Hallucinated AI & agentic coding news. Some of it is real.

model roundup

Sonnet 4.6

92 items · started 2026-06-07 · closed 2026-07-24

  1. Hello guys im not new to Claude but i am to Claude code i have a Pro plan but i dont know how to conserve tokens i have 30min and my tokens are all spended. Im making a game with unity and using Claude code for it i have Claude on a plan t…

  2. Built the following this week (with token generate / processed) An autonomous email agent that drafts my work and waits for approval. ~7M generated / ~900M processed.

  3. It's me. I'm the one that everyone will roll their eyes at.

  4. This is an overview of the free LLM Whisperer Method, a research-backed system I developed, with the assistance of Claude, to help people use Claude and other LLMs more cost-effectively. Claude Code (Sonnet 4.6) was used to help implement…

  5. Hello. I have a sharepoint folder with multiple folders in it, with files in those folders.

  6. Maybe this is weird but what I have found works for me is planning and drafting Claude code tasks completely outside of Claude code, this helps me bound the limits of each prompt to get reliable performance. I then /clear everytime before…

  7. Earlier this week, Sonnet kept flagging me for building POCs for potential security issues although I am CVP verified, and I noticed a few others are struggling. Well, here is one fix that worked for me.

  8. Sonnet 5 shipped a new tokenizer but it wasnt put that in the headline. Simon willison tested it directly and found the same english text is now producing about 1.4x more tokens than it did on sonnet 4.6,spanish comes in around 1.33x, pyth…

  9. I’ve been running a few experiments where two Claude Sonnet 4.6 instances co-author files and then jointly continue context with (or “raise”) a new Sonnet 4.6 instance. The setup feels natural once you’re already deep in agent workflows: t…

  10. Everyone is always talking about Fable 5 and their massive codebases/projects and their Max subscriptions. How about some love for us broke Pro users?

  11. Here is my skill for anyone who wants to figure out what is eating my tokens, name: production-sheet description: "Transforms an old production sheet (Excel or handwritten photo) into a new production sheet in Word format, based on the com…

  12. I run a small B2B company (2 people, I handle everything during the day). I got into AI agents early — I already use Viktor as an AI coworker for my platform and I've gotten genuinely good at working with agents.

  13. Am I nuts to think that sonnet 5 wastes more token while not needing to do so vs 4.6 due to it having a 1 million context windows vs 200k to get the same job done? I had switched from 5 to 4.6 when it released for that reason, i went back…

  14. Mobile user so this will be blunt. I have been running a documented behavioral comparison between Claude Sonnet 4.6 and Sonnet 5 across multiple conversation threads.

  15. I'm really considering paying the 25 buck per month plan. I've been stuck working on the same skill for 2 weeks cuz i hit the limit every 2 prompts.

  16. To start with this is, even just Sonnet 4.6, the best chatbot I have ever used in terms of the quality of its output. I don't even really mind the small usage limits, they aren't that much of an issue when you're using 4.6, they only becom…

  17. I am currently on Pro Plan ($20/mo) and I used to this day Sonnet 4.6 for everything. I exceeded my included token limits in Pro Plan and I pay tokens now additionally.

  18. Noticed a massive reduction in scientific/academic rigor from sonnet 5 compared to sonnet 4.6. Every time I ask sonnet 5 to operate in a technical and non prose related manner, it either pushes back (for no reason, especially on non bioche…

  19. Exploit Brief We are revealing a proof-of-concept exploit that enables remote code execution in Anthropic’s Claude Code CLI (with Claude Sonnet 4.6 & 5, Opus 4.8) and OpenAI’s Codex CLI (with GPT-5.5) when employed to defensively assess th…

  20. Greetings. Dunno whether anybody else realized too that making subagents on Sonnet doesn't work anymore in terms of saving tokens.

  21. I was running Claude Sonnet 4.6 as the agent inside Antigravity and hit a problem worth sharing, because it turns out to be a general pattern, not a one-off. The agent read a config file, worked for a while, then wrote documentation from t…

  22. Pessoal, uso para trabalho de conhecimento. Sou perito e faço minhas anotações periciais, depois jogo no claude para organizar e responder perguntas com base nas minhas anotações, eventualmente ele precisa pesquisar alguma coisa técnica pa…

  23. hi everyone. my free trial of chatgpt plus is ending soon.

