DECAID Academy · October 2026

LLM Model Comparison.
Which model fits your use case?

Over 30 AI models, four tools, one goal: helping you find the right model for your task. As of October 2026.

01 · Recommendation Assistant

Find your matching model.

We help you not with tool selection but with model selection. Which model is best for your use case? Four short questions, plus one optional question about the tool. At the end you get a concrete model recommendation.

How the Assistant Works

Answer four questions about your use case: what you want to do, how deep it should go, how sensitive the data is, and how frequently you use it. Optionally you can specify a tool if you already know you are bound to a particular platform. If not, you get the tool-agnostic best recommendation. If you specify a tool, you get both: the objectively best model and the best model within your tool.

Question 1 of 5
✓
Our Recommendation

Model Recommendation

02 · The Four Tools

Tool Comparison.

Four platforms, four philosophies. Here you see what each is designed for, what it costs, and where the limits are. As of October 2026.

ChatGPT (OpenAI)

Broad feature stack, frontier models first. The original.

Free Go $8 Plus $20 Pro 100 $100 Pro 200 $200 Pro 500 $500 Business $25 Business Premium $125 Enterprise Custom
Strong at:
  • GPT-6 Astra (September 3, 2026) still leads ARC-AGI-2 (95.0%) and GPQA Diamond (96.0%), native computer use
  • ChatGPT Work (July 2026): long-horizon agent that builds docs, sheets, decks, and web apps for hours on end
  • Voice Mode (GPT-Live) full-duplex since July 2026, natural interruption, live translation
  • Native image generation with GPT Image 2.5 (since September 8), leads the text-to-image arena
  • Code Interpreter (Python sandbox) for data analysis
  • GPT-6.1 Sol (DevDay, September 29): close to Astra level for $2 / $10, a fifth of the Astra price
Limits:
  • US default hosting (EU residency only for Enterprise, Edu, and API, not in the Business tier)
  • Regional processing in the API has carried a ten percent surcharge on all newer models since March 2026
  • Training on user data enabled by default in Free, Go, Plus, and Pro
  • Ads in Free and Go in more than 60 countries, in DACH since August 24. Paid plans stay ad-free
  • From October 30, Pro 200 only gets half the Work and Codex allowance (10x instead of 20x Plus)
  • Astra in chat only from Pro, Business, and Enterprise, GPT-6.1 Sol only in Work and Codex. The chat default since October 7 is GPT-6 Sol (Plus and up) or GPT-6 Luna (Free, Go)
  • OpenAI models only, no multi-provider
Best choice for: Frontier models directly, agents, voice and image generation combined. Those who can live with US hosting.

Microsoft Copilot

Deepest Office integration, Graph grounding across emails and files. Known simply as "Microsoft Copilot" since August 2026.

Chat free M365 Premium $19.99 Microsoft Copilot $30 Copilot Business $18 (list $21) Copilot Credits GitHub $10-$100
Strong at:
  • Word, Excel, PowerPoint, Outlook, Teams with Microsoft Graph grounding
  • EU compliance tooling (Purview, Sensitivity Labels, Customer Lockbox)
  • Excel Agent Mode with Python sandbox on tenant data, Python also available while editing since August 25
  • Claude Opus 5.5 and GPT-6 Sol in Word, Excel, PowerPoint, Chat, and Cowork since September 22, GPT-6.1 Sol and Sonnet 5.5 there too since October 6 (the OpenAI and Anthropic subprocessors must be enabled)
  • GitHub Copilot with the widest model choice: GPT-6 Astra and 6.1 Sol, Claude Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5, Gemini 4 Argon (Pro+ and up) and 3.8 Flash, Grok 4.7, MAI
  • In-house MAI stack growing fast: MAI-Code-1.1-Flash, MAI-Image-2.6, MAI-Transcribe-2
Limits:
  • Flex Routing on by default for all tenants created after March 25, 2026, older tenants should check their setting. Can be disabled, otherwise inference may move to the US, Canada, or Australia
  • Anthropic models are entirely excluded from the EU Data Boundary and disabled by default in the EU, EFTA, and UK
  • Grok only in the Frontier program since September 12 and explicitly not in the EU, EFTA, and UK
  • Copilot Business $18 is a promotional price until December 31, 2026, list price $21
  • New billing model since September 25: Cowork, Code, Autopilot as well as Astra and Fable run on Copilot Credits, new Business purchases are usage-based by default from October 19
  • GitHub Copilot has billed by AI Credits since June 2026, Business and Enterprise received fewer included credits as of September 1
Best choice for: Organizations already deeply embedded in the Microsoft 365 ecosystem who want Office-added value.

