[BUG] Claude Code sustained ~80% ECONNRESET failure rate — March 30, 2026 specifically from Boston Area

[kumaakh]

Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

Claude Code CLI has been effectively unusable for an entire day (March 30, 2026). Approximately 80% of prompts fail with API Error: Unable to connect to API (ECONNRESET) after 3+ minutes of waiting. This persists three days after the March 26-27 infrastructure incident that was marked "resolved."

Additionally, Claude Desktop (chat mode) file uploads fail in a size-dependent manner — 28KB PNG succeeds, 70KB PNG fails with "Upload failed due to a network issue" — in the same session where chat text works perfectly. Chat text itself also shows intermittent errors on longer responses.

Debug Log Analysis — Hard Data

Ran Claude Code with -d flag for a 3-hour session (7c909456, 2026-03-30 19:00–22:01 UTC). Full log: 5,612 lines, 118 total errors.

Error breakdown

Error Type	Count	Impact
API "Connection error" (retries)	73	High — blocked all API requests
ECONNRESET (telemetry export)	27 (9 distinct events)	Low — telemetry only
MCP server auth failures	3	Unrelated
Request aborted (user cancel)	3	User-initiated

Burst pattern

Time (UTC)	Attempts	Outcome
19:00	5	User cancelled
19:01–19:02	8	Eventually recovered
19:03–19:05	9	Eventually recovered
19:11–19:17	39	Heaviest — 39 attempts in 6 minutes
22:00–22:01	12	Final burst

Error characteristics

• Error is undefined Connection error — no HTTP status code. TCP socket resets before any HTTP response is received.
• Both the API endpoint (api.anthropic.com) and the telemetry endpoint (Datadog) fail simultaneously — both sit behind the same CDN/edge layer.
• Retry backoff (attempts 1–11) sometimes recovers, sometimes exhausts all retries.
• Requests that succeed show no degradation — responses stream normally. It's binary: connect or ECONNRESET.

Telemetry ECONNRESET events (events lost per failure)

Time (UTC)	Events Lost
19:01:33	28
19:15:57	30
19:16:16	192
19:16:26	279
19:16:36	190
19:16:46	62
19:18:08	74
21:45:02	78
21:50:19	39

This Is Not a Client Network Issue

Claude Code's own debug analysis concluded "network instability on your machine/network." This is incorrect. Counter-evidence:

1. Claude Desktop chat worked fine in the same session, same network, same machine, for hours — including a lengthy troubleshooting conversation with Claude.
2. Claude Desktop file uploads are size-dependent — 28KB PNG passes, 70KB PNG fails. A local network issue would not exhibit a clean size threshold.
3. All standard network diagnostics pass — DNS resolution, TCP port 443 connectivity, TLS handshake, HTTP/2 — all working (consistent with dozens of other ECONNRESET reports: [BUG] ECONNRESET errors - Unable to connect to API despite healthy network #13657, [BUG] Claude code stopped working #35768, etc.).
4. Successful requests work perfectly — when the ~20% that connect do get through, streaming completes without issues. Local network instability would cause mid-stream failures, not binary connect/fail.
5. The burst pattern (39 attempts in 6 minutes, then stable for 2.5 hours, then another burst) is consistent with server-side load balancer rotation hitting unhealthy nodes, not local network drops.

Versions Tested — No Improvement

Version	Channel	Result
2.1.81	Stable	~80% ECONNRESET
2.1.86	Latest	~80% ECONNRESET
2.1.87	Latest (includes "fresh TCP on retry" fix)	~80% ECONNRESET

The fresh-TCP-on-retry fix in 2.1.87 did not help, which is expected if the issue is server-side connection resets rather than stale client connection pools.

Environment

• OS: Windows 11
• Install: Native (C:\Users\akhil\.local\bin\claude.exe)
• Auth: Personal Max Subscription
• Network: Home broadband (Londonderry, NH), no VPN, no proxy, no corporate firewall
• Claude Code version: Tested 2.1.81, 2.1.86, 2.1.87
• Claude Desktop: Latest version, same machine

What Should Happen?

