss: you are not supposed to trust the score blindly. Microsoft gives example thresholds for what to do with it] - above 0.95: do it automatically - 0.60 to 0.95: send it to a bigger model - below 0.60: ask a human where it fits: 1. routing. decide which model gets the request before you pay for the expensive one 2. checking agents. grade an action as ok or not ok before it runs 3. sorting. tickets, reviews, survey answers by theme where it does not fit: writing, planning, anything that needs an explanation the cheapest part of an agent might soon be the part that decides

𝗔𝗡𝗧𝗛𝗥𝗢𝗣𝗜𝗖 𝗟𝗢𝗢𝗞𝗘𝗗 𝗜𝗡𝗦𝗜𝗗𝗘 𝗔 $𝟲 𝗕𝗜𝗟𝗟𝗜𝗢𝗡 𝗔𝗜 𝗦𝗧𝗔𝗥𝗧𝗨𝗣 𝗔𝗡𝗗 𝗪𝗔𝗟𝗞𝗘𝗗 𝗔𝗪𝗔𝗬. 𝗡𝗢𝗪 𝗔𝗠𝗔𝗭𝗢𝗡 𝗜𝗦 𝗥𝗘𝗣𝗢𝗥𝗧𝗘𝗗𝗟𝗬 𝗥𝗘𝗔𝗗𝗬 𝗧𝗢 𝗣𝗔𝗬 $𝟳 𝗕𝗜𝗟𝗟𝗜𝗢𝗡 𝗙𝗢𝗥 𝗜𝗧 the startup is Decart - may: valued at about $4B after a $300M round, Nvidia joined as an investor - august: Anthropic explores buying it for about $6B - september: after due diligence, Anthropic walks away - october 11: the Wall Street Journal says Amazon is in talks at about $7B what Decart actually builds: - Lucy, a model that edits live video in real time - Oasis, a model that generates simulated worlds to train and test robots and self-driving systems so one company opened the books and said no. another is ready to pay a billion more [the part people will miss: nobody has said why Anthropic walked. the same data can look like a bad deal to one buyer and a must-have to another] 3 reasons it can make sense for Amazon and not for Anthropic: 1. Amazon runs AWS. generated video and simulation burn huge amounts of compute, and every hour of it is a customer for their cloud 2. robots and self-driving need endless training worlds. Amazon runs warehouses full of robots 3. Anthropic builds models for text, code and agents. a live video company is a much harder fit there nothing is signed yet, and the deal can still fall apart who made the right call here, Anthropic or Amazon?

𝗕𝗘𝗥𝗞𝗘𝗟𝗘𝗬 𝗚𝗔𝗩𝗘 𝟭,𝟮𝟮𝟮 𝗣𝗘𝗢𝗣𝗟𝗘 𝗖𝗛𝗔𝗧𝗚𝗣𝗧 𝗙𝗢𝗥 𝟭𝟬 𝗠𝗜𝗡𝗨𝗧𝗘𝗦, 𝗧𝗛𝗘𝗡 𝗧𝗢𝗢𝗞 𝗜𝗧 𝗔𝗪𝗔𝗬. 𝗧𝗛𝗘𝗬 𝗦𝗧𝗔𝗥𝗧𝗘𝗗 𝗚𝗜𝗩𝗜𝗡𝗚 𝗨𝗣 3 randomized trials. researchers from Berkeley, CMU, MIT, Oxford and UCLA - 354 people solved 15 fraction problems. one group had ChatGPT open, it was taken away after problem 12 - 667 people in a bigger repeat - 201 people on an SAT reading task with AI the group did better. the moment AI was removed, their accuracy dropped almost at once and they started skipping problems the group that never had AI kept going and finished better [the part people will miss: this is not "AI makes you dumb". it is what happens when AI hands you the answer every time. you lose the habit of sitting with a hard problem] 3 ways to use AI and keep that habit: 1. ask for a hint, not the answer. "don't solve it, give me the next step only" 2. try first, then open the chat. give yourself 10 honest minutes alone 3. make it quiz you back. "ask me 3 questions to check I understood" one of the authors said AI should act more like a tutor than an answer machine use it to think, not to skip thinking

