Gemini 4 Pro / Flash: benchmarks, hands-on tests and leaks

In Robocurve's real-robot tasks, Gemini 4 Argon and Opus 5.5 both score around 48%, GPT-6 Astra about 47%, and Kimi K3.1 Pro only 9%.

Google Cloud and OpenAI announced a partnership: GPT-6 Astra and siblings enter private testing billed through Google Cloud, while OpenAI models in Antigravity remain planned.

A user says Gemini 4 Argon writes excellent prose and jokes that humans trained on Chinese essay formats are outmatched by the model's parameters and TPM.

A developer who spent a full day testing Gemini 4 Argon concludes Google may have no model stronger than Claude Opus 5.5, casting doubt on the rumored Gemini 4 Carbon too.

A user compares subscription quotas in 'pelican riding a bike' units, claiming Gemini 4 Argon leaves 98% weekly quota after one run and could do 50 a week.

A leak says Google has started internal team testing of Gemini 4, with IDs like carbon-g-medium-tf and gemini-4-argon-max-mode-tf carrying a -tf suffix.

A user reports hitting a possible Gemini 4 Flash or Barium canary test where requests route to gemini 4 argon b, with the pelican test showing some improvement.

After a week using Opus 5.5, Sol 6.1, Grok 4.7 and Gemini 4 Argon, a user's first takeaway is that Argon has the tightest usage limits.

Four models ran the same western saloon video prompt; Gemini 4 Argon took 3 minutes and 15 cents and built swinging doors, while Opus 5.5 took 39 minutes and $4.80.

A Turkish user praises Argon 4 for pinpoint accuracy and says global posts over recent days show clear performance gains, urging an end to Gemini pile-ons.

Citing Anthropic data that manual step review catches only 14% of errors, the post shares a three-step method using Opus 5.5 or Gemini 4 Argon as an independent QA chat.

A developer whose zero-shot Dream-1 VLA test failed lists three fixes, one using Gemini Argon or Opus 5.5 for VLM reasoning plus IK tools.