Claude Code CLI should connect reliably. File uploads should work regardless of file size within documented limits.

Error Messages/Logs

Steps to Reproduce

1. Open Claude Code CLI on Windows (claude)
2. Send any prompt
3. Wait 3+ minutes
4. Observe: API Error: Unable to connect to API (ECONNRESET)
5. Repeat — ~80% of attempts fail, ~20% succeed

For the upload size threshold (Claude Desktop):

1. Open Claude Desktop chat (not Code mode)
2. Attach a PNG < 30KB → succeeds
3. Attach a PNG > 50KB → fails with "Upload failed due to a network issue"
4. Send text in the same conversation → works fine

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

No response

Claude Code Version

2.1.87

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

• The March 26-27, 2026 infrastructure incident (networking degradation affecting Opus 4.6 and Sonnet 4.6) was marked resolved on March 27 at 9:30 AM PT.
• A Dispatch issue was identified and resolved on March 29.
• The status page currently shows all services "Operational."
• Hacker News noted 2 days ago that Claude lost its >99% uptime in Q1 2026.
• This issue persists on March 30 — three days after the "resolved" incident.

Related Issues

• [BUG] API Error: Unable to connect to API (ECONNRESET) #39406 — ECONNRESET, Mar 26, open (tagged area:cowork)
• [BUG] Claude code stopped working #35768 — ECONNRESET since Mar 17, closed as duplicate
• [BUG] API Error: Unable to connect to API (ECONNRESET) #35094 — ECONNRESET, Mar 16, closed as duplicate
• [BUG] Unable to connect to API (ECONNRESET) v2.1.50 #27789 — ECONNRESET, Feb 23, closed as stale
• [BUG] Interactive mode fails with ECONNRESET while -p (print) mode works fine #25229 — Interactive mode ECONNRESET while -p mode works, Feb 12
• ECONNRESET kills session — needs auto-retry/reconnect #24614 — ECONNRESET kills session, needs auto-retry, Feb 10
• [BUG] ECONNRESET errors - Unable to connect to API despite healthy network #13657 — Original detailed ECONNRESET report with full network diagnostics, Dec 2025
• [CRITICAL] File upload completely broken - 404 error on all file uploads (claude.com web interface, mobile, desktop) #10462 — File upload completely broken (404), Oct 2025

https://claude.ai/share/d8960f65-da5d-4c61-a330-1cb069374592

[github-actions]

Found 3 possible duplicate issues:

1. [BUG] Claude code stopped working #35768
2. [BUG] ECONNRESET errors - Unable to connect to API despite healthy network #13657
3. [BUG] API Error: Unable to connect to API (ECONNRESET) #39406

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

[kumaakh]

C:\Users\akhil>ping -f -l 1472 api.anthropic.com

Pinging api.anthropic.com [160.79.104.10] with 1472 bytes of data:
Reply from 160.79.104.10: bytes=1472 time=26ms TTL=57
Reply from 160.79.104.10: bytes=1472 time=15ms TTL=57

Ping statistics for 160.79.104.10:
Packets: Sent = 2, Received = 2, Lost = 0 (0% loss),
Approximate round trip times in milli-seconds:
Minimum = 15ms, Maximum = 26ms, Average = 20ms
Control-C
^C
C:\Users\akhil>ping -f -l 1473 api.anthropic.com

Pinging api.anthropic.com [160.79.104.10] with 1473 bytes of data:
Packet needs to be fragmented but DF set.
Packet needs to be fragmented but DF set.
Packet needs to be fragmented but DF set.
Packet needs to be fragmented but DF set.

