# GPT-5.5, DeepSeek V4 preview and Kimi K2.6 in one week

> OpenAI ships GPT-5.5, DeepSeek opens a 1.6T-parameter V4 preview, and Moonshot's Kimi K2.6 targets 12-hour coding runs.

Published 13 Oct 2026. Coverage: 20 Apr 2026 to 26 Apr 2026. By Ali Asif.

## Items

1. **GPT-5.5** (OpenAI, Frontier): A 1,050,000-token context, 128K output, reasoning levels up to xhigh, and $5 in and $30 out per million tokens. [Primary source](https://developers.openai.com/api/docs/models/gpt-5.5)
2. **DeepSeek-V4 Preview** (DeepSeek, Open weights): Open weights for V4-Pro (1.6T total, 49B active) and V4-Flash (284B total, 13B active), both with a 1M-token default context. [Primary source](https://api-docs.deepseek.com/news/news260424/)
3. **Kimi K2.6** (Moonshot AI, Open weights): A 1T-parameter mixture-of-experts model with 32B active, built for long coding runs and agent swarms of up to 300 parallel sub-agents. [Primary source](https://www.kimi.com/blog/kimi-k2-6)
4. **GPT Image 2** (OpenAI, Image, video, world): OpenAI's new image generation and editing model, released April 21, also surfaced in ChatGPT as Images 2.0. [Primary source](https://developers.openai.com/api/docs/models/gpt-image-2)
5. **MiMo-V2.5-Pro** (Xiaomi, Open weights): Xiaomi's V2.5 and V2.5 Pro models arrived the same day as GPT-5.5, with a model page on Hugging Face. [Primary source](https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro)
6. **Memory for Claude Managed Agents** (Anthropic, Agents and coding): Public beta on April 23 of memory that lets hosted agents learn across sessions. [Primary source](https://claude.com/blog/claude-managed-agents-memory)
7. **OpenGame: Open Agentic Coding for Games** (arXiv, Agents and coding): An open system where agents write game code end to end. [Primary source](https://arxiv.org/abs/2604.18394)
8. **LLaDA2.0-Uni** (arXiv, Science): A diffusion language model that handles multimodal understanding and generation in one model. [Primary source](https://arxiv.org/abs/2604.20796)

Three launches this week, and the price tags tell the story as much as the scores.

## GPT-5.5

OpenAI released GPT-5.5 on April 23, with an API snapshot dated the same day and a Pro variant for harder problems. The model page lists a context of 1,050,000 tokens, up to 128,000 output tokens, a December 2025 knowledge cutoff, and reasoning levels from none through xhigh, with medium as the default. It takes text and images and returns text. Pricing is $5 per million input tokens, $0.50 cached and $30 output, and prompts over 272K tokens cost double on input and 1.5 times on output for the whole session.

That long-prompt rule is the line I would put in front of anyone planning to stuff a full tender pack into one call. The headline number is cheap enough. The surcharge makes a big file expensive.

## DeepSeek V4 and Kimi K2.6

DeepSeek published a V4 preview on April 24. V4-Pro has 1.6T total parameters with 49B active, and V4-Flash has 284B total with 13B active. A one-million-token context is now the default across DeepSeek's services, both models have thinking and non-thinking modes, and the API accepts OpenAI and Anthropic request formats. The old deepseek-chat and deepseek-reasoner names retire after July 24. DeepSeek calls Pro the open-source leader on agentic coding. That is the vendor's claim.

Moonshot's Kimi K2.6 came on April 20, a 1T-parameter mixture-of-experts model with 32B active. The announcement stresses long coding runs, with over 12 hours of continuous execution and 4,000 or more tool calls, and swarms of up to 300 parallel sub-agents. The license is a modified MIT.

## Everything else

OpenAI also released GPT Image 2 on April 21, which showed up in ChatGPT as Images 2.0. Xiaomi put MiMo V2.5 and V2.5 Pro out on April 23. Anthropic moved memory for Managed Agents into public beta the same day. Two papers caught my eye on the research side: OpenGame, on agents that write games, and LLaDA2.0-Uni, a diffusion language model aimed at both understanding and generating multimodal content.

What I'd watch next: whether anyone runs V4-Flash on a real set of bid documents and reports cost per page, because that number decides whether it replaces my cheap tier.

## My take

GPT-5.5 at $5 in and $30 out is not cheap, and DeepSeek V4 and Kimi K2.6 are the reminder that I do not have to pay that for every step. I route by task: the expensive model decides, a cheaper one grinds through the 400-line supplier sheet. This week gave me two new open candidates for the grinding.

## Sources

- https://developers.openai.com/api/docs/models/gpt-5.5
- https://api-docs.deepseek.com/news/news260424/
- https://www.kimi.com/blog/kimi-k2-6
- https://developers.openai.com/api/docs/models/gpt-image-2
- https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro
- https://claude.com/blog/claude-managed-agents-memory
- https://arxiv.org/abs/2604.18394
- https://arxiv.org/abs/2604.20796

Canonical: https://heyaliasif.com/ai/2026-04-20-week-in-ai