The 2026 AI Model Field Guide

Why no single model wins

If you only read one paragraph: there is no best AI model in 2026. There are seven or eight very good models, each with a real specialty and a real failure mode. The people doing serious work do not pick one. They route. We route. Our internal tooling sends search to one model, drafting to another, deep research to a third, polishing to a fourth, image generation to a fifth. The output looks like one assistant. Underneath it is a small consortium.

The rest of this guide walks through every major model we have deployed or evaluated, scored on the things that actually matter to a working integrator: hallucination rate, cost per token, data residency, local availability, and what each one is genuinely best at.

The frontier closed-source models

Anthropic Claude (Sonnet 4.5, Opus 4.5, Opus 4.6)

Claude is the model we trust most for writing, code review, and long documents. Hallucination rates on the Vectara leaderboard sit between 4.4% and 4.9% across the 4.x family, which is not the lowest score but is paired with the most useful refusal behavior of any frontier model. Claude says "I do not know" more often than its peers, and that is a feature, not a bug, when the alternative is fabrication.

Anthropic disclosed in November 2025 that a Chinese state-aligned actor used Claude Code in a 30-target espionage campaign before Anthropic disrupted it. We mention this not to dunk on Anthropic but because they actually published the postmortem. Most labs do not. Best for: long-form writing, code, anything legal-adjacent, anything where "I don’t know" is a better answer than a confident wrong one.

OpenAI GPT (GPT-5, GPT-5.2, GPT-5.5)

GPT-5 family is the broadest-capability frontier model. Image generation in GPT-5 is genuinely current-generation. Voice mode is excellent. Raw creative output is excellent. The problem is overconfidence. The AA-Omniscience benchmark recorded GPT-5.5 hallucinating 86% of the time on questions where it did not know the answer. The Vectara summarization score is around 1.4%, which looks great in isolation, but the omniscience number is the one that matters when you ask GPT a factual question about a domain it has not memorized.

OpenAI is also the most legally exposed model maker in the world right now. The OpenAI v Musk trial is set for Oakland federal court in April 2026; Musk is seeking $150B and Sam Altman is on the witness list. The New York Times v OpenAI MDL (25-md-03143) continues to surface training-data regurgitation. None of this stops us from using GPT for the things it is best at; all of it informs how much enterprise data we let touch it.

Google Gemini (3, 3 Pro, 3.1)

Gemini 3 Pro has the lowest summarization hallucination rate on the Vectara leaderboard at the time of writing, and Gemini 2.0 Flash held the record before that at 0.7%. Gemini also has the largest context window in production, the strongest multimodal performance, and the deepest research mode. If we have a 200-page legal package and we need a faithful summary, Gemini gets the job. The cost is data: Google trains on consumer Gemini inputs unless you are on Workspace or Vertex AI with the right toggles set.

xAI Grok (4, 4.1, 4.1 Fast)

Grok is fast and cheap. We have used it for high-throughput, low-stakes work. We do not use it for client-facing output. Independent benchmarks have shown Grok 4.1 Fast hitting hallucination rates above 70% on certain summarization tasks and consistent sycophancy issues across the 4.x line. The model improves quickly. Our position improves with it. As of May 2026 it is on the bench.

The open-source / weights-available models

Meta Llama (4 Maverick, 5)

Llama is the model we deploy locally for clients who want AI without cloud dependency. Llama 4 Maverick runs on a Mac Studio M3 Ultra, a Framework Desktop, or an NVIDIA RTX 5090 box. Hallucination on Vectara sits around 4.6%. The model weights are downloadable. Meta’s license is permissive enough for most commercial uses. For NDAA-aligned government deployments, Llama plus a local stack is often the right answer.

Mistral (Magistral, Mixtral, Le Chat)

Mistral is the European sovereign-AI option. French team, GDPR-aligned, weights available, runs on-prem. We have deployed Mistral for clients who specifically want to keep European data jurisdiction. Capability is one tier below Llama on most public benchmarks, but the deployment story is clean.

DeepSeek V3 / R1

DeepSeek V3 is genuinely impressive at the price point. The R1 reasoning variant hallucinates between 11% and 14% on Vectara, which is not where you want to be for client work. More importantly, the model is Chinese and the data residency question for U.S. enterprise clients is unresolved. We do not deploy DeepSeek for clients with regulated data. We do evaluate it for internal experiments where the input is not sensitive.

Qwen (Alibaba)

Qwen is strong at multilingual work and competitive on coding. Same data residency caveat as DeepSeek. Same use-case ceiling for our client base.

