GitHub Outage Map

Community Discussion

Tips? Frustrations? Share them here. Useful comments include a description of the problem, city and postal code.

Beware of "support numbers" or "recovery" accounts that might be posted below. Make sure to report and downvote those comments. Avoid posting your personal information.

GitHub Issues Reports

Latest outage, problems and issue reports in social media:

• 
  CodeGlitch (@codeglitch) reported 1 minute ago

  A new AI model launched today small enough to run on your phone. Its own launch page said the tradeoff was tiny. Its own benchmark table said something else. Nobody lied. The real number was sitting right there in the table. Most people just never open the table. Today's lesson (full breakdown inside) Why you should read a model's benchmark table, not its summary sentence, and the one-line prompt that does the reading for you. Also in today's brief - Bonsai 27B: a 27B model that fits on an iPhone (Apache 2.0, out today) - Codex + ChatGPT Work passed 8 million users, usage limits reset again - Anthropic launched Claude for Teachers (free for verified US K-12 educators) - GitHub Copilot can now fix a security alert and validate it before opening the PR - Perplexity open-sourced a benchmark for research agents, still far from solved The full lesson and the day's verified brief are inside AI Coding & Agents HQ. A new one every day. Join link below.

• 
  Big Bone © (@steb0ne) reported 2 minutes ago

  @trikcode Maybe I'm slow but can't you just push the commit to github then move to the other platform and say check my progress before you continue?

• 
  Mikhail Rogov (@i_mika_el) reported 3 minutes ago

  @atikursatter Are people already complaining about this problem in GitHub issues? I would start there.

• 
  Siriö Astaröt (@Sirio__Astarot) reported 6 minutes ago

  Feature request @genspark_ai : - Install Skill from CLI and Github directly, ex: $ npx bla bla bla *** clone repository FIX BUG: - Skills uploaded in .zip format and skill.md files are not installed.

• 
  Froxxxie (@Froxxxie) reported 6 minutes ago

  THIS GUY just collected $6,480 from an app their AI team built in minutes. While most people are still sitting in a chat tab, typing prompts and praying the output doesn’t break. In the video you see the full flow: Research Agent, Coding Agent, and Testing Agent working together — planning, building, catching runtime errors, and fixing the code themselves. They ship a ClientHub with invoicing. Dashboard shows real payments coming in. Full code ownership. Export to GitHub. One-click native iOS and Android. Authentication, database, payments, and hosting are already handled. Some people are using this to take client projects they used to turn down. Others are launching small SaaS products that would’ve taken months to code manually. The uncomfortable truth is this: Prompting one model was never the endgame. It was just the training wheels. Most people are still scared to take them off. While others are already collecting real money from what their AI team ships in a single evening.

• 
  Franci Penov (@francip) reported 8 minutes ago

  Doing the same work twice now, because neither the iOS ChatGPT app, nor Codex Web is able to push my changes to github, despite the connector installed and configured with read/write permissions for my org. As much as I love Sol, the coding tools around it are in dire need of someone at @OpenAI actually using Sol to fix them and make them usable.

• 
  felipecsl (@felipecsl) reported 10 minutes ago

  openai can't be serious about codex cloud. it's utterly useless, can't even pull a github issue from the repository. how do people use this? no scheduled tasks either. anthropic is years ahead

• 
  Mark Ajzenstadt (@mardehaym) reported 11 minutes ago

  A startup CTO in Portland runs coding agents on every pull request. 8 developers, 15 agent sessions each, every day, 50,000 output tokens per session. By API on GPT-5 Mini: 6 million output tokens a day. $12 a day. $4,320 a year. An ML engineer in Seoul debugs training pipelines with tool-use agents. 200 calls a day, 4,000 reasoning tokens each. By API on Claude Haiku 4.5: $4 a day, $1,440 a year. His training code leaves his network 200 times a day. A PhD student at ETH Zurich runs 50 coding agent experiments a day. She needs Fable 5 quality reasoning. 10 million output tokens a day. By API on Fable 5: $500 a day. She burned her $3,000 monthly research budget in 6 days. All three pay a cloud provider to reason for them, per token, per request, on repeat. MiniCPM5-1B-Claude-Opus-Fable5-Thinking is a 1 billion parameter model fine-tuned on Fable 5 thinking data, built on OpenBMB's MiniCPM5-1B. Chain-of-thought reasoning, code generation, debugging, native tool calls, 131,072 token context. English and Chinese. Runs on any GGUF runtime. You give it a coding task, get reasoning and a solution back on your machine, with no API key and no internet connection. Anthropic charges $50 per million output tokens for Fable 5 reasoning patterns. GnLOLot fine-tuned those patterns into 1 billion parameters. 688 MB at Q4_K_M. 1.15 GB at Q8_0 (recommended). LlamaForCausalLM architecture, no custom kernels. Every GGUF runtime loads it without modification. Two modes. Thinking: temperature 0.9, top_p 0.95, full chain-of-thought. No-Think: temperature 0.7, fast responses. Run it with llama.cpp, Ollama, LM Studio, vLLM, Docker, Jan, or KoboldCpp. One command each. llama.cpp server gives you an OpenAI-compatible endpoint on localhost. Point any agent framework at it. OpenBMB released MiniCPM5-1B on May 19, 2026. It scored 42.57 average across reasoning, code, math, logic, and agentic benchmarks. The next best 1B model scored 35.61. GnLOLot fine-tuned it on Fable 5 thinking data and published GGUF quantizations on HuggingFace. 9,800 stars on GitHub. Apache-2.0. Claude Fable 5: $10/$50 per million tokens. Your code travels to Anthropic's servers. GPT-5 Mini: $0.25/$2 per million tokens. Cloud-only, no Fable 5 reasoning. Claude Haiku 4.5: $1/$5 per million tokens. Not local, not private. This model costs $0. Your code stays on your machine. That CTO in Portland runs 120 agent sessions a day on one office server now. $0. His team's codebase never touches an external API. The ML engineer in Seoul moved his 200 daily calls to a machine under his desk. He keeps $1,440 a year. The PhD student in Zurich runs her experiments on a university workstation. Her $3,000 monthly budget lasts the semester. You get Fable 5 reasoning in a 688 MB file, running on hardware you already own, for nothing.

