# GPT-5.5 Arrives as DeepSeek Opens V4 and Agents Get More Practical *By AI High Signal Digest β€’ April 24, 2026* OpenAI shipped GPT-5.5 with an efficiency-focused story, while DeepSeek countered with open-source V4 models built for 1M-token context and aggressive pricing. The brief also covers new agent-infrastructure research, orchestration tools, and a U.S. warning on industrial-scale AI distillation. ## Top Stories *Why it matters:* The day’s biggest signal was a two-front competition: OpenAI pushed a more efficient frontier model into products, while DeepSeek answered with a large open-weight release built around million-token context. - **OpenAI launched GPT-5.5 in ChatGPT and Codex, with API access coming soon.** OpenAI says the model is built for real work and agents, matches GPT-5.4 per-token latency, and uses significantly fewer tokens on Codex tasks [^1][^2]. API pricing is **$5 / 1M input** and **$30 / 1M output** with a **1M context window** [^3]. Official results highlighted **82.7%** on Terminal-Bench 2.0 and **81.8%** on CyberGym [^4][^5]. Artificial Analysis says GPT-5.5 now leads its Intelligence Index by 3 points, though its AA-Omniscience hallucination rate remains high at **86%** [^6]. - **DeepSeek open-sourced V4 Pro and Flash.** The new family ships with **V4-Pro** at **1.6T total / 49B active** and **V4-Flash** at **284B / 13B active**, both with **1M context** and live API/web availability [^7]. DeepSeek says V4 uses token-wise compression plus DeepSeek Sparse Attention to cut long-context compute and memory costs [^8]. Pricing is aggressive: **$1.74/$3.48** for Pro and **$0.14/$0.28** for Flash per 1M input/output tokens [^9]. vLLM says the V4 design reduces per-layer KV state at 1M context by about **8.7x**, and Artificial Analysis already ranks V4 Pro as the top open-weights model on GDPval-AA [^10][^9]. ## Research & Innovation *Why it matters:* The most useful research today focused on making training and agent systems more scalable, while clarifying where multi-agent hype breaks down. - **Google DeepMind’s Decoupled DiLoCo** enables training across multiple data centers without stalling on failures; Google says it trained a **12B Gemma** model across **four U.S. regions** and mixed **TPU v6e/v5p** hardware without performance loss [^11][^12][^13][^14]. - **Neural Garbage Collection (NGC)** trains models to manage their own KV cache with RL/GRPO, aiming to stabilize memory for long-horizon reasoning, agents, and tool use [^15][^16]. - A new paper on **diversity collapse** found that multi-agent LLM systems can converge toward near-identical outputs over time because shared context and mutual feedback homogenize the group [^17]. ## Products & Launches *Why it matters:* New launches are less about one-off demos and more about dependable autonomy, memory, and orchestration. - **Codex** added broader browser/computer support plus **auto-review**, letting the agent keep moving through tests, builds, files, and UI tasks while a separate checker reviews higher-risk actions [^18][^19][^20][^21]. - **Sakana Fugu** entered beta as a multi-agent orchestration system; Sakana says it has hit SOTA on **SWE-Pro, GPQA-D, and ALE-Bench** and ships as an **OpenAI-compatible API** with **Mini** and **Ultra** modes [^22]. - **Claude Managed Agents Memory** is now in public beta, giving Anthropic-managed agents a memory layer that learns from prior sessions [^23]. ## Industry Moves *Why it matters:* Competition is shifting toward distribution, enterprise rollout, and on-device deployment. - OpenAI and **NVIDIA** piloted a company-wide **Codex** rollout, and OpenAI says it is now offering whole-company deployments to other enterprises [^24][^25]. - **Liquid AI** signed a multi-year **Mercedes-Benz** partnership to bring embedded speech, language understanding, and reasoning into future MBUX systems [^26][^27][^26]. - **Glif V2** launched alongside a **$17.5M seed** led by **a16z** and **USV**, positioning itself as a creative super agent for ads, films, voiceovers, and more [^28]. ## Policy & Regulation *Why it matters:* The clearest government action today was around model theft and distillation. - A U.S. memo said foreign entities, primarily in China, are running **industrial-scale distillation campaigns** against American AI and said the government will act to protect domestic innovation [^29]. ## Quick Takes *Why it matters:* These are smaller updates, but each points to where competition is moving next.* - **Kimi K2.6** is now the **#1 open model** in both **Vision Arena** and **Document Arena** [^30][^31]. - **Qwen3.6-27B** can run locally on **18GB RAM** and beats the much larger **Qwen3.5-397B-A17B** on major coding benchmarks [^32]. - **Kling 3.0** is live with **native 4K video** generation, without upscaling [^33]. - Anthropic says recent **Claude Code** quality issues were traced to three problems, fixed in **v2.1.116+**, with usage limits reset for subscribers [^34]. --- ### Sources [^1]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2047376561205325845) [^2]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2047376564309115134) [^3]: [𝕏 post by @sama](https://x.com/sama/status/2047379036419014928) [^4]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2047377098483155317) [^5]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2047377269933637774) [^6]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2047378419282034920) [^7]: [𝕏 post by @deepseek_ai](https://x.com/deepseek_ai/status/2047516922263285776) [^8]: [𝕏 post by @deepseek_ai](https://x.com/deepseek_ai/status/2047516936289017964) [^9]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2047547434809880611) [^10]: [𝕏 post by @vllm_project](https://x.com/vllm_project/status/2047520252851105796) [^11]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2047330981145669790) [^12]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2047330989936894350) [^13]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2047330992713589009) [^14]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2047330995309826363) [^15]: [𝕏 post by @cwolferesearch](https://x.com/cwolferesearch/status/2047476297031631102) [^16]: [𝕏 post by @cwolferesearch](https://x.com/cwolferesearch/status/2047476574929420478) [^17]: [𝕏 post by @dair_ai](https://x.com/dair_ai/status/2047326894992081296) [^18]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2047381283358355706) [^19]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2047381465735070050) [^20]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2047436655863464011) [^21]: [𝕏 post by @gdb](https://x.com/gdb/status/2047489218998628780) [^22]: [𝕏 post by @SakanaAILabs](https://x.com/SakanaAILabs/status/2047479445209145785) [^23]: [𝕏 post by @claudeai](https://x.com/claudeai/status/2047421844311949513) [^24]: [𝕏 post by @sama](https://x.com/sama/status/2047395562501411058) [^25]: [𝕏 post by @gdb](https://x.com/gdb/status/2047398726856245623) [^26]: [𝕏 post by @liquidai](https://x.com/liquidai/status/2047315058775220458) [^27]: [𝕏 post by @maximelabonne](https://x.com/maximelabonne/status/2047319469869973577) [^28]: [𝕏 post by @fabianstelzer](https://x.com/fabianstelzer/status/2047359946702880920) [^29]: [𝕏 post by @mkratsios47](https://x.com/mkratsios47/status/2047316220785905948) [^30]: [𝕏 post by @arena](https://x.com/arena/status/2047539004816732257) [^31]: [𝕏 post by @Kimi_Moonshot](https://x.com/Kimi_Moonshot/status/2047543386153140231) [^32]: [𝕏 post by @UnslothAI](https://x.com/UnslothAI/status/2046959757299487029) [^33]: [𝕏 post by @openart_ai](https://x.com/openart_ai/status/2047388742671872107) [^34]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2047371123185287223)