Gemma 4 can't match Opus or ChatGPT. That stopped mattering for most AI workloads.

A software engineer at a Munich fintech pays Anthropic $5 per million input tokens and $25 per million output tokens for Claude Opus 4.6. Her team summarizes regulatory filings, generates compliance checklists, and debugs Python scripts. The work is real. Her invoice last month: $4,200.

Across town, a three-person startup does the same work on a refurbished RTX 4090 that cost them $900 at a Kleinanzeigen listing. Google's Gemma 4, released April 2 under Apache 2.0, processes their queries for the cost of electricity. The model ranks third among all open models on the Arena AI leaderboard, with an Elo score of 1,452, sitting above models 20 times its size.

Both teams ship working products. Both sleep fine. The difference between them is seven benchmark points and several thousand euros per month.

That gap defines the new economics of AI deployment. Not whether open models have gotten good enough, because the latest generation clearly has for routine work. The harder question is narrower: which specific tasks still demand a frontier API, and how much of your budget should keep flowing to Anthropic, OpenAI, or Google's cloud tier? Gemma 4 sharpens that question more than any open model before it. It is the first family that delivers competitive intelligence from a phone to a workstation, under a license that adds zero legal friction.

Buy the GPU or keep renting tokens

Think of Gemma 4 as buying rather than renting your AI capacity. When you call Claude or ChatGPT through an API, you lease intelligence by the token. When you run Gemma 4 locally, you own the inference pipeline. No metering. No rate limits. No data leaving your network.

Google built Gemma 4 on technology from Gemini 3 Pro, its commercial flagship. The family ships in four sizes. The E2B model, with 2.3 billion effective parameters, runs on a Raspberry Pi 5 with 8 GB of RAM. The E4B handles any laptop with 8 GB. The 26B Mixture-of-Experts model activates only 3.8 billion parameters per token and fits on an RTX 4090. The 31B dense model, the strongest in the family, needs roughly 20 GB of VRAM at 4-bit quantization.

The hardware math works out cleanly:

Start at the bottom. A Raspberry Pi 5 with 8 GB squeezes the E2B into about 5 GB of memory, producing usable responses for simple scripts in five to six minutes. Slow, but it works. On a 16 GB laptop the E4B hits 10-plus tokens per second, fast enough for interactive chat. Jump to an RTX 4090 and you can load the 26B MoE or the 31B dense model with room to spare, strong coding and reasoning at consumer prices. If you own a Mac Studio with 64 GB unified memory, the 31B loads at 8-bit quantization. Near full quality, no fan noise drama. Past that, you are looking at an RTX 5090 or an H100 for unquantized weights.

Someone who grabbed a used RTX 3090 for $500 can run the 26B model today. A team with a Mac Studio runs the 31B at near-full precision. Total investment buys infinite inference. No monthly bill. No vendor dependency. That proposition has made Google emboldened enough to release the model under the same Apache 2.0 license used by Linux and Kubernetes, matching what Alibaba does with Qwen while Meta still restricts Llama beyond 700 million monthly active users.

Where seven points vanish

On the benchmarks that matter for everyday work, the gap between Gemma 4 and the frontier closed models is consistent but narrow.

GPQA Diamond, testing graduate-level science reasoning: Gemma 4 31B scores 84.3%. Claude Opus 4.6 hits 91.3%. GPT-5.4 reaches 92.8%. Gemini 3.1 Pro leads at 94.3%. The gap runs seven to ten points depending on the competitor.

LiveCodeBench v6, a coding benchmark: Gemma 4 posts 80.0%. Opus holds the SWE-Bench Verified crown at 80.8%. MMMU Pro, testing visual reasoning: Gemma 4 manages 76.9%. Opus leads at 85.1%.

The numbers tell a consistent story. Gemma 4 trails the frontier by a tier, not a generation. On a chart, you see clear water. In a blind comparison of outputs on typical business tasks, most users cannot tell the difference. Arena AI confirms this: real users voting on blind pairs rank Gemma 4's 31B responses above open models with 10 to 30 times more parameters.

And then there is cost. The FoodTruck Bench, which simulates 30-day business operations requiring multi-step planning, found that Gemma 4 31B produced a 1,144% median ROI at $0.20 per simulation run. Claude Sonnet 4.6 cost roughly 40 times more per run. Gemini 3 Pro cost 15 times more. For teams watching their cloud spend, the math is hard to argue with.

But averages hide the failures that justify premium pricing.

Where the cracks run deep

A developer who tested Gemma 4 31B on real-world coding benchmarks, requiring the model to scaffold a Rails application with specific gem dependencies, watched it spiral into an infinite tool-calling loop after 11 productive steps. Claude Opus 4.6 completed the same task cleanly. GPT 5.4 worked when tested separately, though it failed the benchmark runner itself. The report was blunt: only three model families, Claude, Z.AI's GLM 5, and GPT 5.4, produced code that actually worked. Gemma 4 was not among them.

Speed compounds the problem. Community testers measured the 26B MoE model generating 11 tokens per second on hardware where Alibaba's Qwen 3.5 cranked out 60-plus. Google's MoE architecture was designed for efficiency. In practice, it chokes. The model loads all 25.2 billion parameters into VRAM even though only 3.8 billion activate per token. Routing overhead eats the theoretical savings alive.

