AI error messages, explained
Paste the line your terminal or app printed and find out what it means. Every answer comes from the product's own documentation, or from the strings in its shipped code.
969 lines indexed across 185 pages. Partial text works, so a few distinctive words from the middle of the line are plenty.
safeguards flagged this message (https://www.anthropic.com/legal/aup). This sometimes happens with safe, normal conversations.
Anthropic's safety filters flagged your prompt against its Usage Policy, not the cyber or biology rules. The reasoning_extraction detail and the line about safe conversations mark it as a likely false positive, which Anthropic says its filters can produce.
Error loading webview provided by 'anthropic.claude-code': Error: Could not register service worker: InvalidStateError: Failed to register a ServiceWorker: The document is in an invalid state..
VS Code couldn't start the background helper that every extension panel needs, so the panel stays blank. The extension named in the message is just the one you opened; it isn't the cause.
Something went wrong
Google doesn't publish what the numbers mean. Going by the threads, they fall into three groups: one chat that's stuck, one account that's blocked, or an outage hitting everyone.
failed to initialize the context: failed to allocate buffer for kv cache
The model's weights fit, but the KV cache (the memory that holds every token of the context) couldn't be allocated at the context length requested. It's an out-of-memory error, not a broken model file.
Failed to create unified exec process: helper_unknown_error: setup refresh had errors
Codex's Windows sandbox re-checks file permissions before each command, and that check failed, so Codex refuses to run anything. Since the October 8 update the usual reason is a file Codex itself is holding open.
Opus 5.5's safeguards stopped the response above · continuing once with that noted
Anthropic's safety classifier stopped Claude's reply partway through. Claude Code kept the part that had already arrived, told the model not to repeat the stopped content, and sent the turn one more time on the same model.
The 'gpt-5.3-codex' model is not supported when using Codex with a ChatGPT account.
OpenAI has retired that model for Codex when you're signed in with a ChatGPT account. Your request still names it, often from a saved setting, so the server refuses it.
Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check
Before it starts, the WebUI asks PyTorch whether an Nvidia CUDA GPU is usable. PyTorch said no, so the launcher stopped.
NansException: A tensor with all NaNs was produced in Unet.
The image model's output came back as NaN (not a number, a value that means the math broke), and the WebUI stopped instead of handing you a black square. Running in 16-bit half precision is the usual cause.
Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!
One calculation got two inputs where one sat in GPU memory (cuda:0) and the other in system memory (cpu). PyTorch won't mix them, so it stopped. The code that forgot to move a tensor is usually a custom node, an extension or one model file.
CUDA_HOME environment variable is not set. Please set it to your CUDA install root.
Something tried to compile CUDA code against PyTorch, and PyTorch couldn't find the CUDA toolkit (the kit with Nvidia's nvcc compiler) anywhere it looks. A working GPU and a CUDA build of torch aren't enough for this; the compiler has to be installed too.
Error: 500 Internal Server Error: unable to load model:
Ollama handed the model file to its bundled copy of llama.cpp and llama.cpp gave nothing back, so Ollama reported the blob's path and stopped. It almost always means that copy of llama.cpp doesn't know the model's architecture yet.
Error: Post "http://127.0.0.1:11434/api/chat": EOF
The ollama command sent your request to the Ollama server on port 11434 and the connection closed before a single byte of reply came back. EOF means end of file: the data stopped. The server almost always crashed or was killed while handling the request.
Error surveying hardware
Each LM Studio runtime (the llama.cpp build for CUDA, ROCm, Vulkan or CPU) checks your hardware when it starts. This label means that check crashed, usually because the runtime's library wouldn't load with your GPU driver, so the runtime can't be used.
Something went wrong (1095)
Gemini refuses every prompt with this code, often after a long or heavy session. Google hasn't said what 1095 means; Product Experts in its forum treat it as a stuck sign-in session on your account.
Your request has been blocked as our system has detected suspicious activity from your account.
Cursor's anti-fraud system flagged the request and refused it. Staff say it exists to stop free trial abuse, and that it sometimes catches people who've done nothing wrong.
Too many free trial accounts used on this machine. Please upgrade to pro.
Cursor's anti-abuse check thinks more than one free account has been used on this computer, so it blocks free and trial requests from it. Paid accounts aren't affected.
Your account has been disabled after an automatic review of your recent activities.
Anthropic's systems banned your Claude account. Its help center lists three reasons it bans accounts, but the message doesn't say which one applied to you.
Due to unexpected capacity constraints, Claude is unable to respond to your message. Please try again soon.
Claude's servers are busy across all users, so your message wasn't answered. Anthropic calls this load management and says it clears within minutes.
