Google launches Gemini 4 Argon globally with $2/$10 per million tokens API pricing

One-Line Summary

Frontier AI moved from labs to daily tools: Google opened Gemini 4 Argon globally with low per‑token pricing, Meta launched Muse Wonder for long-running agents, and Google Search’s AI answers now surface more direct links to the web.

Big Tech

Google launches Gemini 4 Argon globally with intro pricing

Google is making its newest top-tier model, Gemini 4 Argon, broadly available to developers and enterprises via the API in AI Studio and the Gemini platform, and to Google AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search. The introductory API price is $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%. ¹

Google positions Argon for deep, multi-step work in areas like coding, finance, legal, cybersecurity, and science, and says it leads on cross-domain benchmarks — including ranking #1 on the Vals Index, which weights sectors such as finance, coding, legal, and tax. ¹

Inside Google, Argon has supported complex engineering: the Chrome team used it to help untangle a 6‑million‑line legacy folder during a wider refactor; engineers reported more than a 35% increase in agentic code submissions with fewer rollbacks; and Google cites a 2.85× inference speedup on Gemma 4 E2B via vLLM on an A100 while maintaining quality. ¹

For everyday use, Mashable notes Pro and Ultra subscribers can access Argon in the Gemini app, with Antigravity’s agentic coding interface now in the macOS app and a new Tasks mode to keep work going across services. It also reports that, as of Oct 13, the Artificial Analysis Intelligence Index scores Argon (High) at 53, alongside detailed comparisons where Google’s own tables show Argon ahead on many tests. ²

Meta debuts Muse Wonder frontier models; top tier limited to Muse Code

Meta’s Superintelligence Lab introduced two state-of-the-art models — Muse Wonder and Muse Wonder Full — its first frontier-class entries. The standard Wonder is available through the Meta Model API, while the more capable Wonder Full is restricted to Muse Code subscribers and not served via the API. Meta emphasizes long-running assistant use cases and stronger “active memory” for retaining user context. ³

Through the API, Wonder is priced at $4 per million input tokens and $20 per million output tokens, and Meta highlights capabilities for agent tasks with “Active Memory” and “Active Focus,” a context window around one million tokens, and multi‑modal inputs. ⁴

Independent evaluations by Artificial Analysis award scores of 56 to Wonder Full and 52 to Wonder on its composite Intelligence Index, placing Wonder Full near leading models on sustained engineering and agent workflows, with differences across benchmarks depending on task type. ⁴

Industry & Biz

DeepSeek updates budget model with public MIT weights

DeepSeek released an updated AI model, V4.1‑Flash‑1017, which it says is faster, leaner, and delivers stronger benchmark performance with shorter outputs. The model is live on DeepSeek’s site, app, and developer platform. ⁵

Analysis notes the weights are available on Hugging Face under an MIT license, off‑peak pricing remains $0.15 per million input tokens and $0.60 per million output tokens, and the update boosts scores across 11 of 11 tests against the prior Flash version, including Terminal‑Bench 4.0 rising to 42.3. Results are provider‑reported and should be validated against your own workflows. ⁶

New Tools

Google Search adds more links and source context in AI answers

Google is updating AI Mode and AI Overviews so you see more inline links higher in responses, a dedicated section that surfaces official sites with context, and more prominent related-article links with thumbnails to help you click through to sources faster. ⁷

Prominent news carousels are expanding to more locations and languages globally, and Google is testing clickable creator and publisher names that link to Search profiles. Google also says usage of Preferred Sources has grown more than 7× since the start of the year. ⁷

Community Pulse

Hacker News (1466↑) — Mixed reactions: some praise Argon’s value for business automation, while others complain about hitting usage limits quickly and unclear plan boundaries. ⁸

"Well, I should mention I'm using it for business automation, not coding. But yes, it is great and cheap." — Hacker News ⁸

"I tried it out and ran out of weekly usage on a $200 plan within a day. Not impressed." — Hacker News ⁸

What This Means for You

For teams doing heavy knowledge work, Argon’s global rollout means you can test a frontier model inside tools you already use, with costs you can estimate: $2 per million input tokens, $10 per million output, and 95% off for cached prompts you reuse. That makes long, multi‑step tasks (drafting complex proposals, code reviews, legal/finance checklists) more feasible — especially if you structure prompts and reuse system instructions to benefit from caching. ¹

If your workflows rely on long‑running agents, Meta’s Wonder family focuses on persistence: “Active Memory” and “Active Focus” aim to retain context over days and keep pushing toward goals even after setbacks. Pricing at $4/$20 per million tokens and an approximately one‑million‑token context window position Wonder as a premium option to evaluate for extended assistant tasks. ⁴

For data‑sensitive or cost‑constrained projects, DeepSeek’s Flash‑1017 offers public weights under an MIT license plus very low off‑peak rates ($0.15 input/$0.60 output per million tokens). It’s designed as a budget workhorse for repetitive agent tasks, but its reported gains are vendor‑run — plan a limited, in‑house comparison before using it for critical work. ⁶

Search distribution is shifting as AI Overviews show more inline links, official‑site sections, and creator profiles. For marketing, content, and comms, this means answering “Which page will AI link first?” becomes as important as “Which blue link ranks highest?” — update your product and help pages so titles, metadata, and on‑page clarity support being surfaced as an “official” source. ⁷

Action Items

1. Try Gemini 4 Argon on a real task: In the Gemini app, give Argon a messy project folder and ask it to produce a client‑ready proposal or spec; compare the output and time-to-finish with your current assistant.
2. Estimate your Argon bill: Take one workflow (e.g., a 30‑page report), approximate input/output tokens, and apply $2/$10 per million tokens; move reusable instructions into cached prompts to capture the 95% discount.
3. Pilot a long‑running agent with Wonder: Scope a limited assistant task that spans days (e.g., weekly competitor brief), then evaluate whether “Active Memory” reduces re‑explaining and hand‑offs.
4. Harden your “official” pages for AI Overviews: Check your top queries in AI Mode, ensure there’s a single, authoritative page for key tasks, and tighten titles, descriptions, and copy so AI can cite you.
5. Request an internal review of DeepSeek Flash‑1017: Ask legal/security to review the MIT license and data path, then run a small, off‑peak evaluation against one repetitive agent task to compare speed and cost.