Gemini 4 Argon, Sonnet 5.5 and What Models You Should Be Using Right Now
Google's Gemini 4 Argon looks frontier-level on benchmarks but is unreleased and unproven, while Sonnet 5.5 is capable but token-hungry, and product experience is now competing with raw model quality.
Watch on YouTube ↗Charlie's summary
- ▶ 10:06
Gemini 4 Argon puts Google back on the benchmarks, but patchy coding scores and a full-price cost problem complicate the story
Gemini 4 scored 68.9% on the VAULT index, ahead of Opus 5.5, and more than tripled the Harvey legal benchmark leader with 19.6%. Coding was mixed: it trailed on frontier suite and was bottom of the pack on Terminal Bench 4.0 at 57.4%. Still, the host sees it as a big catch-up for Gemini's biggest weak point. At full price it would cost more than Astra, and even discounted it is more than twice GPT-6.1 Soul.
- ▶ 13:10
Announcing Gemini 4 without releasing it invites skepticism, given Google's benchmark track record and internal grumbling
Access is limited to trusted cyber defenders, with a gradual rollout starting with API and Ultra subscribers. Skeptics note the official Gemini 3.1 Pro benchmarks also looked dominant. Bloomberg reported employees saying Gemini 4 struggles on some coding tasks, which Google denied. Logan Kilpatrick says new Gemini versions now go through thousands of engineers for weeks. The host's takeaway is that all we have are self-reported benchmarks until people can try it.
- ▶ 17:12
Sonnet 5.5 is a big jump over Sonnet 5, but its token hunger undermines the cost-savings claim
Sonnet 5.5 scored 70.6% on Terminal Bench 4.0, up from 10.3% for Sonnet 5, and beat Opus 5.5's 66.4%. Artificial Analysis ranked it second on its intelligence index. But it cost $7.60 per task, 27% more than Opus 5.5, because it uses so many tokens. Turning settings down to extra high cut cost by two-thirds. Anthropic's 30% savings claim held only on similar tasks.
- ▶ 19:43
Sonnet is not just a cheaper Opus: use Opus to plan, Sonnet to implement, and treat Sonnet as a sub-agent
Builder Kun Chen says Sonnet matches Opus only on problems that don't need much wisdom. Opus asks “what's the real goal here,” while Sonnet focuses on how to get the task done. His setup is Opus for planning, Sonnet for implementation, and Fable when things go sideways. Theo calls Sonnet 5.5 an incredible model you probably shouldn't use standalone, since it overthinks at max settings. He sees it as more efficient as a sub-agent.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.