Ping statistics for 160.79.104.10:
Packets: Sent = 4, Received = 0, Lost = 4 (100% loss),

---

Your MTU is exactly 1500 (1472 payload + 20 IP header + 8 ICMP header = 1500). That's perfectly standard Ethernet — no tunnel, no encapsulation overhead, no ISP shenanigans.
So your network is completely clean:

MTU: standard 1500 ✓
Latency: 15-20ms to Anthropic ✓
Packet loss: 0% ✓
DNS/TLS/TCP: all working ✓
Chat streaming: works fine ✓

Your network is definitively ruled out. Add this MTU test to the bug report as the final nail — it proves there's no path fragmentation issue either.
The upload size threshold (28KB pass / 70KB fail) is therefore also server-side — likely their upload service rejecting or timing out on larger multipart bodies under load, not anything in your network path.
You've been more thorough debugging this than most SREs would be. File it and go get some rest.

[kumaakh]

Every Claude client surface — CLI interactive, CLI -p, VS Code terminal, VS Code extension panel, and Claude Desktop — exhibits ECONNRESET on March 30, 2026. Only the frequency varies. This is a server-side infrastructure issue, not a client bug.

Channel	ECONNRESET Frequency	Notes
CLI interactive mode	>80%	Effectively unusable
VS Code terminal mode (CLI inside VS Code)	>80%	Same CLI binary, same behavior
Claude Desktop chat	~20%	Intermittent, correlates with longer responses
VS Code extension panel (graphical)	Rare but present	Showed up after extended use
CLI -p (print mode)	Rare but present	Most resilient, but not immune
Claude Desktop file uploads >50KB	~100%	Size-dependent — 28KB passes, 70KB fails

[kumaakh]

Regional Finding: BOS (Boston) Cloudflare Edge Degraded — BLR (Bangalore) Unaffected

We have two offices actively using Claude Code CLI. One in New Hampshire, USA and one in Bangalore, India. The NH office has been experiencing ~80% ECONNRESET failure rate all day (March 30, 2026). The Bangalore office has had zero ECONNRESET failures in the same timeframe, using the same Claude Code version and API key authentication.

Cloudflare Edge Traces

Londonderry, NH, USA — ~80% ECONNRESET rate:

h=api.anthropic.com
ip=<redacted>
visit_scheme=https
colo=BOS
sliver=none
http=http/1.1
loc=US
tls=TLSv1.3
sni=plaintext
warp=off
gateway=off
rbi=off
kex=X25519

Bangalore, India — 0% ECONNRESET rate:

h=api.anthropic.com
ip=<redacted>
visit_scheme=https
colo=BLR
sliver=010-tier3
http=http/1.1
loc=IN
tls=TLSv1.2
sni=plaintext
warp=off
gateway=off
rbi=off
kex=X25519

Key Differences

Field	BOS (Boston, failing)	BLR (Bangalore, working)
colo	BOS	BLR
sliver	none	010-tier3
tls	TLSv1.3	TLSv1.2
ECONNRESET rate	~80%	0%

The sliver=none on BOS vs sliver=010-tier3 on BLR suggests the two edges are on different traffic routing paths or backend pools.

[kumaakh]

Even though the bot tagged this as a windows specific bug it does not seem limited to windows. It is an infra issue

[kumaakh]

The ECONNRESET error occurance is reduced by night of 31-March-2026 but it is still there at an HIGH annoyance level.

[kumaakh]

Still continue to be a very very ugly issue

[kumaakh]

This seriously burning tokens for me every other call fails half way throught conversation... this is the first time in months that I have seen 20x plan runout of usage by Thursday

[daoge88]

API Error: Unable to connect to API (ECONNRESET)
Is there a good solution to this problem? It's seriously impacting progress and wasting tokens.

[kumaakh]

Finally some hope to see someone else facing this... I have been sending emails to support@anthropic.com as well but no replies.
Some claude interactions claim that this is a NH specific problem. I amnot asking you @daoge88 to reveal your location, but curious if you are in the general area.