The Frontier

One host found Gemini 4 Argan capable of polished 3D graphics but prone to last-generation coding failures: it added a damage power-up that activated between fights and left food where fish could not reach it.

In codebase audits, Gemini 4 Argan made 47 false claims out of 92 about T3 Code and 24 out of 36 about TS Rust. The hosts said it invented database fields, produced malformed Python, and could not be trusted for routine coding work.

Gemini 4 Argan scored 53 on Artificial Analysis's intelligence index, level with GPT-6 Astra and Fable 5.1, but the hosts said real coding use did not match the benchmark picture. Google listed it at $2 per million input tokens and $10 per million output tokens, then planned to double those prices after a temporary discount.

One host argued that Anthropic and OpenAI are widening their lead because stronger models let them iterate faster, while Google keeps improving at a slower pace without fixing reliability. The hosts also saw Grok as behaviorally more trustworthy than Gemini, but neither as a strong daily driver.

The hosts criticized Android and Google's devices for inconsistent software, clumsy foldable layouts, unreliable app behavior, and poor Gemini integration. They liked the Pixel Fold hardware but said the operating system wasted screen space, Google apps felt worse than their iOS versions, and the built-in Gemini could not act on the computer.

One host expects agents to replace search as people move from Google queries to conversational assistants, and argued that legacy software will lose users when easier alternatives arrive. The hosts compared this shift to music piracy giving way to convenient legal services such as iTunes and Spotify.

The hosts said agent-driven development has moved beyond manual code review: agents can route work, open pull requests, test changes, and review each other. They rejected the claim that agent-written code is automatically slop, citing T3 Code's rapid merges and products that improve through continuous agent work.

The hosts disagreed about how soon models can make product decisions without human direction. One argued that open-ended taste remains a human job; the other said agent panels, competing designs, and training on past feedback could automate more of that judgment.

DeepSeek V4.1 1017 was described as roughly Opus 5-level and usable locally, giving developers a way to run security tests or other tasks frontier services may refuse. A host said Chinese labs are especially good at getting strong performance from smaller models; the version-like suffix reportedly avoids a new Chinese review process.

Meta's Muse Spark Open was harder to run and less capable than leading Chinese open models, but Muse Wonderful stood out for clear communication and useful orchestration. One host said 64% of its changes landed, compared with 71% for Opus 5.5 and none of the Gemini changes tested.

Muse's launch confused users by releasing Wonder and Wonderful together, with different access, speed, and pricing information; Wonderful was the stronger model but lacked API access. In T3 Code, hosts used Muse to direct work while Opus implemented or reviewed changes and Codex handled computer-use tasks.

The hosts credited Anthropic's strong coding with careful training-data selection and a shared development culture between Claude and its builders. They contrasted Anthropic's growth with larger established software firms and said OpenAI needs a major model improvement, while noting that heavy Claude usage can quickly exhaust subscription limits.

One host argued engineers should hand tedious work to agents, even when that means building a bespoke tool rather than improving a general one. Examples included a simulated YouTube API, the Vampire Survivors team's prototype-to-port workflow, and TS Rust, which cut type-check builds while saving substantial CI time.

The Tesla Roadster's flying-car demonstration and stunt impressed the hosts as evidence that engineering can deliver visually striking new machines, though they questioned consumer safety and said the production timeline could slip. They singled out the Roadster's track-focused vacuum system for added traction as more practical than nitrogen thrusters.

Paper's desktop design editor lets agents work on shared HTML and CSS through an MCP server, while General Translation offers localization libraries, a CLI, and shared translation context across web, mobile, and desktop apps. Both sponsors were presented as tools designed to fit agent-assisted development.