---

Apertus vs GPT-5.1, Gemini 3.0 Pro and Claude 4.5 Sonnet

Why a fully open Swiss model is still worth watching in a frontier-model world

This report compares Swiss AI’s Apertus model
https://publicai.co/

with three leading large language models:

• OpenAI – GPT-5.1

• Google DeepMind – Gemini 3.0 Pro

• Anthropic – Claude 4.5 “Sonnet”

The comparison spans:

• Performance on standard benchmarks and a real classroom-style coding task

• Open-source status and licensing

• Training data transparency

• Model architecture and scalability

• Deployment cost and efficiency

• Safety, alignment and guardrails

The conclusion: Apertus is not a “GPT-5 killer”, but a strategically different choice that trades a bit of raw power for openness, sovereignty and transparency.

---

1. Model Overview and Availability

High-level snapshot:

Model	Developer	Scale	Open-Source?	Access / License	Notable Features
Apertus (70B & 8B)	Swiss AI Initiative (EPFL, ETHZ, CSCS)	70B & 8B params, ~15T tokens	Yes	Apache 2.0, weights + training data public on Hugging Face	1,800+ languages, 65k-token context, transparency and EU-AI-Act-ready documentation
GPT-5.1	OpenAI	Not disclosed (successor to GPT-5)	No	Closed API (ChatGPT, OpenAI API)	“Instant” vs “Thinking” modes, ~400k-token context, fully multimodal (text, vision, audio)
Gemini 3.0 Pro	Google DeepMind	Not disclosed (flagship Gemini)	No	Google ecosystem (Gemini app, Vertex AI, etc.)	SOTA reasoning & multimodality, “Deep Think” mode, tops many public benchmarks
Claude 4.5 Sonnet	Anthropic	Not disclosed (latest Claude)	No	Claude API, AWS Bedrock, Vertex AI	Agentic design for long-running tasks, 200k–1M token context, extremely strong coding/tool use

Key structural difference:
Apertus is the only fully open-source model here; weights, training recipes, and even training data are public. GPT-5.1, Gemini 3.0 Pro and Claude 4.5 Sonnet are proprietary, closed-weights models accessible only via API or platform integrations.

---

2. Performance Benchmarks (High-Level)

On standard benchmarks, GPT-5.1, Gemini 3.0 Pro and Claude 4.5 Sonnet sit at the frontier. Apertus aims instead for “LLaMA-3-era” performance with complete transparency.

Approximate picture:

Benchmark	Apertus 70B	GPT-5.1	Gemini 3.0 Pro	Claude 4.5 Sonnet
MMLU (academic knowledge)	~70% (LLaMA-3-level, not SOTA)	~84% (est.)	~90% (SOTA or near)	~89% (near SOTA)
GSM8K (math word problems)	N/A public; likely below SOTA	~90%+ (with tools higher)	~100% with code (AIME), ~95% without tools	~100% with code (AIME)
HumanEval (Python coding)	No public numbers; far lower in practice (see case study)	~92% (SOTA code generation)	~90%+ (competitive)	~90%+ (SOTA for coding agents)

Roughly: Apertus ≈ top open models of 2024.
GPT-5.1 / Gemini 3.0 Pro / Claude 4.5 ≈ frontier proprietary models of 2025.

---

3. Case Study: Can They Actually Build a Real Interactive?

To move beyond benchmarks, I tested all models on a real classroom-style coding task:

Task:
Create a complete, self-contained HTML5 interactive on the Trigonometry Unit Circle for Sec 3–4 students (Singapore level), in a single HTML file with embedded CSS and JavaScript.

Key requirements (abridged):

• Rich educational content (Sec 3–4 trigonometry)

• Interactive controls (sliders, buttons, checkboxes, drag-and-drop)

• Dynamic 2D visualizations (unit circle, graphs, triangle view)

• Real-time feedback and live readings of sin θ and cos θ

• Toggle between degrees/radians

• Touch-friendly (mobile responsive, 44px touch targets)

• Randomised problem generator with step-by-step hints and “Why?” explanations

• Real-time analytics panel for learning assessment (timestamps, action log, state values)

• All in one HTML file, ready to drop into an LMS

Apertus

• Outcome: Repeatedly produced incomplete code and got stuck at essentially the same truncated output each time.

  https://chat.publicai.co/c/75fe5835-66be-4ae4-a0b1-b90c61da0d73

• The model clearly understood the prompt, but could not sustain the long, complex generation needed for a full, production-ready interactive.

