Gemini 4 tops public coding benchmarks but drops to bottom on private tasks
Hesamation · x · 2026-10-30
Artificial Analysis (via Priors, Oct 29) compared two coding leaderboards and the rankings barely overlap:
- Public tasks, Terminal-Bench 4.0 (66 tasks, pass rate): Gemini 4 Carbon leads at 65.2, Gemini 4 Argon second at 57.1, Claude Opus 5.5 third at 56.6; Grok 4.7 last at 24.8.
- Private tasks, FrontierCode 2.0 (132 private tasks, merge rate): Gemini 4 Carbon still #1 (40.6), followed by GPT-6 Astra (40.1) and GPT-6.1 Terra (37.5); Gemini 4 Argon plummets to #9 at 25.7, barely above Claude Sonnet 5.5.
Takeaway: Gemini 4 beats Claude Opus 5.5 on public coding tasks but loses to Grok and most rivals on private ones — a stark divergence between public benchmarks and private-task performance.
More from coding & agent
- effect-harness on Cloudflare: run agent loops, tools and credentials on separate VMs — samgoodwin89 · 2026-10-31
- jh3yy: agents gave me flow state — my tools finally keep up with my brain — jh3yy · 2026-10-31
- CAFE ecosystem: Maple moves 300k lines of React to Foldkit in 3 days — samgoodwin89 · 2026-10-31
- Non-coder builds 100-star Android TV launcher entirely with Claude — SuB8u · 2026-10-31
- Factory acquires typedef_ai to give coding agents an enterprise data context layer — maddiehfaulkner · 2026-10-31
- Browser agent's graph reroute wins 5/5 vs 0/5 for trajectory replay, author retracts own number — JeremyCMorgan · 2026-10-31