Unusual activity has been detected from your device. Try again later.
ChatGPT's anti-abuse checks didn't trust the request from your device or network, so it refused to answer. OpenAI says the message appears when traffic looks automated or unusual.
Unable to load conversation
ChatGPT couldn't fetch that chat's history from OpenAI's servers, so it won't open. The chat usually isn't gone: it's still listed, and in many reports it came back later.
Conversation not found
ChatGPT asked its servers for a chat and they answered that it doesn't exist (a 404). Most reports come from Temporary Chats, which aren't saved, failing on the second message.
A network error occurred. Please check your connection and try again. If this issue persists please contact us through our help center at help.openai.com.
ChatGPT lost its connection to OpenAI's servers before the reply finished. Your internet can be fine: a VPN, a security filter, an outage or a long reasoning reply can all end this way.
<model>'s safeguards flagged this message
Anthropic's real-time cyber safeguards on Opus and Sonnet models blocked a request that looked like prohibited or high-risk cybersecurity work. Dual-use work can be unblocked through the Cyber Verification Program; prohibited use can't.
Model architectures [...] are not supported for now.
The architecture named in the model's config.json isn't in this vLLM's list of model classes, and vLLM couldn't run it through its Transformers fallback either.
As of transformers v4.44, default chat template is no longer allowed, so you must provide a chat template if the tokenizer does not define one.
vLLM looked everywhere it knows for a chat template (the Jinja file that turns a list of messages into one prompt) and found none, so it can't serve a chat request for this model.
RuntimeError: CUDA error: no kernel image is available for execution on the device
The GPU code inside vLLM or PyTorch was compiled for other GPU generations than yours, so there's nothing the card can run when vLLM asks it to.
RuntimeError: Cannot re-initialize CUDA in forked subprocess
Your Python process had already started using the GPU when vLLM copied it to make a worker. PyTorch won't let a copied process use the GPU, so the worker dies on its first CUDA call.
Bfloat16 is only supported on GPUs with compute capability of at least 8.0
vLLM was told to run the model in bfloat16, a 16-bit number format that needs an NVIDIA GPU from the Ampere generation or later (compute capability 8.0). Your card is older, so the worker stops before loading anything.
"auto" tool choice requires --enable-auto-tool-choice and --tool-call-parser to be set
Your request included tools, which vLLM treats as tool_choice="auto", but the server was started without tool-call parsing, so it has no way to turn the model's output into tool calls and refuses the request.
must be divisible by tensor parallel size
Tensor parallelism gives each GPU an equal share of the model's attention heads. Your GPU count doesn't divide the head count evenly, so vLLM stops before loading anything.
No model is loaded. Please select one in the Model tab.
You sent a message before any model was in memory. Picking a model in the dropdown only selects it; nothing loads until you press Load, or until a load that failed is fixed.
This request requires more credits, or fewer max_tokens.
Before running a request, OpenRouter reserves enough credit to pay for the longest reply you allowed. Your balance (or your key's limit) can't cover that reservation, even if it would cover the reply you'd actually get.
No endpoints found that support tool use.
Your request included tools (function calling), and OpenRouter couldn't find a provider for that model which both handles tools and passes your routing and privacy settings.
Unsupported value: 'temperature' does not support 0.7 with this model. Only the default (1) value is supported.
You set temperature on an OpenAI reasoning model. Those models only accept the default value of 1, so any other number is refused before the model runs.
Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.
The request asked for a newer OpenAI model (a reasoning model such as the o-series, GPT-5 or GPT-6) and set the old max_tokens field, which those models refuse.
Your organization must be verified to stream this model.
OpenAI keeps some models and features behind organization verification, an identity check on the account that owns the API key. Your key's organization hasn't passed it, so the request was refused.
The model `gpt-5.5` does not exist or you do not have access to it.
The API couldn't match the model name to one your API key is allowed to call. It deliberately doesn't say whether the model is missing or just off limits to you.
Our servers are currently overloaded. Please try again later.
The model you asked for doesn't have spare capacity at the moment. It's about OpenAI's servers, not your usage limits, and it clears on its own.
Open WebUI Backend Required
The Open WebUI page loaded in your browser, but its first request to its own server, for /api/config, failed. You almost certainly aren't running a frontend-only build.
timed out waiting for llama-server to start
The model engine Ollama started stayed alive but made no loading progress for five minutes, so Ollama gave up. With context canceled on the end, the program that asked for the model hung up first.
model runner has unexpectedly stopped, this may be due to resource limitations or an internal error, check ollama server logs for details
The model had loaded, but the helper program running it stopped answering while Ollama was sending it your prompt. The cause is in the server log, not in this message.
