Week 42, 2026 — AI Briefing CW 42/2026

The short version

Google made Gemini 4 Argon generally available on 13 October. The model now runs in the API, in the Gemini app (Pro and Ultra) and in Antigravity (Ultra). Until then only selected cyber defenders had access.
Independent tests show first place or a tie, depending on the index. The Vals Index puts Argon in first place; Artificial Analysis puts it level with GPT-6 Astra and behind Claude Opus 5.5.
Meta introduced Muse Wonder, models the company describes as frontier models for AI agents. Muse Wonder Full scores 56 on Artificial Analysis; Claude Opus 5.5 scores 58.
OpenAI and Anthropic released no new model through 15 October. Their developer changelogs list only platform and operations changes.

Gemini 4 Argon: available everywhere, no consistent gap in testing

On Tuesday 13 October, Google made Gemini 4 Argon generally available. Developers and enterprises reach it through the API in AI Studio and on the Gemini platform; Google AI Ultra subscribers get it in Antigravity; AI Ultra and AI Pro subscribers get it in the Gemini app and in AI Mode in Search. According to the Gemini app release notes, the app rollout is gradual, and the Gemini API changelog lists gemini-4-argon as "generally available" from the same day. Before that, according to Google's first announcement, Argon was open only to selected cyber defenders in the Fairwind Program, which Google is now expanding by hundreds of partners.

The introductory price is $2 per million input tokens and $10 per million output tokens until 30 November, then $4 and $20. The 1-million-token limit covers input and output together; for output it was 64,000 before, according to Google.

What does the evidence say about capability? Google cites first place on the Vals Index. The Vals Index shows this as of 13 October: Argon 68.90%, Claude Sonnet 5.5 67.04% and Claude Opus 5.5 66.97%. The gap is just under two percentage points. The index measures tasks in finance, coding, law and tax, weighted by share of US gross domestic product. The Artificial Analysis Intelligence Index puts Argon at 53 in its "High" setting, level with GPT-6 Astra (53) and behind Claude Sonnet 5.5 (56) and Claude Opus 5.5 (58). The two indexes measure different things, and the scores apply to specific model settings. The sources we read do not say which setting the Gemini app uses by default.

Google's own examples concern long, autonomous work. Inside Google, agentic code submissions in Antigravity rose by more than 35% and rollbacks fell by almost a third. One team had Argon optimize the open-source Agent Development Kit overnight; session preparation time in long sessions fell by 40%, according to Google. These are vendor figures from its own operations and have not been independently checked.

Our reading: this is mainly a step in availability, not in capability. What is new is access, price and integration into products such as the app and Antigravity; the launch itself adds no new evidence about capability, only the independent tests do. They place Argon among the strongest models, but not clearly ahead of Claude Opus 5.5 or GPT-6 Astra. We do not derive a change to our AGI assessment from this.

Muse Wonder: another provider in the same performance range

On Monday 12 October, Meta introduced Muse Wonder and Muse Wonder Full, which it calls its most capable models to date and describes as frontier models for personal and professional AI agents. Muse Wonder is available through the Meta Model API and Muse Code; Muse Wonder Full only for Muse Code subscribers. Meta highlights two features: "Active Memory" (carrying relevant context across conversations) and "Active Focus" (staying with hard tasks, trying other approaches, recovering from errors). It measures them with its own tests such as FocusBench and SWE-Marathon for long software projects; the results appear in charts in the research post. Meta says a full report with safety evaluations will follow.

The independent indexes measure different tasks than Meta's tests. On Artificial Analysis, Muse Wonder Full scores 56 and Muse Wonder 52, at $4 per million input tokens and $20 per million output tokens. On the Vals Index, Muse Wonder Full reaches 64.57%, behind Argon, Sonnet 5.5, Opus 5.5 and Fable 5.1. Both Wonder models therefore sit in the upper group; these tests show no gap above it.

Also noted

Grok Voice 3.0 from xAI. On Wednesday 14 October, xAI introduced Grok Voice 3.0, a speech-to-speech model available in the API at its predecessor's price. In the Artificial Analysis measurement reproduced by xAI (as of 13 October), it scores 86.9 on the Speech-to-Speech Index, against 81.3 for its predecessor, 83.2 for GPT-Live-1 and 82.6 for Gemini 3.8 Live. On Full Duplex Bench, which tests pauses, turn-taking and interruptions, it reaches 97%. The table comes from the vendor.

Claude Managed Agents can now be traced. On 14 October, Anthropic introduced OpenTelemetry export in beta on its developer platform. It lets teams follow Messages API requests and Managed Agents sessions in their own monitoring system. Scheduled deployments of Managed Agents are now also available on Claude Platform on AWS. No new model came with it.

OpenAI announces the end of older models. The deprecations page states from 14 October that several GPT-4 models, including GPT-4o and GPT-4.1, will be removed from the API on 14 January 2027. OpenAI recommends gpt-6-luna as the replacement.

Number of the week

1.86 percentage points: the gap on the Vals Index of 13 October 2026 between Gemini 4 Argon (68.90%) and Claude Sonnet 5.5 (67.04%), in second place. Claude Opus 5.5 is third at 66.97%. A new, widely available model therefore sits only slightly ahead of the existing top models here. Whether the gap holds on other tasks or settings is not shown by the data.

The short version

Gemini 4 Argon: available everywhere, no consistent gap in testing

Muse Wonder: another provider in the same performance range

Also noted

Number of the week

Recommended reading