Gemini 4 Hydrogen benchmarks leak: near-flagship intelligence at a fraction of the cost

Reviews of Gemini 4 Hydrogen, Google DeepMind's third new model in six weeks, leaked en masse on November 13, with multiple bloggers and Reddit users circulating evaluations and benchmark tables from Artificial Analysis. Taken together, the data suggests the small-sized model delivers near-flagship intelligence at significantly lower cost—cited by several commentators as the latest example of "frontier intelligence rapidly demonetizing." Notably, neither the model name nor some competitor version numbers have been officially confirmed by Google.

Confirmed

• Artificial Analysis (@ArtificialAnlys) published an in-depth review, stating Hydrogen scores just one point below Claude Sonnet 5.5 in intelligence at roughly one-third the cost, and introduces an xhigh reasoning tier for the first time.
• The official comparison table shows: SWE-sweep 13.5%, HLE-Diamond 58.8%. ChrisGPT noted the SWE-sweep score is about 2.3x that of GPT-6.1 Terra (5.8%), also higher than Claude Sonnet 5.5.
• Compiled by markk: an Artificial Analysis Intelligence Index score of 51, surpassing Gemini 4 Argon, with fabricated claims reduced 5x. A leaderboard screenshot (v5.0.1) shared by likeastar20 shows Hydrogen (high) at 54, right behind Gemini 4 Carbon and closing in on Claude Opus 5.5.
• It tops Agents' Last Exam v1.1 at 71.6%, beating Claude competitors; suspected input pricing of $1.5/M tokens (unconfirmed).
• Based on AA data, haider1 compared and found Hydrogen's per-task cost is more than 4x lower than Gemini 3.8 Flash.

Unconfirmed

• The model name "Gemini 4 Hydrogen" and the competitor version numbers involved (e.g., GPT-6.1 Terra, Claude Sonnet 5.5) are unusual; posters such as heypearlai explicitly flagged their authenticity as questionable, and Google has not confirmed them.
• @danielmckinn0n claimed that on rare-disease-related laboratory benchmarks, Hydrogen beats Anthropic Opus 5.5 on most tasks while being faster and cheaper; the post itself notes it is unconfirmed by Google.

Why it matters

• Hydrogen is positioned as the low-cost, high-speed line replacing Flash; marktenenholtz argued its SWE-sweep 13.5% is an outstanding result for a model in the small-size class.
• Using AA's cost-vs-intelligence scatter plot, Electronic-Weekend66 argued frontier intelligence is rapidly getting cheaper, with Cheap-tier models already closing in on Anthropic's flagship. However, poigre's analysis based on ai-pareto.web.app pointed out that Hydrogen's tiers deviate from the intelligence/cost Pareto frontier, so assessments of its value for money remain divided.

2026-11-13 ~ 2026-11-13 · 26 related posts

Full story(5 episodes)→

• Episode 1: Google Launches Gemini 4 Hydrogen Across All Products(2026-11-13, 70 posts)
• Episode 2: Gemini 4 Hydrogen benchmarks leak: near-flagship intelligence at a fraction of the cost(2026-11-13, 26 posts)
• Episode 3: Gemini 4 Hydrogen Ranks 5th on Agent Arena as Fastest Model(2026-11-13, 4 posts)
• Episode 4: Researchers slam Gemini's chemical-element model naming as confusing(2026-11-13, 2 posts)
• Episode 5: Viral "Gemini 4 Hydrogen" Launch Claims Turn Out to Be Fake(2026-11-13, 4 posts)

• [source] Gemini 4 Hydrogen scores 54 on AA Index, 1 point behind Sonnet 5.5 at ~32% of its cost — ArtificialAnlys · 2026-11-13
• [source] Gemini 4 Hydrogen bench table: leads on HLE-Diamond and agent evals, trails Claude Sonnet 5.5 on some — fofrAI · 2026-11-13
• AA index chart: Gemini 4 Hydrogen scores 54, near Pareto frontier on cost — likeastar20 · 2026-11-13
• Blogger: Gemini Hydrogen is 25% cheaper and 2.5x faster than Sonnet as OpenAI goes MIA — bindureddy · 2026-11-13
• Gemini 4 Hydrogen scores 13.5% on SWE-sweep, 2.3x GPT-6.1 Terra, but leads unevenly — ChrisGPT · 2026-11-13
• Gemini 4 Hydrogen Benchmarks: Scores 51 vs Argon's 50, Cuts Fabricated Claims 5X — mark_k · 2026-11-13
• Rumor: Gemini 4 Hydrogen 2.5x faster than 3.8 Flash, orchestrator offloads routine tasks — weskambale · 2026-11-13
• Gemini 4 Hydrogen cuts cost per intelligence task 4x vs Gemini 3.8 Flash, $7.77 to $1.87 — haider1 · 2026-11-13
• Leaked Gemini 4 Hydrogen benchmarks claim agent wins over Sonnet 5.5 at $1.50/M input tokens — heypearlai · 2026-11-13
• Gemini 4 Hydrogen posts 13.5% on SWE-sweep, an standout result for a small-class model — marktenenholtz · 2026-11-13
• Lab claims Gemini 4 Hydrogen beats Opus 5.5 on most rare-disease benchmarks at fraction of price — danielmckinn0n · 2026-11-13
• Gemini 4 Hydrogen hits frontier-level scores at 1/3 the cost of Anthropic's flagship — Electronic-Weekend66 · 2026-11-13
• [source] Gemini 4 Hydrogen early tests: 5pp more accurate, half the time, 12% fewer tokens — demishassabis · 2026-11-13
• Gemini 4 Hydrogen Tiers Fall Well Off the Intelligence-vs-Cost Pareto Front — poigre · 2026-11-13
• Google ships Gemini 4 Hydrogen: beats Argon in benchmarks at ~32% of Sonnet 5.5's cost — APPSO · 2026-11-13
• Unverified post claims Google launched Gemini 4 Hydrogen, 2.5x faster than Flash — goyalshaliniuk · 2026-11-13
• Gemini 4 Hydrogen deep dive: 8% hallucination rate, beats Argon on AA index, 1M-token I/O — 新智元 · 2026-11-13
• Gemini 4 Hydrogen builds a 3D desert ranch in 56s vs 6+ minutes for 3.8 Flash — minchoi · 2026-11-13
• Google's cheaper Gemini 4 Hydrogen scores 51, edging out its own flagship Argon — heyshrutimishra · 2026-11-13
• One-shot Minecraft anatomical bust: Gemini 4 Hydrogen builds skull and ribs, Argon turns zombie — iamfakhrealam · 2026-11-13
• Gemini 4 Hydrogen generates animated 3D desert ranch in under a minute, 3x faster than Flash — iamfakhrealam · 2026-11-13
• Gemini 4 SVG benchmark: Hydrogen finishes in 4 min at $0.42, beats pricier Argon — iamfakhrealam · 2026-11-13
• One-shot HTML zoom from Earth to Statue of Liberty; Gemini 4 Hydrogen does SVG 4 min for $0.42 — iamfakhrealam · 2026-11-13
• Gemini 4 Hydrogen builds 400K-voxel Chinese mountain temple in Blender, zero manual input — iamfakhrealam · 2026-11-13
• Rumored Gemini 4 Hydrogen Testing: 1M Context, $1.50/$7.50 per 1M Tokens — iamfakhrealam · 2026-11-13

1 near-duplicate retellings: XueFz