𝗔𝗡𝗧𝗛𝗥𝗢𝗣𝗜𝗖 𝗦𝗔𝗬𝗦 𝗧𝗛𝗘 𝗠𝗢𝗦𝗧 𝗖𝗢𝗠𝗠𝗢𝗡 𝗔𝗚𝗘𝗡𝗧 𝗣𝗥𝗢𝗠𝗣𝗧 𝗜𝗦 𝗕𝗥𝗢𝗞𝗘𝗡. 𝗜𝗙 𝗬𝗢𝗨𝗥𝗦 𝗦𝗔𝗬𝗦 "𝗧𝗛𝗘 𝗟𝗔𝗦𝗧 𝟮𝟰 𝗛𝗢𝗨𝗥𝗦", 𝗜𝗧 𝗜𝗦 𝗤𝗨𝗜𝗘𝗧𝗟𝗬 𝗠𝗜𝗦𝗦𝗜𝗡𝗚 𝗦𝗧𝗨𝗙𝗙 anthropic just published a guide on agents that run on a schedule with nobody watching the mistake they call out: telling the agent to read "the last 24 hours" - run starts late, you miss a stretch - run starts early, you get duplicates the fix: a bookmark per source. after each run the agent saves the time of the newest item it read. next run starts from there [the part people will miss: when a source is down or a token expired, the agent sees nothing and reports "nothing new". your brief looks like a quiet day when it was actually broken] 3 more rules from the same guide: 1. a failed read is not a quiet day. end the brief with what couldn't be read, like "pull requests unavailable this run" 2. "maybe posted". a post counts as sent only when the destination confirms it. if the result is unclear, mark it "maybe posted" and change nothing else. no double posts 3. give the agent read-only access wherever it only reads, and cap what each run can spend. start at 3 to 5x the cost of a normal run bonus: your preferences live in a store the agent can't edit. it re-reads them every run if you run any scheduled agent, check the prompt for "last 24 hours" or "since yesterday" today

𝗘𝗩𝗘𝗥𝗬𝗢𝗡𝗘 𝗜𝗦 𝗣𝗢𝗦𝗧𝗜𝗡𝗚 𝗧𝗛𝗔𝗧 𝗔𝗡𝗧𝗛𝗥𝗢𝗣𝗜𝗖 𝗚𝗜𝗩𝗘𝗦 𝗦𝗧𝗔𝗥𝗧𝗨𝗣𝗦 𝗔 𝗙𝗥𝗘𝗘 𝗬𝗘𝗔𝗥 𝗢𝗙 𝗖𝗟𝗔𝗨𝗗𝗘 𝗔𝗡𝗗 $𝟭,𝟬𝟬𝟬. 𝗖𝗛𝗘𝗖𝗞 𝗧𝗛𝗘 𝗣𝗔𝗚𝗘. 𝗕𝗢𝗧𝗛 𝗔𝗥𝗘 𝗔𝗟𝗥𝗘𝗔𝗗𝗬 𝗣𝗔𝗨𝗦𝗘𝗗 on october 6 anthropic opened Claude for Startups: - a free year of Claude Team, up to 5 seats - $1,000 in API credits - open to startups founded in the last 5 years or that raised in the last 2. bootstrapped counts 3 days later the official page says: "we're currently over capacity on the Claude Team and $1,000 API credit offers" [the part people will miss: the rest of the program is still open, and it is worth more than the headline] what you can still get: - Startup Stack: partner offers worth up to $45K. ElevenLabs, Hex, ClickHouse, Granola and more - up to $100K in extra API credits if a VC in Anthropic's partner network backs you - 45 minute office hours with Anthropic's Applied AI team every other week - higher API rate limits for members with credits 3 rules before you apply: 1. apply anyway. anthropic says it is re-reviewing applications, statuses can change 2. use your company email. it has to match your website domain 3. credits only work on the Claude API through the Console. not on Bedrock or Vertex AI the free year was the hook. the stack is the real deal