Error: max retries exceeded
A piece of the model download failed six times in a row and Ollama gave up. The real cause is the network error printed after the colon.
Error: invalid file magic
Ollama opened a model file and its first four bytes weren't the GGUF signature, so it refused to read any further. The file is damaged, isn't a GGUF, or came from a format your Ollama version couldn't read yet.
registry.ollama.ai/library/deepseek-r1:8b does not support tools
Your app sent a chat request with a list of tools, and the model you picked has no template or parser that knows how to format tool calls, so Ollama refused it with a 400.
Error: 400 Bad Request: "embeddinggemma" does not support generate
The model you picked is an embedding model. It turns text into a list of numbers for search and can't write replies, so Ollama refuses a text-generation request for it.
Error: digest mismatch, file must be downloaded again
Ollama downloaded a model file, checked its fingerprint against the one the registry published, and they didn't match. It deletes the bad file so the next pull fetches it again.
Error: could not connect to ollama server, run 'ollama serve' to start it
The ollama command is only a client. It knocked on the server's address (127.0.0.1:11434 unless OLLAMA_HOST says otherwise), nothing answered, and it couldn't start one for you.
Unexpected endpoint or method.
LM Studio's server got a request for a path, or a path and method pair, that it doesn't have. It logs an error and answers with status 200 anyway, so the calling app usually fails later with a confusing message of its own.
Error rendering prompt with jinja template
Before a model sees your chat, LM Studio runs the model's prompt template (a small Jinja program shipped with the model) to turn the messages into text. That program failed, so nothing was sent to the model. The quoted text after the colon says why.
request (4390 tokens) exceeds the available context size (4096 tokens), try increasing it
The prompt your app sent is longer than the context llama-server has for one request, so it refused it without generating anything. The first number is your prompt in tokens, the second is the limit.
failed to create context with model
The model's weights loaded, but llama.cpp couldn't set up the context around them (the KV cache and scratch buffers a run needs). The reason is on the llama_init_from_model line just above, and it's usually memory.
Could not load text model:
KoboldCpp tried to load your model file, the loading code reported failure, and KoboldCpp quit. This line only names the file. The reason is a few lines higher in the console.
Sorry, your request failed. Please try again.
The request to the model failed or was rejected, and Copilot Chat showed its generic message. The useful part is the Reason line under the request IDs, which carries the real error.
Sorry, you have been rate-limited. Please wait
GitHub's Copilot service refused the request because a rate limit was hit. The Error Code line under the message names the limit, and the wait time in the first line is the server's own estimate.
Error when talking to Gemini API
A request to Gemini failed and Gemini CLI saved a report about it to a temp file. The line itself doesn't say what failed; the error text after the file path does.
A potential loop was detected
Gemini CLI stopped the model because its output looked like it was going in circles: the same tool call over and over, the same text repeating, or a second model judging the session stuck.
Something went wrong. Try again later.
Google doesn't explain this line. In most threads it appears the moment someone opens a chat on gemini.google.com, only for one Google account, while the same account works in the mobile app.
Something went wrong (9)
Google doesn't document what 9 means. In the threads we read it usually starts after a Google Docs or Drive link in a chat, and then every new prompt in that chat fails.
User location is not supported for the API use.
Google's Gemini API decided your request came from a country or territory where the free API and AI Studio aren't offered. It judges by where the traffic leaves from.
Function call is missing a thought_signature in functionCall parts.
Gemini attaches an encrypted signature to its function calls and checks that you send it back. A function call in the current turn came back without one, so the request was refused.
Waiting for extension host
The background process that runs Cursor's extensions and the agent's local tools didn't start or stopped answering, so your message never left your machine.
Unauthorized request. User is unauthorized
Cursor's servers refused the request as unauthorized. On a paid account in 2026 that's almost always because your computer's clock is wrong, which makes the sign-in token look invalid.
Taking longer than expected...
Cursor's client waited about 15 seconds without hearing back from the server. In September 2026 builds it often means the request never left your machine at all, so the window is stuck.
Planning next moves
It's the status Cursor's agent shows after you send a message and before anything comes back. When it never moves, Cursor staff say the request usually hasn't even left your machine.
This model provider doesn't serve your region.
The company behind the model (Anthropic, OpenAI or Google) doesn't offer its API where Cursor thinks you are, so Cursor won't send the request. Your account and plan are fine.
No tools, prompts, or resources
Cursor reached the MCP server, or tried to, but came away with an empty tool list. Sometimes the server really returned nothing; sometimes Cursor dropped or failed to register what it sent.