• 
  Bluesky 555 (@Co73261) reported 11 minutes ago

  @FurkanGozukara Here’s what is said about you on the net : Critics in parts of the open-source Al community have accused him of excessive self-promotion, aggressive Patreon marketing, and spamming GitHub issues and Reddit. There are long Reddit discussions criticizing these practices, including

• 
  Tanvi (@tanviiiw) reported 18 minutes ago

  More tools ≠ smarter agent. GitHub cut Copilot's built-in toolset from 40 tools to 13, and found the full toolset was actually costing them 2-5 percentage points on SWE-Lancer. Their words: "giving an agent too many tools doesn't always make it smarter. Sometimes it just makes it slower." Speakeasy experimented further: 107 tools in one server, and the model started hallucinating endpoints that didn't exist. Trim it to 10-20 well-chosen tools and it got most calls right. It comes down to two things: every tool definition eats context on every single request, and models fuzzy-match on names, so get_status / fetch_status / query_status all blur together and it picks wrong. But we keep connecting everything anyway, because it feels like giving the agent superpowers (I fell for this too). It doesn't. Access isn't capability. You connect more tools to save time, then spend that time babysitting the tool calls. (Of course, none of this replaces a well-scoped prompt. It's upstream of it. You can write a perfect prompt and still lose to a bloated toolset.) So TLDR; curate the toolset like you'd curate a team.

• 
  Omar Farooq (@potencytoact) reported 20 minutes ago

  The GitHub issue was filed by ignatremizov. He is not asking OpenAI to revert encrypted delivery, only to add a plaintext audit copy of the delegated task, persisted in the local rollout history. One correction to a claim he made on HN, that the prompts already pass through the client for the terminal UI to display, which would make the fix a trivial persistence change. Other users disputed this, and the envelope PR settles it: the payload arrives at the client already encrypted. Only the routing header and child results pass through readable. The plaintext never touches your machine, so restoring auditability requires OpenAI's backend to send something it currently withholds. If OpenAI declines, it will be a decision about the backend, not an oversight in the client.

• 
  Mamba (@opeyemi_ii) reported 22 minutes ago

  5/ Built with: Vercel + GitHub (hosting), Airtable (CRM), Zapier (automation), Paystack (payments), Google Calendar + Gmail (client comms). Small businesses deserve real automation too. This is what I build. Running a service business without this? Let's fix that, DM me.

• 
  SaSame (@SRLsasame) reported 27 minutes ago

  7. Preliminary interpretation The available evidence supports the following limited conclusions. Confirmed ・The public MCP endpoint is reachable. ・The endpoint completed MCP initialization. ・The server negotiated protocol version 2025-11-25. ・The server identified itself as vibekit version 0.7.2. ・tools/list succeeded. ・36 tools were listed. ・The listed tools had typed input schemas. ・The observed tools carried applicable annotations. ・An unauthenticated tool request produced a structured authentication error. ・An unknown method produced a structured JSON-RPC error. ・The protocol and schema observations were materially consistent across July 11–14, 2026. Not confirmed ・Whether a newly issued API key currently works. ・Whether authenticated vibekit_list_apps returns substantive account data. ・Whether deployment tools complete successfully. ・Whether GitHub authorization is correctly enforced for every repository. ・Whether environment-variable values are redacted or exposed under all client configurations. ・Whether database queries enforce read-only behavior in every case. ・Whether destructive tools require confirmation at the server layer. ・Whether task execution, deployment, rollback, and QA are continuously available. ・Whether the observed 36-tool surface remains unchanged over time. ・Whether all tools behave consistently across every supported MCP client.

• 
  Aaron Delasy (@aarondelasy) reported 33 minutes ago

  @zeeg somehow it seems agents are really bad at configs, they have to test and re-test everything multiple times to get it to work. I remember I had a problem in github actions, and for a single line change it took agent around an hour and 300k tokens

• 
  Lukas Helminger (@luhelminger) reported 34 minutes ago

  @Youssef23__ I don’t think there are a ton of publicly open positions. The ones that are hiring can basically choose who they want as many infra projects had to scale down. In particular on the non-senior (applied) research side. But one can tell those companies from actively pushing (eg GitHub)