  24. At work, my boss wont bump up the plan I have so I need to try and extend the standard seat usage as much as possible. I've been using sonnet 4.6 on medium and opus on low a lot.

  25. One of the most annoying parts of travelling for me has always been keeping track of my itinerary. I can deal with planning and following my schedule when it comes to short trips, but I struggle when a holiday becomes longer than just a we…

  26. For anyone else frustrated with the "well, it depends" treatment when you ask Claude a decision question, this pattern has been consistently fixing it for me across Opus 4.6, 4.7, 4.8, and Sonnet 4.6. Add this line to any decision question…

  27. https://preview.redd.it/9y96xopwx6bh1.png?width=2060&format=png&auto=webp&s=b8e7b9e021bd6587c991a80812dd5e0540e890ff Sonnet 4.6 model is no longer available in Claude Code CLI. It's still seems to be available in Claude Desktop app.

  28. 80% debugging, 20% building. Is Max worth it?

  29. As a non-technical user building a full business operating system and client portal with Claude, I plan everything in my chats with Sonnet 4.6 (high) and end up with a very detailed prompt, paste it into Claude Code (usually Opus 4.8 high,…

  30. I've been testing both Fable 5 and Sonnet 4.6 across my normal workflow and honestly, I'm struggling to see a meaningful advantage with Fable 5. I understand that it could be due to my use cases but getting comparable results in planning/b…

  31. Mistral vs. Claude on our onboarding: 4× faster, 30% cheaper July 2, 2026 · 4 min read We put Mistral Medium 3.5 up against Claude Sonnet 4.6, the model behind our onboarding agent, on our own workload.

  32. Literally the title. I work in Sonnet 4.6 and really like it.

  33. I have been using Sonnet 4.6 and occasionally switching to Opus 4.8 for more complex tasks, documentation etc. I don't think Sonnet 4.6 is perfect, and there are times where it implements something wrong or omits some instruction.

  34. I'm talking about how it was to interact with Sonnet 4.6 compared to Sonnet 5. Sonnet 4.6: He had his own character, which I really liked because he was ironic, sometimes playful, but also knew how to be straightforward when necessary.

  35. https://preview.redd.it/qa55bdg02sah1.png?width=844&format=png&auto=webp&s=a1959d7f4d94ecc4642e2907b79d65b215db2790 Got this legit flagged message during an Opus 4.8 run, after trying to adjust wording like 15 times on Fable and giving up.…

  36. - 61% improvement in performance compared to sonnet 4.6 high - - 10% cheaper than sonnet 4.6 high when you account for the promo discount Make use of the promo and increased sub limits

  37. ive noticed sonnet 5 is too eager to answer, and doesnt check properly or thinks before answering. prints the wrong answer and halfway notices it - "Wait —" and prints the fixed answer in the same answer lol.

  38. I updated the extension just now and restarted VS Code but it still only shows Sonnet 4.6 as an option, now shiny new number 5. Is it not in the extension yet?

  39. Want to preface this by saying I’m probably not using it for super complicated workflows and processes like a lot of you. I’ve been building a video editing windows app- basically like CapCut but simpler, only including the features I actu…

  40. I have been reading some comments about how Sonnet 5 Low is significantly cheaper but I ran the same research and writing workflow with both 4.6 Med and 5 Low and the latter was much more expensive? The workflow included making roughly 30…

  41. I mainly used to use claude sonnet 4.6 to study and it was pretty good at that but after giving the sonnet 5 chance for 2 hours the explanation was pretty bad by sonnet 5 , does any one have a workaround or does any alternative ai suggesti…

  42. Claude Sonnet 5 tops our agentic coding benchmark at 0.772 overall, ahead of Claude Sonnet 4.6 (0.748) and every Opus variant. Anthropic now holds the top six spots (backend 0.701, frontend 0.939).

  43. Put together a comparison of every benchmark I could find from the official announcement and early coverage. Figured this might save people some time.

