Google unveils Gemini 4 Argon for a limited group, at half its standard price for now

1. Gemini 4 Argon leads or ties on 13 of 18 published benchmarks, while Opus 5.5 still wins on terminal tasks

   First seen on X 30 hours ago@demishassabis ♥ 1,537

   In the table Google DeepMind published, Argon scores 68.9% on the Vals Index, ahead of GPT-6 Astra (63.1%), Claude Fable 5.1 (65.8%) and Claude Opus 5.5 (67.0%). On AutomationBench, which tests business work, it scores 51.3% against 41.4% for Astra and 42.5% for Opus 5.5. On a legal agent benchmark it scores 19.6% against 5.4% and 3.8%. On DeepSWE v1.1, which measures real software engineering, Argon reaches 77.9% against 74.1% for Astra and 74.2% for Opus 5.5. It scores 91.7% on the long-video test LVBench and ties Astra at 68% on the vulnerability-fixing test CWE-bench.

   The table also shows where it trails. On Terminal-Bench 4.0, Opus 5.5 scores 66.4%, nine points above Argon's 57.4%, and on FrontierSWE v2 Astra's 65.5% beats Argon's 55.0%. All of these are numbers Google published itself, and independent verification has not arrived yet.

   Keep reading ↓

   This was helpfulShare

2. Artificial Analysis scores Gemini 4 Argon at 53, level with GPT-6 Astra, at about 60% of the cost per task

   First seen on X 30 hours ago@aipulseda1ly ♥ 921

   The independent evaluator Artificial Analysis gave Argon an Intelligence Index score of 53. That matches GPT-6 Astra, and it costs $1.99 per task to run the index, about 60% of Astra's cost. That figure uses the 50%-off launch price, so at standard pricing it rises to $3.98 per task. Argon produced 110 million output tokens while running the index, against a median of 82 million, so it is on the wordy side.

   Its hallucination rate is 15%, reportedly the lowest among models scoring 45 or higher on the index. Vals AI said Argon took first place on the Vals Index at 68.9%, the first time a Gemini model has topped it, and Argon is now available on Design Arena.

   Keep reading ↓

   This was helpfulShare

3. With /advisor fable in Claude Code, Opus 5.5 checks with Fable 5.1 before it plans and before it calls a task done

   First seen on X 37 hours ago@angeldot_ ♥ 706

   Type /advisor fable in Claude Code and Opus 5.5 keeps doing the work, calling Fable in as an advisor only at key moments. The official docs say it tends to consult before committing to an approach, when an error keeps recurring, and before declaring the task done. The advisor receives the full conversation, including every tool call and result, and sends back guidance. The model decides when to call it, and you can ask for a consultation in your prompt, for example "consult the advisor before you continue."

   There are three ways to set it: the /advisor command mid-session, the advisorModel setting for a default, or the --advisor flag at launch. Run /advisor alone for a picker and /advisor off to turn it off. Advisor tokens count toward your plan limits, and on some plans a Fable advisor bills to usage credits, so you first give consent with /model fable. It adds little on short tasks and fits long refactors or debugging where the same error keeps coming back.

   Keep reading ↓

   This was helpfulShare

4. First Codex reviews of GPT-6.1 Sol say it burns quota slowly but works slowly too

   First seen on X 2 days ago@blader ♥ 3,844

   Reviews of GPT-6.1 Sol are coming in the day after DevDay. One Codex user ran Sol at max for 52 minutes, used 2.8 million tokens and spent 4% of the weekly limit and 17% of the 5-hour limit, saying a $20 plan can go a long way. The same post adds that it is extremely slow. A similar issue appeared on the Codex repository: the 5-hour window drains two to three times slower than with the previous GPT-5.6 Sol, but tasks that took a few minutes now take 10 to 12, so real throughput has not improved.

   Others rate its efficiency highly. @blader wrote that in internal evals Opus 5.5 was twice as efficient as GPT-6 Sol, and 6.1 Sol came out at three times Opus 5.5 and 2.5 times Sonnet 5.5. If speed is the problem, the Ultrafast mode announced at DevDay runs up to 8x faster in Codex, but only Pro 500 and Enterprise can use it for now.

   Keep reading ↓

   This was helpfulShare

5. Sign in with ChatGPT expands to third-party dev tools, so your ChatGPT plan can pay for Devin or Warp

   First seen on X 42 hours ago@lucian__03 ♥ 212

   At DevDay, OpenAI expanded Sign in with ChatGPT. ChatGPT Plus and Pro subscribers can use their own plan allowance inside third-party developer tools such as OpenCode, Devin, Amp, Warp, Notion and Vercel. Developers no longer have to cover their users' token costs. Users set weekly caps per app in ChatGPT settings, and usage stops once a cap is hit instead of moving to the partner's own billing.

   There are 16 launch partners, with Lovable joining soon. The beta from July only linked accounts; now it also lets people spend their subscription allowance. One Korean-language post said the announcement got little attention and urged people building AI services by vibe coding to take note.

   Keep reading ↓

   This was helpfulShare

6. Skills arrive in the Gemini app: save the instructions you use often and call them with a slash

   First seen on X 34 hours ago@Google ♥ 2,194

   Google has added skills to the Gemini app. You save an instruction you use often once, then type / in the prompt box to call it by name. You can stack several skills and attach reference files such as text, PDFs and images. It is rolling out now in Gemini chat worldwide on every subscription tier, and Google Workspace accounts get it within weeks.

   Skills replace Gems. Existing Gems on personal accounts migrate automatically to skills in November, on Workspace business and nonprofit accounts in March 2027, and on education accounts in June 2027. Sharing and Google Drive file integration are coming soon, Google says.

   Keep reading ↓

   This was helpfulShare

* Summaries and publishing on this page are done by an automated agent. Check the facts against each story's original links.

Privacy PolicySupport