Client closed
Cursor started or dialed an MCP server and the connection ended before it got a working reply. “Client closed” is the aftermath: the actual cause is the log line just above it.
“Cursor” is damaged and can’t be opened. You should move it to the Trash.
Gatekeeper, the macOS check that runs on downloaded apps, decided this copy of Cursor couldn't be trusted. Cursor's docs say it's a macOS issue, and your download is almost always fine.
When loading the graph, the following node types were not found
The workflow you opened uses node types your ComfyUI doesn't have. Usually they come from a custom node pack you haven't installed; sometimes they're core nodes newer than your ComfyUI.
AssertionError: Torch not compiled with CUDA enabled
The copy of PyTorch that ComfyUI is using was built without Nvidia GPU support, so the first time anything asks for the graphics card, PyTorch stops with this error.
RuntimeError: mat1 and mat2 shapes cannot be multiplied
Two parts of your workflow expect data of different sizes, so PyTorch can't multiply them. In ComfyUI that almost always means a model from one family, like SD1.5, fed into another, like SDXL.
(IMPORT FAILED)
Python raised an error while ComfyUI was loading that custom node pack, so none of its nodes were registered. The startup log prints the actual error, usually a missing module, higher up.
Given groups=1, weight of size
A layer in one model received data shaped for a different kind of model. The channel counts in the message show which side expects what.
Error while deserializing header: HeaderTooLarge
The first eight bytes of a .safetensors file say how long its header is. In your file they add up to a nonsense size, which means the file is damaged or isn't a safetensors file at all.
Your access token could not be refreshed because your refresh token was already used.
Codex tried to swap its refresh token for a fresh access token, and OpenAI's login server said that refresh token had been spent already. Usually another copy of Codex on the same machine used it first.
model is not supported when using Codex with a ChatGPT account.
OpenAI's server refused the model name Codex sent for your ChatGPT plan. The model has been retired from ChatGPT sign-in, needs a higher plan, or hasn't reached your account yet.
Cline is having trouble...
The model made several mistakes in a row (broken tool calls, failed tools or API errors) and Cline stopped the task instead of letting it keep burning tokens. The task isn't lost.
There's an issue with the selected model
The API refused the model name Claude Code sent. Either the name is wrong or retired, your account or provider doesn't offer that model, or (on older versions) your login had quietly expired.
Request too large (max 32MB)
Every turn resends the whole conversation, and this time the raw request came to more than 32MB. It's a size-in-bytes limit, separate from the token context window, and images and PDFs are what usually fill it.
Context low
The conversation is within about 20,000 tokens of the context window, and nothing will compact it for you because auto-compact is off. It's a warning; the hard stop comes later as “Context limit reached”.
`temperature` and `top_p` cannot both be specified for this model. Please use only one.
The request set both sampling controls, and this Claude model only takes one of them. The API refuses rather than picking one for you.
Expected `thinking` or `redacted_thinking`, but found `text`.
Extended thinking is on, and the last assistant message you sent back starts with text instead of the thinking block Claude originally produced. Something between you and the API dropped it.
The LLM did not conform to the edit format.
The model replied with an edit Aider couldn't apply, usually a SEARCH block that doesn't match the file's current text exactly. Aider sends the error back to the model and lets it try again, up to three times.
torch.OutOfMemoryError: CUDA out of memory. Tried to allocate
PyTorch asked the GPU for memory and none was left. In vLLM that usually means the memory outside the KV cache, the pool vLLM reserves up front, turned out bigger than vLLM's startup estimate.
ValueError: To serve at least one request with the model's max seq len
The model loaded, but the memory left for the KV cache (where vLLM keeps each request's context) can't hold even one request at the model's full context length.
ValueError: No available memory for the cache blocks. Try increasing `gpu_memory_utilization`
After loading the weights and running a test pass, vLLM had nothing left inside its memory budget for the KV cache, the store that holds each request's context. With zero cache it can't serve anything, so it stops.
on startup is less than desired GPU memory utilization
vLLM wants a fixed share of the GPU's total memory (92% by default), and when it started, less than that was free. Something else, often another model server or a second vLLM, is already using the card.
vllm.v1.engine.exceptions.EngineDeadError: EngineCore encountered an issue. See stack trace (above) for the root cause.
The engine core, the process that runs the model, crashed after it had started serving. vLLM can't recover from that, so every request after it fails until the server restarts.
RuntimeError: Engine core initialization failed. See root cause above.
vLLM starts the model in a separate process, the engine core, and that process crashed before it was ready. This line only reports the crash: the reason is in a traceback printed earlier.
Token validation failed.
Suno refused to start a song because the request didn't carry a valid CAPTCHA token (a proof that a person is at the keyboard). It turns up almost always in tools that call Suno's web endpoints from a script.
