Claude outages, self-hosted agents and two open-weight models

Friday 13 November was a bad day to depend on one AI provider. Sign-in to claude.ai and Claude Code failed for about two hours, and a second incident that also covers the Claude API opened at 13:09 UTC and was still open when this post went out. The day before, Anthropic opened a private beta of self-hosted Claude Managed Agents, and two labs published open weights: Thinking Machines' Inkling 2 Small and Edge0's 145B model.

The three stories answer one question from different sides: what do you run when your usual model is not there? Here is what is confirmed in each, and what a fallback really costs.

Claude on 13 November: two incidents in one day

Everything in this table is from Anthropic's status page. All times are UTC.

Time (UTC)	What the status page said
09:58	Investigating "an issue with users authenticating to Claude.ai and Claude Code"
11:05	A fix has been deployed, monitoring for recovery
11:58	Resolved. The note names claude.ai and Claude Cowork
13:09	New incident: investigating degraded performance on claude.ai, the Claude API and Claude Code
13:52	Update: the investigation now also covers connectors, including Claude Docs and Claude Sheets

What that does and does not tell you:

• The first incident was about signing in. Anthropic lists claude.ai, the Claude Console, Claude Code and Claude Cowork as affected. The Claude API is not on that list.
• It lasted two hours, not ten. One news roundup I read described a ten-hour outage. The status page shows two hours between the first notice and the resolved notice. It gives no earlier start time.
• The second incident is wider. It adds the Claude API, and a 13:52 update extended it to connectors. At about 13:55 UTC the status page showed claude.ai as a major outage, the Claude API and Claude Code as partial outages, and Claude Cowork as degraded.
• The cloud routes stayed green. The same page showed Claude Platform on AWS, Claude in Microsoft Foundry and Claude Models on Google Cloud as operational.
• No cause has been given, and neither incident page mentions a post-mortem.

There is one more open item that is easy to mix up with these. Since 12 November, Claude's Microsoft 365 connector has been failing, and Anthropic's notice puts that down to a Microsoft 365 service incident. Some people on Hacker News guessed that the sign-in failure had the same root. Anthropic has not said so, and its status page lists them as separate incidents.

What to do when Claude Code will not respond

1. Look at the status page before you debug your own machine. A sign-in loop or an "OAuth request failed" message during an incident is not your configuration.
2. Know how your Claude Code signs in. A subscription login, an API key and a cloud provider account are three different routes, and today they failed differently. The morning incident was about authentication and did not list the API. The afternoon one did.
3. Set up a second route before you need it. Claude Code can run through Amazon Bedrock, Google Cloud or Microsoft Foundry instead of Anthropic's own sign-in, and those were the entries that stayed operational on the status page today. I did not test a switch during the incident, so treat that as what the page reported, not as something I measured.
4. Keep a fallback from a different company too. A second route to the same model does not help if the model provider itself has the problem. One AGENTS.md that any agent can read makes that switch cheap. The Cursor vs Claude Code post shows the setup, and yesterday's post covers the newest cheap candidate, Gemini 4 Hydrogen.
5. Commit before long agent runs. An agent that loses its connection halfway through a change leaves you with a working tree to untangle. A commit before each run turns that into one git diff.

Self-hosted Claude Managed Agents: the agent runs in your cloud

Claude Managed Agents is Anthropic's hosted service for running agents: Anthropic runs the agent loop, keeps the session history and, until now, ran the service even when you supplied your own sandbox. On 12 November it announced a version where, in its words, "every part of the service runs in your own cloud or on-premise".

What you run, and what still leaves your network

• You run a server, a PostgreSQL database and a set of workers. The workers run the agent's shell and file tools inside your sandboxes. Session history sits in your database, and credentials can live in your own secrets manager.
• Model calls still go out. For each call, the server sends the conversation so far, including tool inputs and results, to an inference provider. The default is the Claude API. You can point it at Amazon Bedrock, Google Cloud, Microsoft Foundry or your own model gateway.
• The API is the same as the hosted service, so the SDK and the ant CLI work against your server.
• Zero data retention carries over. Anthropic says an existing ZDR arrangement applies to these model requests like any other Messages API call. The hosted service is not eligible for ZDR.

What is missing

Anthropic's "differences" page is long, and worth reading before you plan around this. The ones that would change a design:

• No webhooks. Workers poll the server.
• No Files API, and no prebuilt docx, pdf, pptx or xlsx skills. You upload your own versions.
• No scheduled deployments.
• Web search and web fetch are off by default, and whether they work at all depends on the inference provider you pick.
• MCP servers and Git hosts must be on an allowlist that starts empty.
• The worker does not isolate the agent's shell. The docs say the bash tool is a real shell with the worker's permissions and network access. Isolation is your job.

Who can use it

It is a private beta, and the only way in is to ask your Anthropic account team. I found no price and no list of eligible plans on the announcement, the docs or the release notes. The quickstart asks for a host with at least two CPUs, Docker with the Compose plugin and a Claude API key.

