OpenAI and DeepSeek Ship Coding Agents Hours Apart

GPT-5.5 and DeepSeek-V4 landed within 24 hours, both targeting agentic coding. One's locked inside Codex, one's open-source at 1M context — and teams building with AI now have to pick a side.

OpenAI and DeepSeek shipped dueling coding agents within 24 hours on April 23-24, 2026, and both are aimed at the same target: models that carry multi-step engineering work through to completion. The split between them — one closed behind a subscription, one open-source at 1M context — is about to force a real choice for teams building with AI.

OpenAI announced GPT-5.5 on April 23, 2026, rolling it out to Plus, Pro, Business, and Enterprise tiers in ChatGPT and Codex. OpenAI claims GPT-5.5 matches its predecessor's per-token latency while performing better on nearly every evaluation they ran, with specific gains called out in agentic coding and computer use. A GPT-5.5 Pro variant targets demanding tasks on the same tiers.

Hours later, DeepSeek released V4 Preview as open-source, with a 1.6T-parameter Pro model (49B active) and a 284B Flash model (13B active). Both support 1M context by default — DeepSeek calls this the new standard — and both plug into Claude Code, OpenClaw, and OpenCode via drop-in OpenAI or Anthropic-compatible APIs. DeepSeek claims SOTA among open models on agentic coding benchmarks, trailing only Gemini-3.1-Pro on world knowledge as of April 2026.

The contrast is the story. GPT-5.5 is a platform play — it runs where OpenAI runs it. DeepSeek-V4 runs anywhere and keeps existing tooling intact; a team already on Claude Code can swap the underlying model without re-plumbing the agent loop. That's a new pressure on closed-model pricing that didn't exist a month ago.

The capability ceiling is also rising faster than the marketing. A developer thread documented that Codex generated its own launch video — converting the official After Effects motion to Remotion and React components, then using computer use to map out clicks in a copy of the desktop app's source. The agent built the assets to announce itself.

For builders, the practical question is whether 1M context and open weights beat the integration polish of Codex inside ChatGPT. For buyers, DeepSeek-V4 just reset the negotiating floor on any closed-model contract. For creators, the tools are now capable enough to produce the collateral that sells them.

The bottom line: If you're locked into a single coding-agent vendor as of April 2026, this week is the signal to run a bake-off before your next renewal.

Sources