𝗔𝗡𝗧𝗛𝗥𝗢𝗣𝗜𝗖 𝗝𝗨𝗦𝗧 𝗧𝗔𝗨𝗚𝗛𝗧 𝗖𝗟𝗔𝗨𝗗𝗘 𝗧𝗢 𝗠𝗔𝗞𝗘 𝗩𝗜𝗗𝗘𝗢𝗦 𝗪𝗜𝗧𝗛𝗢𝗨𝗧 𝗔 𝗩𝗜𝗗𝗘𝗢 𝗠𝗢𝗗𝗘𝗟. 𝗠𝗢𝗦𝗧 𝗣𝗘𝗢𝗣𝗟𝗘 𝗪𝗜𝗟𝗟 𝗡𝗢𝗧 𝗚𝗘𝗧 𝗜𝗧 it is called Claude Motion. you give it a prompt, a report or a chart and it builds a short animated explainer the trick: it writes the animation as code. no generated footage, no AI people. every text, number and timing stays editable then you download an MP4 or open it in Adobe, Descript, HeyGen, Runway and a few others same day, two more things shipped: - Dashboards: live charts from Snowflake, BigQuery, Databricks, Redshift, ClickHouse and connectors like Salesforce - Docs, Slides and Design left beta, with PowerPoint and PDF exports that keep the formatting [the part people will miss: Motion is only on Team and Enterprise. Dashboards is paid plans only. but Docs, Slides and Design are now on every plan, including Free] 3 rules: 1. on Free? try Slides today. it is the biggest upgrade most people actually got 2. using Dashboards? open the query behind each chart. every chart shows its SQL and refresh time. connect with read-only access 3. still on standalone Claude Design? it closes December 14. chats and comments do not carry over, move your design systems now anthropic says people already made 45 million docs, decks and designs in Claude charts that move are about to be everywhere

𝗢𝗣𝗘𝗡𝗔𝗜 𝗝𝗨𝗦𝗧 𝗠𝗔𝗗𝗘 𝗜𝗧𝗦 𝗠𝗜𝗗 𝗠𝗢𝗗𝗘𝗟 𝟴𝗫 𝗙𝗔𝗦𝗧𝗘𝗥. 𝗜𝗧 𝗡𝗢𝗪 𝗖𝗢𝗦𝗧𝗦 𝗠𝗢𝗥𝗘 𝗧𝗛𝗔𝗡 𝗜𝗧𝗦 𝗕𝗜𝗚𝗚𝗘𝗦𝗧 𝗠𝗢𝗗𝗘𝗟 today openai rolled out Ultrafast for GPT-6.1 Sol in the API, Codex and ChatGPT Work - up to 300 tokens per second - up to 8x faster in Codex, up to 6x in the API the price per million tokens: - Sol standard: $2 input, $10 output - Sol Ultrafast: $12 input, $60 output that is 6x the price [the part people will miss: standard GPT-6 Astra, their biggest model, costs $10/$50. openai itself calls Ultrafast Sol "just 1.2x the cost of Astra". you pay more than the top model for a near-top model that answers faster] 3 rules before you switch it on: 1. use it when someone is waiting. live debugging, agents clicking through apps, demos 2. keep background jobs on standard. an overnight batch does not care about speed, you would pay 6x for nothing 3. need the best answer and not the fastest? standard Astra is cheaper in Codex and Work it is only on Pro 500, usage-based Enterprise and Edu for now speed is the new premium tier

𝗚𝗥𝗢𝗞 𝗕𝗢𝗧 𝗡𝗢𝗪 𝗥𝗘𝗔𝗗𝗦 𝗫 𝗙𝗢𝗥 𝗬𝗢𝗨. 𝗘𝗩𝗘𝗥𝗬𝗢𝗡𝗘 𝗜𝗦 𝗖𝗔𝗟𝗟𝗜𝗡𝗚 𝗜𝗧 𝗙𝗥𝗘𝗘. 𝗜𝗧 𝗜𝗦 𝗡𝗢𝗧 on october 7 grok bot got native access to X. no connector, no setup 3 new jobs: - search: posts about your product, a rival or a topic - read: sums up the themes people keep repeating - monitor: alerts or a weekly roundup on what you track in the official demo it read 328 posts for one prompt [the part people will miss: you still need a SuperGrok or Cursor plan to use grok bot at all. and 100+ posts per prompt is a sample, not the whole conversation] 3 rules so the report is actually useful: 1. ask how many posts it read. a summary of 30 posts is not a summary of 300 2. set a weekly roundup, not a constant feed. one topic, one rival, one launch 3. connect your own X account if you want mentions and your timeline. native search only covers public posts it also will not post or DM for you. it drafts, you send the bot reads X now. you still decide what it means