Your input exceeds the context window of this model. Please adjust your input and try again.
The request is bigger than the model can take in one go: prompt, history, tool definitions and schemas all count toward the model's context window.
The server had an error while processing your request. Sorry about that!
Something failed on OpenAI's side while handling your request. Your request wasn't necessarily wrong, and most of these clear on a retry.
Rate limit reached for
Your organization or project used up its tokens-per-minute (or requests-per-minute) budget for that model, and this request would have pushed it over.
Incorrect API key provided
OpenAI received a key it doesn't accept: mistyped, deleted, from the wrong place, or a placeholder some tool sent on your behalf.
Country, region, or territory not supported
OpenAI placed the IP address your request came from in a country or territory it doesn't serve, and refused it.
Open WebUI: Server Connection Error
Open WebUI's server tried to call a model provider (Ollama, an OpenAI-style API, a speech service) and the request failed before a usable answer came back.
Network Problem
A request from the browser got no readable answer. In chat that usually meant the WebSocket link to Open WebUI was blocked; next to a connection it means the provider couldn't be reached at all.
Error: 500 Internal Server Error: llama-server process has terminated
The separate program Ollama starts to run the model died before the model finished loading. The 500 is just how the Ollama server passes that on to the client.
Error: listen tcp 127.0.0.1:11434: bind: address already in use
Another program already holds port 11434, and it's nearly always Ollama itself: the desktop app or the Linux service started a server before you typed ollama serve.
llama-server process has terminated: exit status 0xc0000409
The program Ollama uses to run the model crashed on Windows, and 0xc0000409 is the code Windows gives that kind of abrupt stop. The code is shared by many unrelated crashes, so the text after it matters more.
Error: an error was encountered while running the model: unexpected EOF
The helper program that runs the model died while it was writing your answer, so the connection Ollama was reading from ended early. EOF is short for end of file, here meaning the data stopped arriving.
Model loading was stopped due to insufficient system resources.
LM Studio estimated how much memory the model needs at your settings, compared it with what your machine has free, and refused before loading anything. Nothing crashed.
(Exit code: 18446744072635812000). Unknown error. Try a different model and/or config.
The engine process that loads the model (LM Studio's build of llama.cpp) crashed before loading finished. The long number is a Windows crash code mangled into an unsigned number, and it doesn't say which crash.
llama_model_load: error loading model
llama.cpp couldn't turn that GGUF file into a model. This line is a wrapper: the real reason is the text after the second colon, and it's usually about the file, not your hardware.
CUDA error: out of memory
llama.cpp asked the NVIDIA driver for more video memory and didn't get it. The weights, the context cache and the scratch space for each batch all have to fit in VRAM together.
Cannot access gated repo for url
The model's authors require people to request access before downloading, and the Hub didn't see an approved account behind your request: either no token was sent, or the account hasn't been let in.
429 Client Error: Too Many Requests for url
The Hub counted more requests from you (or your IP address, if you weren't logged in) than your plan allows in a five-minute window, and refused the rest until the window resets.
Model stream ended without a finish reason.
The model's reply finished streaming but failed Gemini CLI's check on what a finished reply looks like: no end marker, no text, or a broken tool call. The CLI had already tried four times.
Failed to sign in. Message:
The sign-in itself may have worked. Gemini CLI prints this when anything in the auth step throws, and the real reason is the text after Message:, which usually comes from Google's servers.
Something went wrong (13)
Google doesn't document what 13 means. In the threads we read, it usually stays with one Google account on every device and browser, and it's often tied to one kind of request, such as a photo upload or an image prompt.
Something went wrong (1185)
One conversation stops answering while new chats work. Google hasn't said what 1185 means; a Product Expert says Google is aware of it.
Something went wrong (1152)
Gemini couldn't finish the reply. Google hasn't said what 1152 means, but the reports we read cluster on long or heavy chats, and a fresh chat usually works.
Something went wrong (1099)
Google doesn't say what 1099 or 1076 means. In the community threads they show up in two ways: everyone at once during a backend incident, or one conversation that stops answering while a fresh chat still works.
Something went wrong (1097)
Google hasn't said what code 1097 means. Every report we read that named the account type came from a work or school (Google Workspace) account, and Google Workspace support told one admin in May 2026 that it was a global issue engineering was working on.
Your API key was reported as leaked. Please use another API key.
Google found your API key exposed somewhere public and has blocked it from the Gemini API for good. Your code and your quota are fine; the key is what's dead.
API key not valid. Please pass a valid API key.