Gemini (Google)

Massive context window (1M tokens), strong multimodality, included in Workspace.

Free AI Plus $4.99 AI Pro $19.99 AI Ultra $99.99/$199.99 Workspace from $14
Strong at:
  • 1M token context at 99.7% recall, clearly more than ChatGPT
  • Native multimodality: video, audio, and image on equal footing
  • Gemini Omni Flash and Veo 3.1 (video), Nano Banana 2 and Pro (image), Lyria 3.5 (music) first-party
  • Deep Research and Gemini Notebook (formerly NotebookLM) currently leading for research
  • Gemini 4 Argon generally available since October 13 (in the app from AI Pro), first place on Arena Text, also in the EU multi-region
  • Flash models at $0.75 / $3.75 are currently the cheapest in the frontier class
  • Included for free in Workspace from Business Starter upward
Limits:
  • A single region in Germany is only available with Gemini 3.5 Flash, which has been deprecated in the Gemini API since October 8 (requests are routed to 3.6 Flash) and is still listed on Vertex for now
  • Gemini 2.5 Pro will be shut down on Vertex AI on October 20, 2026, in the Gemini API only existing users still get access
  • The Flash introductory prices expire on December 31, 2026, after which they double to $1.50 / $7.50
  • The Argon introductory price of $2 / $10 only applies until November 30, from December 1 Argon costs $4 / $20
  • Gems are being replaced by Skills (Workspace since October 5, app since October 13), Business and Enterprise reportedly can no longer use Gems from March 2027
  • Enterprise market share only approx. 18%
Best choice for: Google Workspace customers, long documents, multimodal use cases, video and audio generation.

Langdock

EU platform from Berlin. Multi-provider under one interface.

Business 25 EUR Business Max 99 EUR Workflows from 539 EUR Governance 3.50 EUR from 2027 Enterprise Custom
Strong at:
  • Claude, GPT, Gemini, Mistral, Llama, DeepSeek in one interface (over 40 models in the marketplace)
  • EU hosting on Azure, ISO 27001 + SOC 2 Type II, professional-secrecy add-on available
  • German contract, German invoice, German support
  • Integrations for M365, Google, Slack, Salesforce, and MCP, plus add-ins for Excel and Outlook
  • Workflows and agents more capable than in ChatGPT Business or Copilot
  • New models fast and EU-hosted: Claude Opus 5.5 and Sonnet 5.5 in September, Claude Haiku 5.5 on October 8, Gemini 4 Argon on October 13
  • Native desktop app (September 10) and image generation with GPT-Image 2.5 (September 11), EU-hosted
  • Composable Workflows since September 18: workflows call other workflows, up to five levels deep
Limits:
  • No Voice Mode on web or desktop, voice input only in the mobile app
  • No video generation, image generation runs through model partners rather than a proprietary model
  • Not every model is EU-hosted: DeepSeek and Llama 4 Maverick run globally, GPT-6.1 Sol is listed as "Global" on the models page although the changelog announced EU hosting
  • Governance features (agent review, compliance rules) paid from January 1, 2027 (3.50 EUR/user, 2.80 EUR billed annually)
  • No ISO 42001 certification, unlike the three US providers
Best choice for: DACH SMBs with compliance requirements, multi-model needs, and anyone who wants to use Claude without a separate subscription.
03 · Model Landscape

Which model comes from which stable?