For a solo developer this is not something to adopt this week. It matters for a different reason: it separates where the agent runs from where the model runs. Today's incident shows why that split is useful. An agent service in your own cloud, with model calls routed through a cloud provider, has a different set of things that can fail than one login page.

Two open-weight releases, and what they need to run

Both of these landed on 12 November. Neither is a drop-in replacement for the model inside your coding agent, and the reasons are different.

Inkling 2 Small from Thinking Machines

Thinking Machines announced Inkling 2 Small on 5 November and added a note on 12 November that the weights are on Hugging Face. The model card gives the licence as Apache 2.0, with a link to an acceptable use policy.

• Size: 594 billion parameters in total, with 17 billion active per token. It takes text and images in and gives text out, with a context window of up to 1 million tokens.
• Hardware: the model card asks for at least 500GB of combined GPU memory in the smaller format, and suggests two B300 or four H200 cards. This is a model for a server rack or a hosting provider, not a workstation.
• The lab's own numbers: 30.6 on the Artificial Analysis index, where the best model in the same table scores 36.5, and 72.1% on DeepSWE v1.1 against a best of 77.0%. The lab chose the comparison set, which is mostly small and fast models.
• One loose end: the safety section of the launch post still says the weights release is "pending further safety evaluations", while the update note above it says they are out. The model card says safety evaluations ran ahead of release.

What an Apache 2.0 licence gives you here is the right to have it hosted by anyone, including yourself. It is also available for sampling and fine-tuning on the lab's Tinker service. The pages I read list no Tinker price.

Edge0-145B

Edge0's pitch is the opposite: a 145 billion parameter model on your own computer, with about 10 billion parameters doing the work on each token. The launch post claims more than 80 tokens per second and a "reported peak memory of 29GB".

Read the rest of the post and the model page before you trust that number:

• The 29GB is not the whole footprint. The same post says an expert library of about 40GB sits in system memory and about 48GB of vocabulary data sits on the SSD. Edge0's own measurement on an RTX 5090 D with 24GB shows 23.25 GiB of GPU memory plus 38.27 GiB of system RAM.
• Mac support is unclear. The post mentions a MacBook Pro and the model page lists speeds on an M2 Ultra, but the build instructions target NVIDIA cards with CUDA 13 on Linux or Windows, and the repository's platform table leaves macOS blank.
• It is not open source. The weights use the Edge0 Community Model License 1.0, and the model page says commercial products and services need a separate written licence.
• The benchmarks are Edge0's internal runs, and the post's summary is kinder than its own table. On GPQA Diamond, for example, Edge0-145B scores 90.4, below all three of the larger models it is compared with.

The part I find useful is the server. Edge0's tool starts a local endpoint that it says speaks both the OpenAI and the Anthropic API formats, which is the shape a local fallback needs. I have not run either model, so everything in this section is from the vendors' pages.

What a fallback really costs

Fallback	What it protects against	What it costs you
Same model through a cloud provider	A sign-in or API incident at the model company	A cloud account, set up in advance
A model from another company	A provider-wide outage, or a contract or court decision	A second subscription or key, and portable instructions
Hosted open weights	A vendor withdrawing or repricing a model	Finding a host, and a weaker model for now
Open weights on your own machine	Any network or vendor failure	A 24GB GPU and about 40GB of RAM, plus a licence check

The first two rows are an afternoon of setup. If you do one thing after today, do those.

The short version

• Claude had two incidents on 13 November. Sign-in to claude.ai and Claude Code failed from 09:58 to 11:58 UTC. A second incident covering claude.ai, the Claude API and Claude Code opened at 13:09 UTC and was unresolved at 13:55 UTC. Anthropic has given no cause.
• Self-hosted Claude Managed Agents is in private beta through Anthropic account teams. The server, database, workers and session history run in your infrastructure. Model calls still go to an inference provider you choose. No price is published.
• Inkling 2 Small is Apache 2.0 and needs at least 500GB of GPU memory. Edge0-145B runs on one machine but needs far more than its headline 29GB, and its licence restricts commercial use.
• The practical step is a second route to your model and one tested fallback from another provider, set up before the next outage.

Frequently asked questions

How long was the Claude outage on 13 November 2026?

Anthropic's status page shows the first incident, a sign-in failure on claude.ai and Claude Code, running two hours from the first notice at 09:58 UTC to the resolved notice at 11:58 UTC. A second incident covering claude.ai, the Claude API and Claude Code opened at 13:09 UTC and was still open, with no fix announced, at 13:55 UTC, when this post was published.

Can I use self-hosted Claude Managed Agents today?

Only by request. Anthropic opened it as a private beta on 12 November 2026 and its docs say to contact your Anthropic account team for access. The pages read for this post state no price and no list of eligible plans.

Can Inkling 2 Small or Edge0-145B run on a laptop?

Inkling 2 Small cannot. Its model card asks for at least 500GB of combined GPU memory. Edge0-145B is built for one machine, and its own measurement on a 24GB RTX 5090 D shows 23.25 GiB of GPU memory plus 38.27 GiB of system RAM, which is more than the 29GB figure in the launch post's headline claim.