Reference: Apertus test log
https://chat.publicai.co/c/75fe5835-66be-4ae4-a0b1-b90c61da0d73

GPT-5.1, Gemini 3.0 Pro, Claude 4.5, DeepSeek v2.3

• GPT-5.1, Gemini 3.0 Pro and DeepSeek v2.3 were all able to generate complete, runnable HTML5 interactives, albeit with varying polish and bugs that a human developer would still need to fix.

• Claude 4.5 Sonnet is particularly strong at refactoring and improving such interactives, even if not always perfect on first pass.

• These models could:

  • Maintain the long prompt requirements

  • Create a working canvas/unit circle + sliders

  • Wire up basic analytics logging in JavaScript

Sample transcripts (for reference, not necessary in the blog):

• GPT-5.1: https://chatgpt.com/share/69269c75-42a8-8008-8268-5389633343ab 

  https://chatgpt.com/share/69269c75-42a8-8008-8268-5389633343ab

• Gemini 3.0 Pro: (Gemini project link)

  https://gemini.google.com/u/1/app/16fdb1e578a3431c?pageId=none

• DeepSeek v2.3: https://chat.deepseek.com/a/chat/s/b485bcee-29c1-46eb-bf3b-26a25d32b829 

  https://chat.deepseek.com/a/chat/s/b485bcee-29c1-46eb-bf3b-26a25d32b829

• Claude 4.5 Sonnet: https://claude.ai/chat/edb2b8ac-86fb-4f9a-9bf0-ce3fdeabf6ce 

  https://claude.ai/chat/edb2b8ac-86fb-4f9a-9bf0-ce3fdeabf6ce"}">https://claude.ai/chat/edb2b8ac-86fb-4f9a-9bf0-ce3fdeabf6ce

Takeaway from the case study
For complex, long-form, production-grade coding tasks, the frontier proprietary models are clearly ahead. Apertus behaves more like a strong open model that still needs careful chunking, scaffolding and possibly external tools to match this level of output.

---

4. Open-Source Status and Licensing

This is where Apertus truly stands out.

• Apertus

  • Fully open-source: weights, training code, and training data are public.

  • Licensed under Apache 2.0, allowing commercial use, modification and redistribution.

  • Downloadable from Hugging Face in 8B and 70B variants; you can self-host, fine-tune and even publish derivatives.

• GPT-5.1

  • Closed-source, proprietary model.

  • Access only via OpenAI’s APIs / ChatGPT.

  • No public weights or full training dataset.

• Gemini 3.0 Pro

  • Proprietary model embedded in Google’s ecosystem (Gemini app, Workspace, Vertex AI).

  • No weights or training data released; license is proprietary.

• Claude 4.5 Sonnet

  • Proprietary model; accessed via Anthropic API, AWS Bedrock, Vertex AI, etc.

  • No public weights or training corpus.

If you need self-hosting, auditability, and no vendor lock-in, Apertus is in a category of its own.

---

5. Training Data Transparency

Another area where Apertus is very different:

• Apertus

  • Trained on 15 trillion tokens from publicly available data only.

  • About 60% English / 40% non-English; explicitly includes many under-represented languages (Swiss German, Romansh, African and Asian languages).

  • Respects opt-out requests from websites.

  • Provides scripts to reconstruct the training dataset and a detailed record of the training pipeline.

  • Designed explicitly to comply with the EU AI Act transparency requirements.

• GPT-5.1, Gemini 3.0 Pro, Claude 4.5

  • Training data is described only in broad terms (web text, books, code, images, etc.).

  • No detailed list of sources, no way to reconstruct the corpus.

  • Users must largely trust the provider’s assurances about privacy, copyright and bias mitigation.

For researchers, regulators and public institutions, this level of openness from Apertus is rare and powerful: you can actually know what went into the model.

---

6. Architecture and Scalability

Apertus

• 70B-parameter decoder-only Transformer (plus an 8B variant).

