Gemini 4 Argon Goes Global as Carbon Joins a Crowded AI Model Race

A month of frontier releases

In its October roundup, Google made model releases the main story. Gemini 4 Argon is a frontier reasoning model, and it reached global general availability. About two weeks later, Google announced Gemini 4 Carbon, which it calls its strongest model so far. Carbon is aimed at long-horizon engineering work and cybersecurity defense.11 Both launches landed in a busy autumn. Anthropic finished its Claude 5.5 family, OpenAI filled out its GPT-6 lineup, and open-weight labs had one of their most active months in recent memory.32337

The main takeaway from October is that the leading labs now ship in two stages. A capable model goes out broadly, while a stronger or riskier model is held behind a trusted-access program. Google, Anthropic and OpenAI all follow some version of this pattern, and the gap between public and gated models is now part of how each company positions itself.

Google: Argon ships, Carbon waits

Gemini 4 Argon became generally available on October 13. It is the first model in the Gemini 4 family, which Google introduced in late September.1517 Google says Argon is built for complex coding, financial analysis and legal workflows. It is available through Google AI Studio, the Gemini API, Google Cloud, Search and the Gemini app, and a separate robotics variant handles reasoning about the physical world.1120 Consumers get it through Google AI Ultra and Pro subscriptions, including AI Mode in Search.13 One account says Argon reached AI Mode in 19 countries by October 19.32

Enterprise rollout moved quickly. Gemini Enterprise made Argon generally available in its web and mobile apps on October 13 and in Workflow Builder the next day. Argon also replaced Gemini 3.8 Flash as the default model for new AlphaEvolve experiments on October 26, which changes token costs for customers who use the default.1213 Google also tidied its older lineup. Gemini 3.7 Flash and 3.5 Flash were deprecated, with requests automatically routed to newer Flash versions. Nano Banana 2.1, a faster image model, reached general availability on October 6.14

Carbon was announced on October 28 and has a narrower audience for now. It is in pre-release testing with the U.S. government and a select group of cyber defenders through what Google calls the Fairwind Program. Google says it will offer Carbon to developers, businesses and consumers once its safety guardrails are refined.1115 The coverage agrees on the facts but describes them differently. Google presents the two-week gap between Argon and Carbon as proof of its pace.11 One independent tracker calls Argon the month's gated flagship.32 Our reading: "gated" fits Carbon far better than Argon, which is now widely available.

The model news also connects to a larger product push. On October 8, Google introduced the Gemini agent, a single tool meant to plan and carry out work across Workspace, Slack and Microsoft 365. Each agent gets its own Workspace account and email address.1618 Notably, the agent's model picker starts with Anthropic's Claude models as third-party options, and Google plans to add open-source models later.16 Reports disagree on which model powers the agent. One analysis says it launched on the Gemini 3 family. That is plausible given the timing, since it launched five days before Argon.19 Separately, Gemini CLI added a notice that it is not yet optimized for Gemini 4.13

Anthropic: the Claude 5.5 family is complete

Anthropic finished its Claude 5.5 rollout in about two weeks. Opus 5.5 launched September 22, Sonnet 5.5 on September 28, and Haiku 5.5 on October 7.13 Anthropic says Opus 5.5 performs at about the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.2 Its price is $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5.8 Sonnet 5.5 is described as more than 30% faster and up to 30% cheaper than Sonnet 5.1 Anthropic calls Haiku 5.5 its cheapest and fastest small model ever, priced well below Haiku 4.5.5 One timeline lists a 1-million-token context window and up to 128K output tokens for Haiku 5.5.6

The safety framing stands out. Opus 5.5 is Anthropic's first release since it publicly called for pacing frontier development. Because the model's abilities in biology and cybersecurity are comparable to the restricted Claude Mythos 5.1, Anthropic is shipping it with safeguards similar to those on Fable 5.1.28 Mythos 5.1 is still limited to approved trusted-access programs.4 Google is also spreading Claude more widely: Opus 5.5 and Sonnet 5.5 arrived in Gemini Enterprise's developer tools on October 8.12

OpenAI: GPT-6 reaches everyone, and pricing tiers spread out

OpenAI launched GPT-6 Astra on September 3 and calls it the most intelligent and aligned model in the world.2127 By one account, Astra meets OpenAI's "Critical" cybersecurity threshold. Its more permissive capabilities are available only to vetted organizations through a program called Daybreak.27 That makes OpenAI's setup closely resemble Google's Fairwind and Anthropic's Mythos programs.

OpenAI then spread GPT-6 down its product line. GPT-6 Sol and Luna arrived on September 22 at half the API price of the 5.6 series. Anthropic shipped Opus 5.5 just 90 minutes earlier that day.30 At DevDay on September 29, OpenAI introduced GPT-6.1 Sol, which it says comes close to Astra on agentic coding and computer use at one-fifth of Astra's token prices. Its standard rate is $2 per million input tokens and $10 per million output tokens.2229 TechCrunch noted that the expected GPT-6.1 Astra did not launch.29 On October 7, GPT-6 with "Intelligent UI" began reaching ChatGPT's general chat. The feature lets ChatGPT answer with interactive tools and visuals, and OpenAI says ChatGPT has more than 1.2 billion weekly users.23 GPT-6.1 Terra followed in late October at $1 and $5 per million tokens. OpenAI says it matches Astra on a computer-use benchmark at one-tenth of Astra's per-token price.2432

OpenAI's own pages show some date inconsistencies, so dates from independent outlets are the more reliable guide. The overall direction is clear, though. OpenAI is competing on cost per task as much as on raw capability, and its Sol pricing is the same as Anthropic's Sonnet 5.5 pricing.622

Open weights: the busiest stretch in a long time

Open-weight releases may be October's most important story. On October 15, Meta released Muse Spark Open, an 816-billion-parameter mixture-of-experts model with 42 billion active parameters and a 1-million-token context window. It is the first Muse model with downloadable weights.3637 It uses Meta's own Muse Model License rather than a standard open-source license, and running it locally reportedly takes about 1 TB of memory.3537 DeepSeek released an update, V4.1 Flash 1017, on October 17 under the MIT license. API pricing starts at about $0.10 per million input tokens.3537 MiniMax launched M3.1 Flash on October 19 and published its weights on October 23.37

The largest jump came from Moonshot. Kimi K3.1 Pro, released October 23, reportedly scored 55 on the Artificial Analysis Intelligence Index. That puts it three points behind Claude Opus 5.5, a gap that stood at eleven points three weeks earlier.32 Moonshot promised the weights by November 4. One ranking noted they had not appeared as of its update.3239 Mistral Large 4 also illustrates the gap between announcing weights and releasing them. Its Hugging Face page listed the weights as upcoming as late as October 27.33 A later ranking cites a Mistral open-weights release dated October 29.39 Smaller models matter too. Liquid AI's LFM3 family fits roughly 1 GB at 4-bit precision, and Aleph Alpha's Kolibri-1 targets sovereign German-English deployments.3433

What it means

Taken together, October shows the closed frontier pulling apart into public and restricted tiers. Gemini 4 Carbon, Mythos 5.1 and Astra's cyber capabilities all sit behind vetted-access programs. Meanwhile, the best open-weight models are closing in on the closed models that the public can actually use.1142732 For buyers, these three things now matter more than any single benchmark lead: price per task, licensing terms, and whether a model can be self-hosted at all.