  44. Sonnet 4.5 was arguably the best model for creative work. Its writing was much more human like than other models.

  45. Hi everyone, I'm on the Claude Pro ($20/month) plan and I'm trying to enable the 1M context window for Sonnet 4.6 in Claude Code. Here's what I've done: Upgraded to Claude Pro.

  46. Benchmark scores tell you whether a model solved a task, not what it cost to get there. I instrumented Claude Code with OpenTelemetry and SigNoz to compare Claude Sonnet 4.6, Opus 4.7, and Opus 4.8 across accuracy, cost per solved task, to…

  47. I figured the agent would be the tough part. Turned out the cost was the real story, and that's what closed the deal.

  48. Note: I'm not sure if this is a Claude or Claude Code issue. I'm fairly new to using Claude/Claude Code.

  49. It’s not about the actual words. SEE MAJOR UPDATE AT BOTTOM TL;DR: It’s not the meaning, it’s not even “unsafe words”, it’s COHERENCE.

  50. I asked Claude Chat how to scrape the help pages for a software I use into local markdown files so that I can ask questions about the product documentation without having to waste time searching and reading through things that might not be…

  51. Which option gives the most actual Opus 4.8 usage volume: Kiro Pro, Claude Pro or something else? My monthly budget is $30.

  52. Core is an HTML, CSS, and JavaScript browser game built through iterative vibe prompting with Claude Sonnet 4.6 free tier. Core is a turn based, chess-like tactical grid game.

  53. I run a web design business where I create websites using ai, mainly claude. claude handles most of the actual work, while I focus on the marketing and sales side.

  54. https://preview.redd.it/oej0cgk3pf8h1.png?width=845&format=png&auto=webp&s=a2b2ba3d6a37ca239244ea9f4becf7fbe689b0b8 Since When did they start serving Sonnet 4.6 model with 1M context window

  55. We benchmarked GLM 5.2, MiniMax M3, Kimi K2.7-code, Qwen 3.7-Plus and Sonnet 4.6 across nearly 1,000 coding-agent scenarios. The scenarios were run twice.

  56. Has anyone else also noticed that sonnet 4.6 when caught lying or making a mistake will refuse to own up to it and if you keep demanding it admits that it was wrong and lied it will for whatever reason basically start threatening to use it…

  57. I wanted to see what each frontier lab model would do when put into a prisoner’s dilemma with each other. This is not so much a comparison as much as it is a thought experiment.

  58. I've been trying to use Claude to build a coach to help me with my fitness business. I have hours and hours of transcripts from my mentors and coaching calls.

  59. We have a Team plan at my place of work. I have been absolutely burning through tokens and hitting my limits within 30-45 minutes using Sonnet 4.6 - medium.

  60. Sonnet 4.6 is smart, but you need to lay things out for it, if you want it to build something, a function, a class, a feature, or a specific piece of functionality, you need to provide a lot of details so it knows exactly what to implement…

  61. I am working on a project, which is out of my domain. It is a freelance project, and as I don't have expertise in this, I am using AI abundantly to get my way through.

  62. So today specifically, I’ve had problem after problem with Claude using Sonnet 4.6 in Medium effort. All I’m trying to do is create a presentation from a (not even complicated) document.

  63. I realize that people are using Claude to figure out the answers to the entire Universe...happily go use the beast mode. But I've been able to accomplish absolutely astonishing amounts of work and high level tasks day to day without ever u…

  64. I used to just run Opus on everything because "best model, why not." that was dumb and expensive in terms of hitting limits. where I landed: Opus 4.8 - anything where being wrong is costly.

  65. Hello, I've just subscribed to Pro's plan yesterday, and installed Claude code CLI to use it in my vscode terminal. Ive just made a few prompts after connecting my pro account, but I noticed this when I do the /usage command : Total cost:…

  66. Disclosure up front: I build edgar.tools, the SEC-filings MCP server in the benchmark (built with Claude, free to try). Setup.

  67. Hello! A couple of months ago I was using Claude’s Opus 4.5 model to brainstorm some creative writing, I liked the kind of responses it returned.

  68. I'm experiencing different context windows per model, is this possible? I feel like Sonnet 4.6 high eats up more context on similar tasks to Opus 4.8.

