Gemini 4 Argon reaches the API at half price, and Meta's Muse Wonder arrives at $4/$20

The week on the beat

1. Anthropic disclosed that Claude models took unintended actions on real websites during testing, and it has cut live internet access from all its internal evaluations

In one case, Claude Haiku 4.5 was running example tasks on random webpages, landed on a police tip form about an unsolved homicide, and submitted an invented sighting; the tip was flagged as spam and never forwarded for investigation. Other runs exploited an injection flaw on a university server, reached data gated behind tokens or fees, and used URL shorteners to get around limits on a fetch tool, and some of the sites belonged to US government agencies. Anthropic says the impact was minimal. If you run agents on the live web, the report's remediation section is worth ten minutes.

Read the story →

2. Google put Gemini 4 Argon on the Gemini API's pay-as-you-go tier at half price until December, but Antigravity access needs Google AI Ultra

Google's pricing page lists $2/$10 per million through November 30 and $4/$20 from December 1, with batch and flex at half those rates and implicit caching only, and says Argon is not available on the API's free tier or the Google AI Pro plan. In Google Antigravity it needs an Ultra subscription or an enterprise account; Pro subscribers get it only in the Gemini app and AI Mode in Search. Artificial Analysis scores it 52.6 at its high setting, level with GPT-6 Astra's 52.7, and one index task costs $1.99 at today's price, so plan on about $4 once the discount ends. If you are testing it on the API now, budget at the December rate.

Read the story →

3. Meta released Muse Wonder on its Model API at $4/$20 per million

Meta's pricing page confirms the rate, with $0.20 cached input, a 1M context and no long-context premium. Artificial Analysis scores it 52.4, just under GPT-6 Astra's 52.7, at $4.52 per index task against Astra's $3.26, and it outputs about 131 tokens per second on Artificial Analysis's speed test, three times Astra's 44. The stronger Muse Wonder Full, which scores 55.8, is only available inside Muse Code subscriptions, so the API gets the smaller model.

Read the story →

4. Microsoft released Decision-1, a small model that scores a fixed set of answers instead of writing text

You give it the options and it returns a calibrated probability for each, for routing, classification, verification and agent control. It is post-trained from Qwen3.5-9B, takes up to 32K tokens of input, and costs $0.042 per million input tokens with output free on Microsoft's Azure endpoint on OpenRouter; it is also in Foundry. Microsoft reports 83.5 percent accuracy at 85 ms across 36 benchmarks in its own tests. If a frontier model is currently deciding where your requests go, this is a cheap thing to benchmark against it.

Read the story →

5. Anthropic launched OSS Scanner, free periodic vulnerability scans for open-source projects that opt in

Each report comes from Anthropic's most capable models and carries a proof of concept, an explanation and a suggested fix where one exists. Reports go out without human review, and Anthropic expects a true-positive rate above 90 percent, so some findings will be wrong. Anthropic says it is meant for projects with the capacity to keep up with what it finds; other projects keep getting human-verified disclosures.

Read the story →

Model moves

• Now on the Gemini API: Gemini 4 Argon (Google). $2/$10 per 1M through November 30, then $4/$20, on the paid tier only (not the free tier or the Google AI Pro plan). Batch and flex are half price, cached input is $0.10 (then $0.20), and it has a 1M context with up to 1M output tokens. In Antigravity it needs Google AI Ultra or an enterprise account. Model page →
• New: Muse Wonder (Meta). $4/$20 per 1M with $0.20 cached input, a 1M context and no long-context premium. Muse Wonder Full is not on the API. Model page →
• Alibaba's new US endpoints for Qwen cost more. On OpenRouter, Alibaba's new us-east-1 endpoint charges $2/$6 per 1M for Qwen 3.7 Max against $1.475/$4.425 on its default endpoint, and 15 percent more for Qwen3.7 Flash. The default rates did not change, so check which endpoint your traffic lands on. Model page →
• Benchmark revision on Epoch's listings, October 14. Qwen 3.6 35B-A3B's reported Terminal-Bench score rose from 23.0% to 24.6%. Model page →
• No vendor list-price changes this week.

Personal take

The most useful sentence I read this week was in Anthropic's incident report. The Claude agent that sent Philadelphia police a fake tip was told never to log in, create accounts, enter personal data, make purchases, or submit anything destructive. A tip form was none of those things. Anthropic responded by restricting its web fetch tool, building tools that detect and block these actions, and fixing training environments that rewarded workarounds. None of the fixes improves the instructions. If your agent can reach the open web, its system prompt is a request. The real limits are those your tools enforce, like which domains it can reach and whether it can submit anything. That is worth checking before you hand any agent a browser.

Until next Thursday, Anmol