Google DeepMind released Gemini 3.8 Flash on September 2, internal codename Skimaki. Third Flash update in three months. Nobody else in the frontier tier ships this fast.
It’s a coding-first model aimed at long-horizon software engineering and autonomous agents — the exact territory Anthropic owns. Per the WSJ, Google engineers ran it against Claude Opus on Jetski, Google’s internal coding platform, and preferred Skimaki. Google credits a heavy reinforcement learning push since January. The claim landed on Hacker News’ front page with 139 points.
The numbers behind the trash talk
On Vals Finance Agent v2, 3.8 Flash scores 61.4% — ahead of Opus 5 (58.6%) and GPT-5.6 Sol (53.8%). Best score on Harvey’s Legal Agent Benchmark too. On DeepSWE v1 hard tasks it hits 71.0% versus Opus 5’s 74.0% — still behind, but at Flash prices. That gap-per-dollar is the real story.
API access
Live now in the Gemini API: $0.75 per 1M input tokens, $3.75 output, intro pricing through December 2026 — unchanged from 3.7. Obvious move: swap it into your coding agent or finance/legal agent pipeline and see if the internal hype holds.
You Might Also Like
- Claude Opus 4 7 Goes Wall Street First 64 4 on Vals Finance 1m Context in Claude Code
- Anthropic Ships Claude for Legal Going Head to Head With Harvey
- Google Gemini Enterprise Gives Every ai Agent a Cryptographic id its Real Fight With Openai and Anthropic
- Hiveterm Bets on the Multi Agent Workspace Claude Codex and Gemini in one Terminal
- Google Rolls out File Generation in Gemini Chasing Chatgpt Canvas and Claude Artifacts
Top AI Product 
Leave a comment