The most important models per provider, as of October 2026. With context window, API price (USD per 1M tokens in/out), and sweet spot. "Available in" shows where to get each model.

OpenAI

San Francisco. Market leader in frontier models, birthplace of GPT.
ChatGPT Copilot (via Azure) Langdock API
Model
Type
Context
Price $/1M
Sweet Spot
GPT-6 AstraSeptember 3, 2026Flagship · 1.05M · $10 / $50
Flagship Reasoning
1.05M
$10 / $50
OpenAI's top model with 53 points in the Artificial Analysis Intelligence Index, behind Claude Opus 5.5 and Sonnet 5.5 since late September. Still leads ARC-AGI-2 (95.0%), GPQA Diamond (96.0%), and FrontierMath v2 Tier 4 (97.6%). Native computer use. Caution: prompts above 272K tokens cost double on input and 1.5x on output. In ChatGPT chat (as GPT-6 Pro) from Pro, Business, and Enterprise, on Plus only in Work and Codex.
GPT-6.1 SolSeptember 29, 2026Reasoning · 1.05M · $2 / $10
Reasoning
1.05M
$2 / $10
Unveiled at DevDay: close to Astra level (on par on DeepSWE, OSWorld 2.0 71.4% vs. 73.5%) at a fifth of the price. Replaces GPT-6 Sol, which is only a week old (September 22, same price) and on several benchmarks even scored below GPT-5.6 Sol. In ChatGPT only in Work and Codex, chat runs on GPT-6 Sol. Since October 8 there is an ultrafast mode, also with EU residency. Above 272K tokens the input price doubles.
GPT-6 LunaSeptember 22, 2026Fast + affordable · 1.05M · $0.10 / $0.50
Fast + affordable
1.05M
$0.10 / $0.50
Half the price of GPT-5.6 Luna with a larger context. High-volume tasks, classification, routing. Chat default for Free and Go since October 8. Replacement for GPT-4o, 4o mini, 4.1, and 4.1 mini, which will be shut down in the API on January 14, 2027. One of the cheapest near-frontier models available.
GPT-5.6 Sol / Terra / LunaJuly 9, 2026Chat models · 1M · from $0.20 / $1.20
Predecessor
1M
$4 / $20 and $2 / $12
No longer the ChatGPT chat default since October 7, chat now runs on GPT-6 Sol and Luna. Sol costs $4 / $20 as a promotion until at least November 21, 2026, Terra $2 / $12, Luna $0.20 / $1.20. In the API the GPT-6 siblings are cheaper and the better choice for new projects.
GPT Image 2.5 (Flare / Sunburst)September 8, 2026Image generation · Flare fast, Sunburst precise
Image generation
—
Included in ChatGPT
Leads the text-to-image arena, now with Sunburst (Elo 1,197) just ahead of Flare (1,190), GPT Image 2 in third place. Flare is the default with half the latency, Sunburst is for editing workflows that need tight control. The old gpt-image-1 will be shut down on October 23, 2026, gpt-image-1-mini and gpt-image-1.5 on December 1, 2026.
GPT-Live / GPT-Realtime-2.1July 2026Full-duplex voice · real-time translation
Voice-to-Voice
—
Per audio minute
GPT-Live is the voice experience in ChatGPT: speaks and listens at the same time, natural interruption, live translation. In the API the models are called GPT-Realtime-2.1 and -mini. The old audio family will be shut down on January 20, 2027, the Whisper and 4o-Transcribe models on February 26, 2027.