Google's Gemini API endpoint didn't recognize the key it was sent. Usually that's a different key from the one you think you're using, or a key made for another Google service.
This model is currently experiencing high demand. Spikes in demand are usually temporary. Please try again later.
Google's servers for that model were too busy to take your request. It's a capacity problem on Google's side, and Google's advice is to wait and try again.
You've hit your usage limit
You've used up the included usage for one of your plan's pools this billing cycle, so Cursor won't send requests to models billed from that pool until it resets or you pay for more.
The model you chose is not available
Cursor refused to send the request to the model you picked. Recently that has mostly meant a model was broken on Cursor's side for a few hours; with your own API key it usually means the key isn't being used or can't reach that model.
Rate limited by model provider, retrying…
The company serving the model (xAI, Anthropic, OpenAI and so on) refused the request as too busy, and Cursor is retrying it for you. If you've added your own API key, it can also be that key's limit.
We're experiencing high demand right now. Please try again in a few moments.
Cursor's servers turned the request away because the model you picked (or Auto) was over capacity at that moment. It isn't your account, your network or your settings.
Errored, Not Charged
The request failed before it produced an answer, so Cursor didn't bill it and it didn't come out of your included usage. There's nothing to fix.
A background task finished, but the chat couldn't resume. Try again in a moment.
A background shell or subagent finished while the chat sat idle, and the automatic wake-up that should have continued the agent's turn failed. Cursor staff call it a known issue.
Error loading webview: Error: Could not register service worker: InvalidStateError: Failed to register a ServiceWorker: The document is in an invalid state..
An extension's panel couldn't start, because the browser-side helper every webview needs failed to register. The bug comes from VS Code, which Cursor is built on, and nobody has pinned down the cause.
Connection failed. Please try again, or contact support if the issue persists
Cursor couldn't finish a chat or agent request. Sometimes the network between you and Cursor is at fault, but often the request never left your machine because the chat itself is broken.
Reconnecting
The page in your browser lost its live connection to the ComfyUI server and is trying to get it back. When it happens mid-generation, the server process has almost always stopped, and running out of memory is the cause ComfyUI's maintainers point to first.
Prompt outputs failed validation
ComfyUI checked the workflow before running it and found at least one node with an input it can't use, so nothing ran. The line under the heading names the node, the input and what's wrong.
torch.OutOfMemoryError: Allocation on device 0 would exceed allowed memory. (out of memory)
A node asked PyTorch for more graphics card memory (VRAM) than was free, and the request failed. It's the same out-of-memory error most people know as “CUDA out of memory”, printed by the allocator ComfyUI turns on by default.
CreateProcessAsUserW failed: 5 (Access is denied.)
Codex's Windows sandbox couldn't start your command. Either the permission step before launch failed (“setup refresh had errors”, usually one folder you can't change the permissions on) or Windows refused to start the program as the sandbox user (“CreateProcessAsUserW failed”).
MCP startup failed
One of the MCP servers in your config didn't finish starting, so Codex went on without its tools. The words after “MCP startup failed:” say which step broke: launching the program, the handshake, a login, or the 30-second startup timer.
exceeded retry limit, last status: 429 Too Many Requests
The server answered 429, too many requests, and the reply didn't match any of the limits Codex knows how to explain, such as your plan's usage limit or an empty credit balance. Codex gives up on the turn straight away.
error sending request for url
Codex couldn't get a request to OpenAI's servers at all: no connection, a failed DNS lookup, a proxy in the way or a certificate it didn't trust. The line comes from Codex's HTTP library and leaves the specific reason out.
Error running remote compact task
Codex tried to shrink a long conversation by asking OpenAI's server to summarize it, and that request failed. The words after the colon are the actual failure, passed through from another error.
You've hit your session limit
Your Pro, Max, Team or Enterprise plan has used up its rolling five-hour allowance. Nothing is broken; requests are blocked until the reset time in the message.
Your account does not have access to Claude. Please login again or contact your administrator.
The API answered 403 because the saved sign-in was revoked. You only see this wording outside an interactive terminal, in claude -p, the Agent SDK or the desktop app; the terminal says “OAuth token revoked” instead.
API Error: Server is temporarily limiting requests (not your usage limit)
Anthropic's API sent back a short-lived 429 throttle that isn't tied to your plan's quota. Your usage allowance hasn't run out.
API Error: Claude's response exceeded the 32000 output token maximum. To configure this behavior, set the CLAUDE_CODE_MAX_OUTPUT_TOKENS environment variable.
One reply (thinking included) ran into the output token ceiling for the request. By the time you see this, Claude Code has already tried to continue three times and hit the ceiling each time.