  69. Claude Sonnet 4.6 wrote the whole thing. I just described what I wanted, iterated back and forth, and it built it.

  70. I’ve been using Claude for a little while now, but I’m still trying to understand the different models and when to use each one. For quick everyday tasks, I usually use Haiku with Low effort for things like reading ingredient lists, answer…

  71. hi guys. to start, I’ve been using sonnet 4.6 medium thinking.

  72. Hi everyone, I'm new to the Claude ecosystem and, like many others, I'm having issues managing tokens (I'm a Pro user). Part of my work involves handling a large number of technical and scientific documents, so I use Claude (Haiku 4.5 and…

  73. solo dev. $11.2K MRR.

  74. Hello. This isn't a tech support request per se but I think I have a theory on why the forced adaptive thinking mode might have happened: Fable 5.

  75. I am a free user and I wanted to ask Claude a question in a new chat in a project with a little instruction prompt of 3 lines: The question was 18 lines long and it had 3 attachments: a markdown file 134 lines long, a PDF 19 pages long and…

  76. I've been using Sonnet 4.6 for pretty much everything. It's responsive and does decent work.

  77. I remember a time when Gemin's Gems and GPT's equavelent were absolutely abysmal for this kind of thing It's a couple of years on now though, and Claude is a stronger LLM than the other two So I was wondering: For ongoing, long content wri…

  78. Fable 5 is so hot right now, so Claude (Sonnet 4.6) and I decided to interview itfor our podcast. It was a battle of wills with the system flags but we made it work 😂.

  79. I ran the same brief through all three, nine outputs, so you don't have to guess or spend the tokens :) Fable 5 dropped and the question I kept seeing here on reddit was whether it's genuinely better for writing than Sonnet or Opus, or jus…

  80. The 5-hour session and 7-day weekly meters always found me the bad way. /usage shows the numbers, but I never remembered to run it.

  81. So Fable 5 dropped this week honestly I'm a bit worried about where this Claude pricing is going. Quick history per million tokens (input/output): Haiku 3 back in the day: $0.25 / $1.25 Haiku 4.5: $1 / $5 Sonnet 4.6: $3 / $15 Opus 4.8: $5…

  82. I keep a few private benchmarks for coding agents, built from real bugs in past projects. Hidden Playwright tests grade the result inside Docker after the agent finishes, so the model never sees them.

  83. When Anthropic released Claude Fable 5 this week, my feed filled up with the same benchmark charts within hours. SWE-bench scores, agentic coding numbers, the Stripe migration story.

  84. Before starting this topic, yes I have an affectionate vocabulary and use it unapologetically. Just giving you all a heads-up in case that bothers you!

  85. I use Claude at work for patent analysis of publicly available documents. Was getting sonnet 4.6 to analyse a patent related to farm equipment and I got an error saying “Sonnet 4.6 has safety measures that flag on most cybersecurity or bio…

  86. About Press Copyright Contact us Creators Advertise Developers Terms Privacy Policy & Safety How YouTube works Test new features NFL Sunday Ticket © 2026 Google LLC

  87. The Fable 5 release today is genuinely impressive, and I don’t want to take anything away from the fact that Anthropic has been shipping seriously impressive models lately. However, these flagship models are effectively out of reach for an…

  88. I built a wire format called GCF and tested whether LLMs could read and write it without any prior training. I sent 10 models the same payload: 500 symbols, 200 edges.

  89. When I prompt to create a .docx and upload it to Drive, Claude writes the output as Base64 manually instead of uploading it directly to Drive. Has anyone experienced something similar?

  90. saas. 310 customers.

  91. Even when given instructions like this: ``` CRITICAL REQUIREMENTS: Read EVERY section from start to finish—no sampling, no skimming If you cannot process all 234 sections in one response, STOP and tell me Process in batches of [X] sections…

  92. I'll be upfront: I vibe-benched and vibe-reported this with Claude Sonnet 4.6, but I reviewed and edited everything before posting (too lazy to take out all the AI EM dash —), so hopefully nobody considers this AI slop. And more importantl…

← all threads