• Uses modern tricks: a custom activation (xIELU) and optimizer (AdEMA-Mix).

• Trained from scratch, then aligned with supervised fine-tuning and QRPO.

• Supports 65,536-token context – unusually long for an open model.

• All architectural details are fully documented.

• Not multimodal in practice: while UI may allow image upload on some demos, Apertus currently behaves as a text-only model in real use.

GPT-5.1

• Architecture details are proprietary, but likely a very large Transformer with mixture-of-experts and extensive tool-calling support.

• Provides:

  • ~400k-token context window.

  • Instant vs Thinking modes (adaptive compute – spending more “thinking tokens” on hard questions).

  • Full multimodality: text, images, and audio/video.

Gemini 3.0 Pro

• “Built from the ground up” as a multimodal model (text, images, code, video).

• Multiple variants (Pro, Flash, Lite, Deep Think).

• Very long context (hundreds of thousands of tokens), plus strong tool orchestration.

• Designed as a “one model for many domains” system.

Claude 4.5 Sonnet

• Transformer-based, optimized for agentic, long-horizon tasks.

• Context window: 200k tokens, extendable up to 1M for special use cases.

• Highly optimized for long sessions (e.g. 30+ hours of coding with stable memory and “checkpoints”).

• Strong code execution and file-creation tools built in.

Summary
All four models push context and scale in different ways. Apertus demonstrates what public institutions can do at 70B / 65k context; the proprietary models push further with trillions of parameters, multimodality and more aggressive context scaling – but behind closed doors.

---

7. Cost of Deployment and Inference

Apertus

• No API fees – but you pay for hardware.

• 70B model typically needs multiple high-VRAM GPUs for low-latency inference, especially with long contexts.

• 8B model can run on a single 16–24 GB GPU or even CPU (slow).

• Can be quantized, pruned and fine-tuned by the user; supports efficient inference engines like vLLM/SGLang.

• Attractive if you:

  • Already have GPU infrastructure (e.g. national supercomputers, institutional clusters), or

  • Expect very high usage and want to avoid per-token API bills.

GPT-5.1

• Available as a paid API service.

• Typical pricing (late-2025 ballpark): roughly US\(1.25 / 1M input tokens and US\)10 / 1M output tokens, cheaper than GPT-4 and lower than Claude for output.

• Adaptive reasoning (Instant vs Thinking) reduces cost for simple tasks.

Gemini 3.0 Pro

• Accessed via Google Cloud (Vertex AI) and the Gemini app.

• Pricing is broadly in line with other frontier models; may be bundled into enterprise contracts.

• Highly optimized on Google TPUs; benefits from caching and batching in production.

Claude 4.5 Sonnet

• Premium pricing: ~US\(3 / 1M input tokens and US\)15 / 1M output tokens.

• Positioned as a high-end model for complex, long-running tasks.

• No self-hosting; entirely cloud/API based.

Cost pattern

• For small to medium usage: GPT-5.1 / Gemini / Claude are much easier — no hardware to manage, just pay per token.

• For very large-scale, sensitive or long-term deployments: Apertus may become cheaper and more strategic, especially if you can amortise hardware costs and use the 8B model for simple tasks while reserving 70B for harder ones.

---

8. Safety, Alignment and Guardrails

Apertus

• Training pipeline deliberately filters personal and toxic data; respects opt-outs.

• Post-training alignment with QRPO, but no strict output filter is shipped by default.

• Model card explicitly says that real-time content moderation is the user’s responsibility (filters are planned but not enforced yet).

• Focus is on legal and ethical compliance via transparency, rather than heavy, baked-in refusal behaviour.

GPT-5.1

• Strong, built-in content filters and refusal behaviour.

• Extensive RLHF, red-teaming, and continuous safety updates; significantly fewer successful jailbreaks compared to GPT-4.

• Tends to be helpful but will decline disallowed content consistently.

Gemini 3.0 Pro

• Marketed as Google’s “most secure model yet”.

• Subjected to rigorous internal and external safety evaluations (e.g. UK AI Safety Institute).

• Improved resistance to prompt injection, sycophancy and misuse; deeply integrated with Google’s safety stack.