Anthropic (Claude)

San Francisco. Strong at reasoning, long documents, and coding. Primary path for DACH users: Langdock.
Langdock claude.ai directly GitHub Copilot Microsoft Copilot (opt-in)
Model
Type
Context
Price $/1M
Sweet Spot
Claude Fable 5.1September 1, 2026Top tier · 1M · $10 / $50
Top tier
1M
$10 / $50
Anthropic's most expensive model, Terminal-Bench 4.0 55.8%, ARC-AGI-2 90.0%. Overtaken in the Intelligence Index since September 22 by Opus 5.5, which costs less than half. Cache reads cut by 75%. Retention: the standard is 30 days for safety monitoring, until Enterprise Frontier Safeguards (monitoring data in your own cloud storage) are available, eligible customers can use Fable 5.1 with zero data retention.
Claude Opus 5.5September 22, 2026Top reasoning + coding · 1M · $4 / $20
Top reasoning + coding
1M
$4 / $20
The new default for companies and since September 22 first place in the Artificial Analysis Intelligence Index (58 points), ahead of GPT-6 Astra and Fable 5.1. Terminal-Bench 4.0 66.4%, ARC-AGI-2 93.3%, OSWorld 2.1 81.8%. List price 20% below Opus 5, around 40% cheaper per task according to Anthropic. Available with zero data retention. Adaptive thinking cannot be switched off. Strategy, audit, legal, coding, creative writing.
Claude Sonnet 5.5September 28, 2026All-rounder · 1M · $2 / $10
All-rounder
1M
$2 / $10
Daily driver at an unchanged price and, with 56 points in the Intelligence Index, in second place, 18 points above Sonnet 5. Terminal-Bench 4.0 70.6%, OSWorld 2.1 80.1%, more than 30% faster output, zero data retention available. Caution: at the highest effort level it uses more output tokens than any other model Artificial Analysis has measured. Editing, emails, sales coaching, agentic coding.
Claude Haiku 5.5October 7, 2026Fast + affordable · 1M · $0.10 / $0.50
Fast + affordable
1M
$0.10 / $0.50
Around 75% cheaper than Haiku 4.5 and a generational leap: OSWorld 2.1 72.4% (Haiku 4.5: 15.7%), GDPval-AA 1,620 Elo. The price applies up to 100K prompt tokens, above that $0.50 / $2.50. Caution: the new tokenizer counts around 30% more tokens, temperature, top_p, and prefill return an error. Classification, routing, subagents, mass processing. Haiku 4.5 ($1 / $5, 200K) stays active for now without a shutdown date.
Claude Opus 5 / Sonnet 5July and June 2026Legacy · still available
Legacy
1M
$5 / $25 and $2 / $10
Legacy since the 5.5 releases, shutdown no earlier than June 2027 (Sonnet 5) or July 2027 (Opus 5). Opus 5 remains relevant for anyone who does not want to migrate prompts right away: Opus 5.5 has breaking changes (thinking cannot be disabled, forced tool use returns an error). Fable 5 and Opus 4.8 are legacy too. Sonnet 4.5 will be shut down on November 30, 2026.

Google (Gemini, Veo, Nano Banana, Lyria)

