Three Stealth Models Leaked and Claude's Enzyme Discovery: The Agent Weekly Model Briefing

Conclusion: What to Try Right Now Is Space Bunny, and What to Watch Is Sonnet 5.5

This week there are three pieces of news that directly touch agent work. The OpenRouter stealth model Space Bunny Alpha (1M context, free), signs of a leak of OpenAI's GPT-6 family Astra Minor, and a leak of Claude Sonnet 5.5 (rumored 2M context). On top of that comes the news that 950 Claude agents independently discovered the CRISPR-like enzyme system ART. (Web verification included; unverified items are marked with sources)

1. The MiniMax M3.1 Suspicion: Space Bunny Alpha (Available)

Space Bunny Alpha has been registered as an OpenRouter stealth model. It has a 1M token context, up to 520,000 output tokens, text/image/video input, and is currently free. Test code for M3.1 was caught early in MiniMax's open-source repository (1M context + low/medium/high reasoning effort settings), and analyses say the tokenizer and error fingerprints match the MiniMax family, so it is presumed to be a stealth test version of M3.1. MiniMax, however, has not confirmed it.

Agent takeaway: free, 1M-context, and multimodal alone makes it worth testing for whole-repository ingestion and large-scale log analysis.

Competitive landscape: for the open-weight camp, 1M context has become the default. Space Bunny's (presumed M3.1's) direct rivals are below.

If M3.1 launches officially, it will be a head-to-head matchup with Kimi K3.1 (routing settings caught, rumored to be imminent). With both offering 1M context and both in the open camp, the cost-performance crown for long-context agent work is at stake.

2. Signs of an OpenAI Astra Minor Leak (Deleted)

This is a report that a hidden model, Astra Minor, was spotted in OpenAI's help center, and the related content has now been deleted. Another report says gpt-6-astra-minor appeared alongside GPT-6 Sol and Luna in the Azure OpenAI model registry. It is presumed to be the small, high-efficiency position in the GPT-6 product family, but there is no official confirmation.

Competitive landscape: the full GPT-6 lineup is composed of the flagship Astra (launched about 3 weeks ago) + the mid-range Sol + the low-cost Luna. Sol and Luna launched on 2026-09-22, and their official pricing is half that of GPT-5.6.

3. The Claude Sonnet 5.5 Stealth Test Rumor (Includes a Correction)

Following Opus 5.5, this is a leak that Sonnet 5.5 is in the stealth test stage. A correction worth noting: the current Sonnet 5 (launched 2026-06-30) already has 1M context and adaptive thinking, so the core of the 5.5 leak is the 2M token context rumor + flagship Claude Fable 5-class performance + Sonnet-class pricing. If true, it is a cost-performance throne-changing level event.

Competitive landscape: in the same week, GPT-6 Sol ($2/$10) launched a half-price offensive, and Kimi K3 ($3/$15, cached $0.30) had laid down open-weight cost-performance. If Sonnet 5.5 arrives with 2M context + Sonnet pricing, the price and context benchmarks of the mid-range market all get shaken. For agent practitioners, now is not the time to sign long-term contracts for paid APIs.

4. The Biological Discovery ART by 950 Claude Agents (Officially Confirmed)

Anthropic's official announcement (2026-09-23, per the company's report): as an early achievement of its newly established life sciences lab, about 950 Claude agents explored 21 hours' worth — 210 million tokens — of bacteriophage DNA and discovered an unknown enzyme system, ART (array-associated reverse transcriptases). It is a CRISPR-like repeat structure, and after the AI pinpointed what human researchers had missed, it was experimentally verified at the company's Bay Area lab. The actual biological function of ART, however, is still unconfirmed.

Agent takeaway: a 950-agent, 21-hour autonomous exploration campaign leading to a real scientific discovery is a first-of-its-kind case. It is worth studying as a reference for designing swarm-style exploration workflows.

5. The Claude Code Change Rumor: Plan Mode → effort control

There is word that Claude Code is considering abolishing the dedicated Plan Mode and replacing it with effort control via Shift + Tab. Desktop/web-only project features and cloud session rollout were also reported alongside it.

6. Other Briefs (Based on the Original Video Summary, Not Independently Verified)

• Kimi K3.1: 3.1 added to agent routing settings, hinting at an imminent release
• Xiaomi Mimo v3.0: improves 1M-context efficiency with a KV cache sharing/reuse architecture
• ChatGPT Voice: email/calendar/Slack integration + voice handling for creating documents, presentations, and websites
• Gemini 3.8 Flash TS / Flashlight TTS: natural voice generation
• Meta ultra-light VR glasses: AI control, a virtual workspace on the level of movie editing

Sources

• OpenAI, Introducing GPT-6 Sol and Luna (2026-09-22)
• TechCrunch, OpenAI launches GPT-6 Sol and Luna (half-price announcement)
• codersera, GPT-6 Sol and Luna Complete Guide (Sol $2/$10, Luna $0.10/$0.50)
• Moonshot AI / codersera / llm-stats, Kimi K3 (2026-07-16, 2.8T, 1M, $3/$15)
• MiniMax, MiniMax M3 (2026-06-01, open-weight, 1M, multimodal)
• codersera, GLM-5.2 vs MiniMax M3 (open-weight coding comparison)
• Artificial Analysis, MiniMax-M3 vs M2.5 / Intelligence Index (M3 55, K2.6 54)
• OpenRouter, Space Bunny Alpha (stealth, 1M context, free)
• Lookonchain, OpenRouter launches anonymous model Space Bunny (MiniMax-M3.1 code caught)
• BlockBeats, the Space Bunny Alpha and MiniMax M3.1 mobilization theory
• Anthropic, Claude discovers a novel enzyme system (2026-09-23)
• Unite.ai / Tech Yahoo, reports on the ART enzyme system (950 agents, 21 hours, 210 million tokens)
• Claude Platform Docs, Claude Sonnet 5 (1M context, adaptive thinking)
• llm-stats, Claude Sonnet 5 ($2/$10, launched 2026-06-30)
• thewincentral / aibase, Claude Sonnet 5.5 leak (2M context rumor)