Claude 4.5 Sonnet

• Anthropic’s signature “Constitutional AI” approach; possibly the strongest public emphasis on safety and alignment.

• High refusal rates for dangerous content; reduced sycophancy and deceptive behaviours.

• Designed to stay safe even over very long, agentic sessions with complex tool use.

Bottom line

• The three proprietary models come with strong, maintained guardrails out-of-the-box, but you can’t inspect or alter their safety layers.

• Apertus gives you full visibility and control, but you must build your own moderation and policy layer if you need strict guardrails.

---

9. Why Apertus Is Still Worth Exploring

Given all this, why should anyone bother with Apertus when GPT-5.1, Gemini 3.0 Pro and Claude 4.5 are clearly stronger on raw capability?

1. Open and Sovereign AI

Apertus embodies a “public infrastructure” model of AI:

• Built by public institutions (EPFL, ETHZ, CSCS) for the public good.

• Can be deployed entirely on national or institutional infrastructure, keeping sensitive data on local soil.

• Avoids dependence on foreign vendors’ policies, pricing and availability.

For governments, universities and regulated sectors (banks, hospitals, schools), this is strategically different from renting a black-box API from overseas.

2. Transparency and Regulatory Compliance

• Full architectural and data transparency makes Apertus auditable.

• Easier to answer questions like: “Was this model trained on copyrighted X?” or “How did we mitigate bias against group Y?”.

• Aligns naturally with emerging AI governance regimes (e.g. EU AI Act) that demand provenance and explainability, not just performance.

3. Multilingual and “Long Tail” Strengths

• Apertus dedicates a large portion of its training to non-English and under-represented languages.

• For many “long-tail” languages and local dialects, it may be stronger out of the box than frontier models that are mostly optimised for English and a few major languages.

• This makes it especially interesting for countries and communities where such languages matter for education and public services.

4. Adaptability and Research Value

• With Apache 2.0, you can:

  • Fine-tune Apertus on your own domain data.

  • Experiment with new alignment techniques.

  • Add modalities (e.g. vision) or integrate into open-source toolchains.

  • Release derivative models.

For researchers, Apertus is a live laboratory for studying large-scale LLM behaviour, bias and safety in a way that closed models simply cannot offer.

5. Cost Benefits at Very Large Scale

• For organisations consuming billions of tokens per month, API costs from GPT-5.1 / Gemini / Claude add up quickly.

• Running Apertus on existing GPU clusters can be cheaper in the long run, especially with:

  • Quantisation,

  • Tiered deployment (8B for simple tasks, 70B for hard ones), and

  • Careful batching.

6. Community and Ecosystem

• Apertus lives within the open-source ecosystem: Hugging Face, community fine-tunes, open research projects.

• Over time, we can expect:

  • Domain-tuned Apertus variants (coding, legal, educational, etc.)

  • Shared prompts, evaluation sets and safety tooling.

That community-driven evolution is already what keeps open models like LLaMA competitive; Apertus extends that story with a much higher level of transparency.

---

10. Conclusion

If your question is simply “Which model is the most capable right now?”, the answer in late 2025 is still:

Gemini 3.0 Pro, GPT-5.1 and Claude 4.5 Sonnet lead on frontier benchmarks and complex tasks.

Apertus does not beat them on raw performance. My own classroom-style coding test (the Trigonometry Unit Circle interactive) made that very clear: Apertus repeatedly failed to complete the full HTML5 simulation, while the proprietary models could.

But that is not the only question that matters.

If your questions are:

• “Who controls this model?”

• “Can we audit the training data?”

• “Can we run it on national infrastructure?”

• “Can we adapt and publish our own versions?”

then Apertus offers something the frontier models cannot:
complete openness, sovereignty, and a blueprint for transparent AI at scale.

In a world where GPT-5.1, Gemini 3.0 Pro and Claude 4.5 Sonnet dominate the performance charts, Apertus stands out as a credible, high-end public alternative. It deliberately trades a few percentage points of benchmark score for “sunlight” in every layer of the stack – from data to weights to training code.

As AI regulation, digital sovereignty and educational use cases mature, that trade-off is looking less like a weakness and more like a necessary second pillar alongside frontier proprietary models.