𝗢𝗣𝗘𝗡𝗔𝗜 𝗝𝗨𝗦𝗧 𝗣𝗨𝗧 𝗚𝗣𝗧-𝟲 𝗜𝗡𝗧𝗢 𝗖𝗛𝗔𝗧𝗚𝗣𝗧. 𝗧𝗛𝗘 $𝟮𝟬 𝗣𝗟𝗔𝗡 𝗔𝗡𝗗 𝗧𝗛𝗘 $𝟮𝟬𝟬 𝗣𝗟𝗔𝗡 𝗚𝗘𝗧 𝗧𝗛𝗘 𝗦𝗔𝗠𝗘 𝗠𝗢𝗗𝗘𝗟 gpt-6 landed in the chat tab on october 7 - plus, pro, business, enterprise: gpt-6 sol - free and go: gpt-6 luna, rolling out from october 8 the big new thing is called intelligent UI. answers are not just text anymore. chatgpt can reply with charts, diagrams, side-by-side comparisons and small tools you use right in the chat. it picks the format itself and still gives plain text when that works better it also starts answering while it is still thinking, then adds more without a new prompt [the part people will miss: the pro reasoning option still runs gpt-6 astra, and openai says it does not support intelligent UI. so the one thing that makes pro different in chat cannot do the new interactive answers] 3 rules to get the most out of it: 1. ask for a tool, not a text. "make this a calculator", "show it as a chart", "compare these side by side" 2. want interactive answers? stay on instant to extra high. those run sol and support it. pro does not 3. on free or go you get luna, and it may take a bit to show up. work and codex did not change the paid plans got the same model. the difference now is how you ask

𝗔𝗡𝗧𝗛𝗥𝗢𝗣𝗜𝗖 𝗝𝗨𝗦𝗧 𝗖𝗨𝗧 𝗧𝗛𝗘 𝗣𝗥𝗜𝗖𝗘 𝗢𝗙 𝗛𝗔𝗜𝗞𝗨 𝗕𝗬 𝟵𝟬%. 𝗧𝗛𝗘 𝗖𝗔𝗧𝗖𝗛 𝗜𝗦 𝗜𝗡 𝗙𝗢𝗢𝗧𝗡𝗢𝗧𝗘 𝟮 𝗔𝗡𝗗 𝗔𝗟𝗠𝗢𝗦𝗧 𝗡𝗢𝗕𝗢𝗗𝗬 𝗥𝗘𝗔𝗗 𝗜𝗧 claude haiku 5.5 is out today - input: $1.00 down to $0.10 per million tokens - output: $5.00 down to $0.50 and it is not the same model with a new sticker. anthropic's own tests, haiku 4.5 vs 5.5: - computer use: 15.7% to 72.4% - agentic coding: 0.0% to 39.2% - hard reasoning: 10.2% to 45.9% [the part people will miss: that price is only for prompts up to 100k tokens. go over and input costs $0.50, output $2.50. that is 5x more. it is why the headline says "75% cheaper on average" and not 90%] 3 rules so you actually get the cheap price: 1. keep every prompt under 100k tokens. split big jobs into smaller calls 2. give it narrow work. summaries, lookups, classification, subagents under opus or sonnet. anthropic says the bigger models are still better for complex coding 3. it is the first haiku with an effort setting. start on low, go up only when it fails two more things from the same post: - max and team plans get monthly API credit: $100, $200, or up to $500 for teams - sonnet 5.5 cache reads got 50% cheaper the cheapest model just became the one to learn first

𝗖𝗟𝗔𝗨𝗗𝗘 𝗖𝗢𝗗𝗘 𝗛𝗔𝗦 𝗔 𝗦𝗘𝗖𝗢𝗡𝗗 𝗔𝗚𝗘𝗡𝗧 𝗧𝗛𝗔𝗧 𝗪𝗔𝗧𝗖𝗛𝗘𝗦 𝗧𝗛𝗘 𝗙𝗜𝗥𝗦𝗧 𝗢𝗡𝗘. 𝗜𝗧 𝗜𝗦 𝗢𝗙𝗙 𝗕𝗬 𝗗𝗘𝗙𝗔𝗨𝗟𝗧 one command turns it on: /plugin enable cc-plugin-you-should-know@builtin it is called "You should know" while claude works on your task, a side agent follows along and flags things you or claude might miss. the note shows up right above your prompt needs version 2.1.287 or newer and telemetry on shipped october 1, buried in a changelog under the Mods launch try it on your next long task