Google cuts Gemini 4 Carbon prices before Gemini 2.5 sunset

Google is tightening its developer migration timeline with aggressive pricing ahead of next month’s legacy endpoint cutoff: Gemini 4 Carbon now sits #3 on the overall leaderboard while promotional $2 / $10 API rates remain locked through November 30 — one day before gemini-2.5-pro and gemini-2.5-flash shut down permanently on December 1.

At the same time, Gemini 4 Hydrogen (google/gemini-4-hydrogen) reached general availability on November 12 at $0.50 input / $2.50 output per million tokens, giving enterprise teams a clear 1M-context replacement ladder alongside ultra-fast google/gemini-4-argon.

Where the Gemini 4 lineup lands on quality and cost

Across more than 378 million evaluated sessions on Arena and Vigas model benchmarks, Google’s fourth-generation family splits neatly by workload:

• Gemini 4 Carbon (High): Ranks #3 Overall at 16.90% win share ($1.42 estimated cost per completed task) and claims #1 in Text ahead of competing proprietary and open-weight releases.
• Gemini 4 Hydrogen (High): Debuts directly at #5 Overall (14.80% win share, $0.92 per task), combining strong reasoning accuracy with 4× lower token pricing than standard Carbon rates.
• Gemini 4 Argon: Sits cleanly at $0.30 / $2.50 per 1M tokens with 1M context for high-volume multimodal extraction and Flash-era legacy replacements.

Avoiding breaking migrations before December 1

Teams still hitting gemini-2.5-pro or gemini-2.5-flash in production face a strict December 1 shutdown deadline across Vertex AI and Gemini API keys. Promotional pricing on Carbon doubles from $2/$10 back to $4/$20 after November 30, making early cutovers twice as valuable: run staging evals during the discount window before standard billing resumes.

Engineers upgrading apps via the Vigas gateway or compatible research and coding tools should audit three areas during switchover: structured JSON output schemas, thinking-budget latency caps, and automatic fallback chains (gemini-4-hydrogen -> gemini-4-argon) so a single provider deprecation never breaks production API traffic again.