# GPT-Realtime-2, Anthropic's SpaceX compute deal, MolmoAct2

> OpenAI ships a reasoning voice model, Anthropic doubles Claude Code limits after a compute deal, and Ai2 opens a robot action model.

Published 13 Oct 2026. Coverage: 4 May 2026 to 10 May 2026. By Ali Asif.

## Items

1. **GPT-Realtime-2** (OpenAI, Frontier): OpenAI's most capable speech-to-speech model, with configurable reasoning, tool use, a 128K context and image input, released May 7. [Primary source](https://developers.openai.com/api/docs/models/gpt-realtime-2)
2. **Multiagent orchestration and Outcomes for Managed Agents** (Anthropic, Agents and coding): May 6 brought multiagent orchestration and Outcomes into public beta, plus Dreams as a research preview. [Primary source](https://claude.com/blog/new-in-claude-managed-agents)
3. **Codex Chrome extension** (OpenAI, Agents and coding): Codex can now work inside the user's own Chrome, launched May 7. [Primary source](https://developers.openai.com/codex/chrome-extension)
4. **MolmoAct 2** (Allen Institute for AI, Robotics): A fully open robot action model with weights, code and data, reporting a 20.6% average on its MolmoBot household benchmark against 10.3% for pi0.5. [Primary source](https://allenai.org/blog/molmoact2)
5. **Higher usage limits and a SpaceX compute deal** (Anthropic, Policy and business): Claude Code five-hour limits doubled for paid plans after Anthropic agreed to use all of SpaceX's Colossus 1 capacity, over 300 megawatts. [Primary source](https://www.anthropic.com/news/higher-limits-spacex)
6. **ARIS: Autonomous Research via Adversarial Multi-Agent Collaboration** (arXiv, Science): A system where several agents argue with each other to carry out research on their own. [Primary source](https://arxiv.org/abs/2605.03042)

A quieter week for new models, and a busy one for the plumbing around them.

## GPT-Realtime-2

OpenAI shipped GPT-Realtime-2 on May 7, alongside realtime translate and whisper variants. The model page calls it the most capable realtime voice model so far: speech-to-speech, configurable reasoning effort, function calling and prompt caching, with a 128,000-token context and 32,000 max output tokens. It accepts text, audio and image input and answers in text or audio. Text costs $4 in and $24 out per million tokens, and audio runs $32 in and $64 out. The text output price is up from $16 on the previous Realtime model.

The thing I noticed is the reasoning dial. Voice agents used to trade intelligence for latency. Here OpenAI lets you pick, and warns that higher effort adds delay and output tokens.

## Anthropic's plumbing

On May 6 Anthropic doubled Claude Code's five-hour rate limits for Pro, Max, Team and seat-based Enterprise plans and removed the peak-hours reduction. The reason was capacity: Anthropic agreed to use all of SpaceX's Colossus 1 data center, which adds more than 300 megawatts, over 220,000 NVIDIA GPUs, within the month. The same post lists commitments with Amazon, Google and Broadcom, Microsoft and NVIDIA, and Fluidstack.

The same day, Managed Agents gained multiagent orchestration and Outcomes in public beta, and Dreams arrived as a research preview. Over at OpenAI, the Codex Chrome extension launched on May 7, putting the agent inside the browser the user already has open.

## MolmoAct 2

Ai2 released MolmoAct 2 on May 5, an action reasoning model for robots with weights, code and data all open. On its MolmoBot household benchmark Ai2 reports a 20.6% average against 10.3% for pi0.5. The caveat in the blog is honest: out of the box it works on the robot setups it was trained on, which are SO-100, a bimanual YAM and Franka arms.

## Papers

ARIS, a multi-agent system where the agents challenge each other to run research, ranked second on Hugging Face's weekly list behind MolmoAct 2.

What I'd watch next: whether the doubled Claude Code limits hold once the Colossus capacity is fully online.

## My take

Realtime voice that can call tools is the first thing in this category I can picture on a supplier call, in Arabic or English. I would still want a human to approve the first message before it speaks for the company. But at $32 per million audio input tokens the price is no longer the thing stopping me.

## Sources

- https://developers.openai.com/api/docs/models/gpt-realtime-2
- https://claude.com/blog/new-in-claude-managed-agents
- https://developers.openai.com/codex/chrome-extension
- https://allenai.org/blog/molmoact2
- https://www.anthropic.com/news/higher-limits-spacex
- https://arxiv.org/abs/2605.03042

Canonical: https://heyaliasif.com/ai/2026-05-04-week-in-ai