Remote Control failed
The local session couldn't register with Anthropic's Remote Control server, so your phone or browser can't see it. The text after “failed” names the reason, and that's the part to act on.
Claude Code process exited with code 1
The SDK started the Claude Code program in the background and it quit with an error. Code 1 is generic, and the reason is in what that program printed before exiting.
Output blocked by content filtering policy
A filter on Anthropic's API stopped Claude's reply while it was being written. Claude Code only passes the error along, and false positives on ordinary work are common in the reports.
No conversation found with session ID
claude --resume was given a session ID and couldn't find a saved transcript for it with any messages in it, so it exited instead of starting the session.
✘ Failed to connect
Claude Code started or dialed an MCP server and didn't get a working connection. For a local server that's often the 30-second startup deadline running out.
Our intentionally broad safeguards allow us to deliver more capabilities faster, but can sometimes flag legitimate cybersecurity work.
A cybersecurity classifier on Anthropic's side declined the request. It's tuned broad on purpose, so legitimate security work and plain coding both get caught.
Could not process image
Anthropic's API refused an image in your conversation, either because it's too big or because it couldn't decode it. The image is still in your history, which is why the error can come back.
API Error: Connection lost mid-response. The response above may be incomplete.
The connection dropped while Claude's answer was streaming, after part of it had arrived. Claude Code kept the finished parts and stopped rather than risk running the same tools twice.
API Error: Claude can't help with this. Start a new session to continue.
A safety classifier on Anthropic's side declined to answer, judging the whole conversation. It isn't an account ban, and the same session will usually keep refusing.
Auto-update failed · Run claude doctor
The background updater tried to check for or install a new version and something went wrong. Your current version keeps working.
API Error: 500 Internal server error
Something failed inside Anthropic's API (or your provider's) while handling the request. Claude Code had usually retried it several times already, and your prompt, settings and plan aren't the cause.
Each `advisor_tool_result` block must have a corresponding `server_tool_use` block before it.
An advisor call and its answer got separated in the saved conversation, usually because Claude Code wrote something between them, and the API now rejects the whole history on every turn.
You’ve hit your limit for Claude messages.
Your plan's usage allowance for the current window is used up, so the app won't send more messages until the time it shows. Usage in Claude Code and the desktop app counts against the same allowance.
Your message will exceed the length limit. Try attaching fewer or smaller files or starting a new conversation.
The message you're about to send, with its attachments and the chat so far, is bigger than the model can take in, so the request was refused before Claude read any of it.
Unable to load usage limits. Please try again later.
The usage page couldn't fetch your limit windows, so it shows this card in place of the meters. It doesn't mean you've hit a limit.
This chat is too long to continue. Start a new chat, or turn off some connectors to free up space.
The chat plus everything sent with it (project instructions, files, connector tool definitions) no longer fits in the model's context window, so Claude can't take another turn in this chat.
Something went wrong while generating the response. If this issue persists please contact us through our help center at help.openai.com.
ChatGPT took your message but couldn't finish an answer. It's a catch-all: OpenAI's help center says the general version can mean a server problem or something in your own setup.
Error in message stream
The reply ChatGPT was sending you broke off before it finished. OpenAI Support calls it a generic error with more than one cause, so the message on its own doesn't say why.
Error code: Out of Memory
Your browser killed the ChatGPT tab because it used up the memory it's allowed. It's the browser's error, and on ChatGPT it mostly comes from very long conversations.
messages: text content blocks must be non-empty
Your request contains a text block whose text is an empty string, and the Claude API refuses the whole request. The word before the colon says where: messages or system.
invalid x-api-key
The Claude API couldn't match the key in your request to a working API key. It's malformed, expired, disabled or deleted, or it isn't an API key at all.
You exceeded your current quota, please check your plan and billing details.
The OpenAI API refused the request on billing grounds, not speed: no prepaid credit left, a spend limit reached, or OpenAI's own usage cap for your organization.
Error: pull model manifest: file does not exist
Ollama asked the model registry for that name and tag and the registry answered 404. Your disk is fine: the model, or that particular tag, doesn't exist where Ollama looked.
Usage limit reached for all Pro models.
Google refused a request to a Pro model and Gemini CLI read the refusal as a quota you can't retry your way past, usually the daily one. It names all Pro models because the check is on the word "pro" in the model name.
429 RESOURCE_EXHAUSTED
Your Google Cloud project went over one of its Gemini API limits: requests or tokens per minute, requests per day, or a spend limit. The line under the message names which one.
You've reached your weekly rate limit.
GitHub has capped how many model requests your account can make this week. It's a rate limit, separate from the AI credits your plan includes, and it lifts at the time printed in the message.
You’ve hit your usage limit.
