DeepSeek V4 Bugs, GPT-5.5 vs DeepSeek, & Multi-Team Inference Fixes
Welcome back. Let's get into it.
DeepSeek V4 is everywhere right now — and it's having a week. A messy, complicated, fascinating week. Let's start with the good, because there is some genuinely impressive stuff here.
A user called HealthRanger was working with code originally generated by Claude Opus 4.7. The code had problems — crashes, instability. So they pointed DeepSeek V4 at it through a tool called OpenCode, asked it to analyze and fix things, and it came back with eight memory leaks identified and patched. In minutes. Total cost? About three cents. That's not a typo. Three cents to debug what Claude couldn't get right the first time. That's a compelling use case for V4 as a code review layer — even on top of other models' output.
But not everyone's having that experience.
A user named MaxBlade ran a head-to-head test. Same prompt to both DeepSeek V4 and GPT-5.5 — build a Pokémon-style battle app. DeepSeek V4 gave him a crashing game with what he described as garbage UI. Every single run. Runtime errors, broken interface. GPT-5.5, on the other hand, apparently produced something he called a masterpiece. Same prompt. Wildly different results. And he posted a demo video showing the V4 errors in real time — it's not a vibe, it's documented.
So you've got V4 fixing other models' bugs for three cents on one hand — and completely fumbling a game build on the other. The model is clearly capable, but inconsistent in ways that matter if you're building anything serious.
And there's a more technical problem brewing underneath all of this.
A developer named karminski3 pinpointed a reproducible API bug in both DeepSeek V4 Pro and Flash. Here's what's happening — the `reasoning_content` field is returning an empty string about 59% of the time on obvious tool calls. That sounds like a minor quirk until you realize what it causes downstream. When you pass that empty string back, you get HTTP 400 errors — "must pass back reasoning_content." It's crashing workflows in Claude Code terminals, in Cursor IDE, and in agents. The catch is the workaround is almost absurd: you have to transmit the empty string verbatim, exactly as returned. If you don't, it breaks. If you do — sometimes it works. This is the kind of silent failure that wastes hours before you even understand what's going wrong.
Now — some people are shipping in spite of all this.
The team behind openclaw just dropped version 2026.4.24. They've integrated DeepSeek V4 Flash as the default onboarding model, with Pro also available. The release adds voice calls to their full agent, browser coordinate clicks, and recovery features. They also specifically fixed Flash's replay and thinking issues around tool-call turns — which, given the bug karminski3 just described, is well-timed.
And speaking of fixing things fast — there's a genuinely impressive infrastructure story here.
The open-source inference framework SGLang had a garbled output bug with DeepSeek V4. The team at lmsys.org flagged it — and what followed was a cross-team sprint that finished in under 48 hours. We're talking engineers from Ant Group, Ollama, NVIDIA, Meta, FireworksAI, and DeepSeek itself all converging on a single pull request over a weekend. That kind of cross-org coordination on an open-source fix — in under two days — is not normal. It says something about how much is riding on V4 working correctly across the inference stack.
So here's where we are with DeepSeek V4: serious capabilities, real bugs, active fixes, and a community that's clearly invested in making it work. It's not a smooth launch — but it's not abandoned either. Watch this space closely over the next few weeks.
That's your AI digest for 01 May 2026.