Mountain View. Largest context window, native multimodality, complete creative stack.
Gemini App Workspace Langdock Vertex AI
Model
Type
Context
Price $/1M
Sweet Spot
Gemini 4 ArgonOctober 13, 2026Flagship Reasoning · 1M · $2 / $10
Flagship Reasoning
1M
$2 / $10
Generally available since October 13 and Google's new top model: first place on Arena Text (1,533), 53 points in the Intelligence Index, on par with GPT-6 Astra. Also runs in the EU multi-region on Vertex, in the Gemini app from AI Pro, in Langdock, and in GitHub Copilot (Pro+ and up). Caution: $2 / $10 is an introductory price until November 30, 2026, from December 1 $4 / $20. Coding, long documents, multi-step workflows.
Gemini 3.8 FlashSeptember 2, 2026Flash flagship · 1M · $0.75 / $3.75
Multimodal Default
1M
$0.75 / $3.75
Current top Flash model and, since the Argon launch, Google's affordable default. Second place on Terminal-Bench 2.1, GPQA Diamond 95.3%. Replaces 3.7 Flash (August 13) and 3.6 Flash (July 21). Caution: $0.75 / $3.75 is an introductory price until December 31, 2026, then $1.50 / $7.50. According to the Vertex docs it now also runs in the EU multi-region, there is no single region in Germany.
Gemini 3.1 ProFebruary 2026 (preview)Reasoning Multimodal · 1M · $2-4 / $12-18
Reasoning Multimodal
1M
$2-4 / $12-18
Still the only Pro model of the 3 series and still in preview. Long documents, charts (84.2% CharXiv), Deep Research, ARC-AGI-2 77.1%. Tiered pricing: $2 / $12 up to 200K, above that $4 / $18. Its successor Gemini 4 Argon has been generally available since October 13, 3.1 Pro remains interesting mainly for chart analysis.
Gemini 3.5 FlashMay 19, 2026Single region Germany · 1M · $1.50 / $9
EU-capable all-rounder
1M
$1.50 / $9
Still important for DACH: the most recent Gemini with a single region in Germany (europe-west3) including ML processing. Terminal-Bench 2.1 76.2%, computer use, Live Translate. More expensive than the newer Flash models, but the first choice when data must stay in Germany. Deprecated in the Gemini API since October 8 (requests are routed to 3.6 Flash), still listed in Europe on Vertex for now. Gemini 2.5 Pro, Flash, and Flash-Lite will be shut down on Vertex on October 20, 2026.
Gemini Omni Flash / Veo 3.12025/26Video with native audio
Video generation
—
In AI Ultra / API
Gemini Omni Flash is Google's strongest video model and can extend and interpolate scenes, but has dropped to eighth place in the text-to-video arena. Alibaba's Wan 3.0, Utopai X, and ByteDance Seedance 2.5 lead, followed by MiniMax H3 and FLUX 3. Veo 3.1 and Veo 3.1 Lite deliver lip-sync and 4K upscaling. Veo 2 and Veo 3 were shut down on June 30, 2026.
Nano Banana 2.1 / Pro, Lyria 3.52026Image and music generation
Image + music
—
From approx. $0.03 per image
Nano Banana 2.1 (GA since October 6) is the volume image model: around 3.4 cents per 1K image, up to 4K, fourth place in the text-to-image arena. Nano Banana Pro remains the first choice for text in images and infographics. Lyria 3.5 has been GA since September 3 and produces songs up to three minutes with vocals for $0.08. The Imagen 4 endpoints were shut down on August 17, 2026.

Microsoft (MAI, Phi)

Redmond. In-house model stack since Build 2026 (June 2): reasoning, coding, image, voice, transcription. MAI models replace tens of thousands of prompts per week in Excel and Outlook and provide the default coding model in GitHub Copilot. The heavy reasoning tasks still run on licensed models from OpenAI and Anthropic.
M365 Copilot (Subtasks) GitHub Copilot Copilot+ PCs Microsoft Foundry
Model
Type
Context
Price $/1M
Sweet Spot
MAI-Thinking-1June 2026Reasoning (MoE, 35B active) · 256K context
Reasoning flagship
256K
Mid-weight (n/a)
First in-house reasoning model, in public preview in Microsoft Foundry since August 12, 2026. Preferred over Claude Sonnet 5 in blind tests, trained without distillation from third-party models.
MAI-Code-1.1-FlashAugust 11, 2026Coding/agentic · with vision · efficient
Coding efficient
—
Cheap (n/a)
Default coding model in GitHub Copilot, now with native image processing and around 73% cheaper than its predecessor. The old MAI-Code-1-Flash was deprecated on September 10, 2026.
MAI-Image-2.6 / -FlashSeptember 4, 2026Image generation · Copilot in PowerPoint
Image generation
—
n/a
Public preview in Foundry. Microsoft reported second place in the text-to-image arena at launch, in the ongoing Artificial Analysis ranking 2.6 now sits in sixth place (Elo 1,164) behind Nano Banana 2.1 and Grok Imagine Image 2.0, but well ahead of its predecessor 2.5. Clearly improved text rendering, the Flash variant is tuned for throughput.
MAI-Voice-2 / 2.1June 2026TTS · 15 languages · voice cloning
TTS
—
n/a
Natural speech synthesis in 15 languages, voice adaptation from a short sample. Powers Copilot Voice and the podcast feature. Since October 1, MAI-Voice-2.1 and 2.1-Flash have been available in Foundry.
MAI-Transcribe-2September 3, 2026STT · 60 languages · diarization
STT
—
$0.10 / hr
Public preview. Speech-to-text across 60 languages with speaker separation, word timestamps, and code-switching, at a promotional price of $0.10 per audio hour until year end. Replaces MAI-Transcribe-1, which was deprecated on August 20, 2026. Since October 1 also available as a streaming variant for real time, $0.54 per audio hour until year end.
Phi-4 / Phi-4-multimodal2025/26Small + on-device · 14B parameters
Small / on-device
128K
Self-host free
Runs on-device on Copilot+ PCs (NPU). Background tasks in Office: auto-summaries, slide layouts. Microsoft still lists the Phi-4 family as the current generation.