VRAM hunger creates a second constraint. On identical hardware with identical quantization, Qwen 3.5 27B supported a 190,000-token context window. Gemma 4 fit roughly 20,000 on the same card. If your workflow depends on processing long documents or maintaining extended conversations, that difference stings more than any benchmark score.

Fine-tuning tooling arrived broken on day one. HuggingFace Transformers did not recognize the architecture. PEFT could not handle a new layer type in the vision encoder. Teams planning domain-specific adaptation found themselves waiting for upstream patches while Qwen models worked out of the box.

These are not edge cases you can engineer around. They define the boundary where "good enough" limps into "not ready."

The real competitor is not Claude

Here is the part most coverage misses. If you are choosing a model to run locally, your realistic alternative to Gemma 4 is not Claude Opus or GPT-5.4. You cannot run those on your hardware. Your alternative is Alibaba's Qwen 3.5.

The head-to-head comparison at the 30B scale is a draw on benchmarks, with Qwen winning slightly more categories. Three differences tip the balance depending on what you build.

Speed. Qwen 3.5 27B pushes roughly 35 tokens per second on an RTX 4090 with Q4 quantization. Gemma 4 31B manages about 25. For interactive use, that gap is noticeable.

VRAM efficiency. Qwen uses a hybrid attention architecture that produces a 75% smaller KV cache than standard transformers. Longer conversations and bigger context windows become practical on the same GPU.

Multilingual quality. Gemma 4 outperforms Qwen 3.5 on non-English tasks by a clear margin. Community testing across German, Arabic, Vietnamese, and French shows Gemma 4 producing markedly better translations and multilingual reasoning. One tester called it "in a tier of its own" for non-English work.

At the laptop tier, Qwen dominates outright. Its 4B model beats Gemma 4 E4B by double-digit margins on most benchmarks. Gemma 4's only advantage at that scale is native audio input, valuable for on-device voice applications but irrelevant for text work.

The Qwen comparison should make Anthropic and OpenAI more anxious than the Gemma 4 launch itself. Two open model families, both Apache 2.0, both running on consumer hardware, both trading blows at quality levels that were proprietary-only 18 months ago.

Owners, renters, and the split stack

The ownership question breaks the decision into three groups.

Own confidently. Teams with clear privacy requirements, multilingual workloads, or high-volume inference where per-token costs compound. A team processing 10 million tokens daily pays nothing with Gemma 4 versus roughly $2,500 monthly for Opus at blended rates. That math resolves itself. Edge deployers building offline products on phones or IoT devices have no closed-model alternative at all. Gemma 4 E2B is the only model in its performance class that fits on a phone.

Rent selectively. Anyone whose core workflow depends on complex multi-step code generation, agentic reliability, or sustained reasoning across very long contexts. The infinite tool-call loops and the speed cliff make Gemma 4 a poor fit for production agentic systems today. VRAM constraints narrow the gap further. Opus 4.6 at $5 per million input tokens remains the cheapest path to 80.8% on SWE-Bench Verified. GPT-5.4 leads on professional knowledge work across 44 occupations at 83% on the GDPval benchmark. Those gaps look small on paper. In production, they separate shipping from debugging.

Split the stack. The smartest teams in 2026 route tasks by complexity. Gemma 4 handles summarization, translation, document Q&A, and simple code generation locally. Claude or GPT handles the hard cases through API. The local model absorbs the majority of volume, and cloud bills drop proportionally. One analysis estimated that about 70% of typical enterprise AI workloads fall within what current open models handle well. The remaining 30% is where frontier APIs earn their price.

The invoice wins

Google released Gemma 4 into the most competitive open-model field AI has ever seen. Six families now ship competitive weights under permissive licenses. The closed-model premium buys you seven to ten points on the hardest benchmarks and meaningfully faster inference for agentic work. It also buys tooling that works without day-one patches.

For the Munich fintech engineer paying $4,000 a month, the question is concrete. She could provision a $2,000 GPU, run Gemma 4 for routine document processing, and cut her Anthropic spend by 60% inside a month. The hard work, the compliance edge cases that need surgical precision, keeps going to Claude.

That split is where enterprise AI budgets head over the next 12 months. Not a clean replacement of closed models. Not loyalty to the incumbents either. A division of labor that matches the model to the stakes, and stops paying frontier prices for work that does not need frontier intelligence.

Gemma 4 did not win on the leaderboard. It won on the invoice.

Alibaba Releases Third Closed-Source AI Model in Three Days, Signals Profit-First Strategy

Three proprietary AI models in three days. That's what Alibaba shipped this week, with Qwen3.6-Plus landing Thursday as the third, Bloomberg reported. This one handles agentic coding and multimodal re

The Implicator

Vienna Builds. San Francisco Hires.

San Francisco | Monday, February 16, 2026 Peter Steinberger built OpenClaw into the fastest-growing open-source project in GitHub history from a Vienna apartment. 196,000 stars in three months, 600 c

The Implicator

OpenAI Launches Frontier to Manage AI Agents From Rival Vendors in One System

OpenAI on Thursday released Frontier, a platform that lets companies build, deploy, and manage AI agents from multiple vendors inside a single system. The product works with agents built by OpenAI, by

The Implicator

Harkaram Grewal

New Delhi

Freelance correspondent reporting on the India-U.S.-Europe AI corridor and how AI models, capital, and policy decisions move across borders. Covers enterprise adoption, supply chains, and AI infrastructure deployment. Based in New Delhi.