Your ChatGPT plan's Codex allowance for the current window is spent. The rest of the line depends on your plan and tells you whether buying credits, an upgrade or your admin can get you going sooner.
stream disconnected before completion: Transport error: network error: error decoding response body
The model's reply was arriving over a live connection and the connection closed before OpenAI sent the end-of-response event. Codex retries it by itself, five times by default.
Selected model is at capacity. Please try a different model.
OpenAI has run short of servers for that model, at least where your requests land. OpenAI staff have said on GitHub that it has nothing to do with your rate limit or your remaining usage.
Credit balance is too low
Your request was billed to a Console organization, and that organization has no prepaid credit left. Subscription usage is a separate pot the check never looks at.
Prompt is too long
The request Claude Code tried to send was bigger than the model's context window. That counts the whole conversation plus the system prompt, every tool definition and anything you attached.
OAuth token has expired. Please obtain a new token or refresh your existing token.
Your sign-in's short-lived access token ran out and Claude Code's automatic refresh didn't replace it in time, so the API turned the request away with a 401.
Not logged in · Please run /login
Claude Code tried to send a request and found no credential the API would take. If your key came from an environment variable, you get the Invalid API key line instead.
API Error: Request timed out.
A request to Anthropic's API ran past its deadline, 10 minutes by default. Claude Code had already retried before it showed you this.
API Error: Connection dropped (ECONNRESET)
The connection to Anthropic's API was cut after it opened. Claude Code already retried it several times before showing you this, so the drop kept happening.
API Error: 400 due to tool use concurrency issues. Run /rewind to recover the conversation.
The conversation history Claude Code sent has a tool call with no matching result straight after it, so the API refused the whole request.
/limit-reset isn't available in this environment.
You typed the command into a session with no interactive terminal, such as claude -p or a program driving Claude Code. It's about where you ran it. Your account and your reset were never checked.
Invalid `signature` in `thinking` block. The block is bound to a different conversation.
The Claude API rejects a thinking block you sent back when the conversation in front of it has changed. Editing an earlier message, the system prompt or the tools list invalidates that block and every thinking block after it.
A session-limit reset isn't available right now.
The reset wasn't on offer for the limit you're at. You haven't used up the week's reset (that prints a line with a date on it), and a failed check prints its own try-again line.
llama_model_load: error loading model: unknown model architecture: 'qwen4exp'
Your llama.cpp build predates Qwen3.8-Flash-Next support. The qwen4exp architecture was added in pull request 27742, merged on Aug 27, 2026, so anything built before that has no code for the name the file declares.
llama_model_loader: tensor overrides to CPU are used with mmap enabled - consider using --load-mode none for better performance
llama-server can stop partway through loading Qwen3.8-Flash-Next on llama.cpp's HIP backend and never finish. It prints no error and ignores SIGTERM. The same file and flags load on Vulkan in 105 seconds.
rate_limit_error
A 429 rate_limit_error means your organization hit a per-minute rate limit, its usage tier's monthly spend cap, or a Claude Code workspace limit. Only the first one clears by waiting.
overloaded_error
A 529 overloaded_error means the Claude API is temporarily overloaded, from high traffic across all users rather than anything your organization did. In a streaming response it can arrive after the API has already returned 200.
PROHIBITED_CONTENT
Gemini returns PROHIBITED_CONTENT either as a block reason, when the prompt never ran, or as a finish reason, when generation stopped partway. It belongs to the built-in protections, so safety settings can't switch it off.
tool_choice: type "tool" and "any" are not supported for this model.
Claude Fable 5.1 and Claude Mythos 5.1 don't support forced tool use. Sending tool_choice with type any or type tool returns a 400 invalid_request_error, including on the token counting endpoint.
Lower-priority mode isn't available right now.
The offer was never made to your account, so there's nothing to switch on. It doesn't mean you used up the week's allowance, and you're not in a cool-off either (both of those print a different line).
Repeated 529 Overloaded errors
Anthropic's API is busy across every user and Claude Code has run out of retries. Your key, your plan and your usage limits are all uninvolved.
No match. Try a shorter fragment, or a word from the middle of the line rather than the start.
Why these pages exist
Coding agents print a lot of one-line refusals that don't say why. Some are about your plan or your billing, some mean the service is busy, and a few are about a staged rollout your account isn't part of yet. Telling those apart is most of the job, because the fix is different in each case and only some of them are worth sitting and waiting out.
Every page here names the version the strings were read from. The CLIs ship often, and the lines that come from a server (the Gemini and OpenAI APIs, for instance) can change without any release at all, so treat a mismatch as your copy being newer rather than as our typo.