Mistral (EU vendor, France)

Paris. Most important EU provider. Open-weight strategy, multilingual, strong in German.
Langdock Le Chat Azure AI Foundry Self-Host
Model
Type
Context
Price $/1M
Sweet Spot
Mistral Large 4October 6, 2026 (preview)MoE multimodal · approx. 1T total / 52B active · $1.36 / $4.18
MoE Multimodal (Preview)
up to 1M
$1.36 / $4.18
In the API as a public preview since October 6, Mistral's new top model with around one trillion parameters. Up to 100 images per request, 38 points in the Intelligence Index and thus clearly behind the US frontier. 50% launch discount until around October 20, open weights are announced for the end of October. Mistral states a 1M context, Artificial Analysis 512K. Large 3 remains regularly available.
Mistral Large 3December 2025MoE multimodal · 41B active / 675B total · 256K
MoE multimodal
256K
$0.50 / $1.50
EU sovereignty, very strong in German, open-weight for on-prem. The price now sits at $0.50 / $1.50, a quarter of the original list price. Since August 11, 2026 with a selectable EU or US endpoint (regional inference, surcharge approx. 10%).
Mistral Medium 3.5April 30, 2026Generalist multimodal · 256K · $1.50 / $7.50
Generalist multimodal
256K
$1.50 / $7.50
Strong on agentic and coding (77.6% SWE-bench Verified), open-weight under a modified MIT license. Default model in Vibe, the agent interface that has replaced Le Chat since May 28, 2026.
Mistral Small 4March 16, 2026Hybrid · MoE 119B / 6B active · 262K
Efficient all-rounder
262K
$0.15 / $0.60
The price-performance highlight in the Mistral stack: combines instruct, reasoning, and agentic coding in one model, replacing the Magistral, Pixtral, and Devstral lines. Apache 2.0, so freely self-hostable. Ideal for business volume with a GDPR argument.
Codestral 25.08August 2025Coding specialist · 256K · $0.30 / $0.90
Coding specialist
256K
$0.30 / $0.90
Coding model with fill-in-the-middle, for IDE integrations. Self-host possible. Mistral Medium 3 and 3.1 were shut down for good on August 31, 2026.

Meta (Llama)

Menlo Park. Self-hosting standard for critical infrastructure, public sector, banks. The Llama line has been dormant since April 2025, Meta's models now ship under the Muse brand: Muse Spark is proprietary, Muse Glimmer has been open again since August 2026 and even under Apache 2.0.
Langdock Self-Host AWS Bedrock Hugging Face
Model
Type
Context
Price $/1M
Sweet Spot
Muse WonderOctober 12, 2026Frontier · 1M · $4 / $20 · closed-weight
Frontier multimodal
1M
$4 / $20
Meta's first true frontier model, input text, image, video, audio, and PDF. 52 points in the Intelligence Index, the stronger Wonder Full variant 56, on par with Claude Sonnet 5.5, fifth place on LMArena Text. Wonder Full is only available in the Muse Code subscription, not via the API. No cheap contributor tier as with Spark.
Muse Spark 1.3September 2, 2026Frontier · 1M · $1.25 / $4.25 · closed-weight
Frontier agentic
1M
$1.25 / $4.25
Meta's top model until the Wonder launch, trained for long-horizon agents. Terminal-Bench 2.1 88.8%, DeepSWE 1.1 75.4%. The weights are closed. Caution when comparing prices: the often-cited $0.10 / $0.20 rate is the contributor tier, where Meta may use prompts for training and the rate limit drops to one thirtieth.
Muse GlimmerAugust 10, 2026Open-weight agentic · 30B dense · 131K
Open-weight
131K
Self-host free
Meta's first open model since Llama 4 and the first ever under Apache 2.0. SWE-bench Verified 76.0%, multimodal, runs in 4-bit under 20 GB on a consumer GPU or a Mac. That makes it the realistic self-host option from the Meta stack.
Llama 4 MaverickApril 2025MoE multimodal · 17B active / 400B total · 1M context
MoE multimodal
1M
Self-host free
Open-weight frontier tier of its time, 128 experts. Still available and not deprecated, but technically superseded.
Llama 4 ScoutApril 2025RAG sniper · 10M context · on a single H100
Extreme context
10M
Self-host free
10M-token context, runs on a single H100. For RAG use cases with huge knowledge bases, still unmatched on context to this day.

Other relevant models

xAI, DeepSeek, Moonshot, Aleph Alpha, Suno. Specialists and sovereignty options.
Model
Type
Context
Price $/1M
Sweet Spot
xAI Grok 4.7September 21, 2026Reasoning/coding · 500K · $2 / $6
Reasoning + coding
500K
$2 / $6
Current xAI flagship, replaces Grok 4.6 at the same price. DeepSWE 1.1 71.0%, but with 46 points in the Intelligence Index clearly behind Claude, GPT-6, and Gemini 4 Argon. Above 200K prompt tokens the rates double for the entire request. US hosting, check carefully for DACH compliance.
DeepSeek V4.1 FlashSeptember 10, 2026Reasoning MoE · 1M context · peak/off-peak pricing
Reasoning MoE
1M
from $0.15 / $0.60
Extremely cheap: off-peak $0.15 / $0.60, peak $0.30 / $1.20. According to recent reports V4 Pro is not being phased out after all: DeepSeek has reversed the redirect to V4.1 Flash, V4 Pro still costs $1.32 / $3.96 (peak) or $0.66 / $1.98 (off-peak). For DACH compliance: check the China hop, choose EU hosting.
Moonshot Kimi K3July 16, 2026Open weights · 2.8T parameters · 1M context
Open-weight frontier
1M
$3 / $15
Largest open-weight model on the market, downloadable under a modified MIT license since July 27. That makes it a genuine sovereignty option for anyone who wants to self-host frontier quality. Also selectable in GitHub Copilot.
Aleph Alpha Pharia (now Cohere)2024/25Made in Germany · sovereignty platform
Sovereignty platform
—
Custom
Heidelberg. Aleph Alpha and Cohere signed the definitive merger agreement on September 16, 2026 (headquarters in Berlin and Toronto, closing after regulatory approval), a combined Command-Pharia 1 is announced for Q4 2026. New since October 3: Kolibri 1, an open MoE model (78B, around 3.5B active) under Apache 2.0 with a high share of German training data. Still the go-to when "Made in Germany" is a legal and political requirement (defense, interior ministries).
Suno v6September 9, 2026Music generation with vocals · up to 8 minutes
Music generation
—
Subscription
Full songs with vocals, voice cloning, tracks up to eight minutes. Crucial for organizations: v6 is the first Suno model built on licensed training data from Warner Music, BMG, and Believe. The predecessor models are being phased out. Since October 1, Suno Speech (beta) combines speech and music in one track.
04 · Legal and Data Privacy

What do you need to watch with each tool?

The compliance picture for all four tools side by side. As of October 2026. EU AI Act: GPAI enforcement and Art. 50 transparency obligations have been in force since August 2, 2026, the high-risk obligations have been pushed to December 2027 by the Digital Omnibus on AI. In Germany, the Bundesnetzagentur has been the competent authority since July 29, 2026.

Criterion
ChatGPT
M365 Copilot
Gemini
Langdock
EU data residency
Partial
Enterprise/Edu/API: yes, since Jan 2026 also in-region GPU inference, in the API with a 10% surcharge. Plus/Pro/Business: no.
With caveat
Standard yes, but Flex Routing on by default for tenants created after March 25, 2026 (can be disabled). Anthropic models entirely outside the EU Data Boundary. New setting for in-country storage (including Germany), deadline December 14, 2026.
Partial
Workspace yes. On Vertex Gemini 4 Argon and the Flash models 3.6 to 3.8 run in the EU multi-region, single region Germany only with Gemini 3.5 Flash (deprecated in the Gemini API).
Yes
EU hosting on Azure, by default. Individual models such as DeepSeek run globally.
DPA available
Yes
From the Business tier upward.
Yes, with caveat
Microsoft Products DPA standard. It does not apply to Anthropic models with data retention, there Anthropic is a separate controller.
Yes
Google Workspace DPA standard.
Yes
German contract, German DPA.
Training on customer data
On by default in Free to Pro
Free, Go, Plus, and Pro train by default. Business/Enterprise: no.
No
Contractually excluded.
Workspace no, consumer yes
Can be disabled manually.
No
Contractually excluded.
EU vendor
No (USA)
No (USA)
No (USA)
Yes (Germany)
ISO 27001 / SOC 2
Yes
Yes
Yes
Yes (both)
EU compliant out of the box
No
Configuration required.
No
Disable Flex Routing.
No
EU Workspace setup required.
Yes
Model sovereignty
OpenAI only
OpenAI, MAI, Anthropic, and Grok
In the EU only after admin opt-in and outside the EU Data Boundary.
Google only
Multi-vendor
ISO 42001 (AI management system)
Yes
Yes
Yes
No
Only ISO 27001 and SOC 2 Type II.

EU AI Act deadlines: Art. 4 (AI literacy, mandatory documented employee training) has applied since February 2, 2025. Since August 2, 2026, GPAI enforcement and Art. 50 transparency are in force: the Commission can enforce GPAI obligations with fines of up to EUR 15M or 3% of global turnover, chatbots must identify themselves as AI, and synthetic content must be marked in a machine-readable way. For Art. 50, a grace period applies until December 2, 2026 for systems placed on the market before August 2, 2026, and for GPAI models from before August 2, 2025 even until August 2, 2027. The Digital Omnibus on AI was published in the Official Journal as Regulation (EU) 2026/1744 on July 24, 2026 and has been in force since July 27, 2026. It postpones the high-risk obligations: standalone high-risk systems (Annex III) to December 2, 2027, high-risk in products (Annex I) to August 2, 2028. New is an Art. 5 ban on AI-generated non-consensual intimate images and CSAM, to be complied with from December 2, 2026. The relief on technical documentation and quality management now also applies to "small mid-caps" with up to 750 employees, but this changes nothing about the GPAI and transparency obligations themselves. In Germany, the AI Market Surveillance and Innovation Promotion Act has designated the Bundesnetzagentur as market surveillance authority and central point of contact since July 29, 2026, with BaFin for the financial sector and the state media authorities for media. First publicly known review case: since September the Commission has been examining why OpenAI did not report a security incident to the AI Office (Art. 55). On October 9, 2026 the Commission also held a special meeting of the Scientific Panel on loss-of-control incidents with frontier models, the AI Office and the panel have prepared questions for the providers concerned.