Microsoft Phi (3, 4)

Phi is the small-model story. Phi-3-mini fits on a phone. Phi-4 fits on a Raspberry Pi cluster. We use Phi for edge devices and intent classification in voice pipelines where we do not need a frontier model to interpret "turn off the lights."

Specialty, voice, and small models

Apple Intelligence

Apple Intelligence is the most private mainstream AI. On-device on M-series and A17 Pro and later, Private Cloud Compute for the few queries that need a server, no training on user data. The capability ceiling is real: Apple Intelligence is roughly two model generations behind GPT-5 or Claude Opus on most tasks. For a luxury homeowner who values privacy over raw capability, it is genuinely the right choice for personal use. For a Crestron-grade automation system, it is a complement, not the engine.

MiniMax, Kling, Veo, Sora

Video generation models. We use Veo 3.1 for visual mockups when a client wants to "see" a room before lighting design. We do not deploy any of these for production marketing without a human editor on top.

Whisper, Voxtral, Parakeet

Speech to text. Whisper is OpenAI but the model is open-weight and runs locally; we use it for offline voice transcription. Voxtral (Mistral) is faster on CPU. Parakeet (NVIDIA) is the most accurate on noisy environments. Pick based on hardware.

The retrieval and answer engines

Perplexity Sonar and Sonar Pro

Perplexity is the answer engine we use most. The Sonar API is what we hit programmatically. Perplexity has the highest citation rate of any major answer engine and the broadest source diversity (1,430 unique news sources cited per the company’s own data, vs Google’s 881 and OpenAI’s 707). Sonar scores 0.86 F-score on SimpleQA.

The honest caveat: a Columbia Journalism Review audit found Perplexity answers contained errors 37% of the time despite the citations. Citations are not accuracy. We treat Perplexity as a research head-start, not a verdict. Full detailed look at our Perplexity page.

Perplexity Comet and Perplexity Computer

Comet is the AI browser that browses for you. Perplexity Computer is the multi-agent system. Both deserve their own treatment, which they get on the Perplexity page.

Hallucination leaderboard, May 2026

Compiled from the Vectara public leaderboard and AA-Omniscience benchmark coverage. Lower is better. Numbers move; check sources for current values.

What each lab does with your inputs by tier

The training-on-your-data question deserves more detail than a single bullet. The honest summary as of May 2026:

• OpenAI free / ChatGPT free: may train on your conversations unless you opt out in settings. Opt-out is buried under Data Controls. Most users have not opted out.
• OpenAI Plus / Team / Enterprise: Plus may still allow improvement-use unless you opt out; Team and Enterprise are contractually no-train by default.
• OpenAI API: default no-train since March 2023. Verify with current terms.
• Anthropic Claude consumer: by stated policy, conversations are not used for training by default. Anthropic also does not retain conversations in long-term storage past their stated window.
• Anthropic API: contractually no-train.
• Google Gemini consumer: may train, with retention windows you can adjust in My Activity. Default is on.
• Google Workspace and Vertex AI: contractually no-train when configured properly.
• xAI Grok via X: trains on tweets you post. Standalone Grok app is more conservative but still less private than peers.
• Meta Llama running locally: by definition, no training on your data, because the data never leaves your network.
• Perplexity Pro / Max / Enterprise: stated no-train policy. Free tier may use queries for product improvement.
• Apple Intelligence: stated no-train, on-device when possible, Private Cloud Compute when not. Cleanest privacy story in the industry.

The terms change. Read them when you renew. Read them when you onboard a client. We re-verify our enterprise contracts annually.

The router pattern

Here is how a real Restrepo workflow runs. Client asks: "Find me three Crestron Home dealers in Litchfield County, Connecticut, and draft outreach emails." Old way: an associate spends an hour. New way:

1. Perplexity Sonar searches and returns three dealers with verified addresses and phone numbers, with citations.
2. Claude Sonnet 4.5 drafts three personalized outreach emails based on the dealer profiles.
3. GPT-5 polishes for tone and brevity.
4. A human (us) verifies, signs, and sends. Always.

Total elapsed time: under two minutes. Cost in API tokens: under twenty cents. Quality: better than the associate, because no single model owns the whole pipeline and each one is doing what it is best at.

Cost reality

"ChatGPT is free" is the most expensive sentence in your stack. The free tier may train on your data, has rate limits that break automations, and changes terms unilaterally. Real deployment runs on API tokens at roughly the following 2026 prices:

• Claude Sonnet 4.5: $3 per million input tokens, $15 per million output
• GPT-5: $1.25 per million input, $10 per million output
• Gemini 3 Pro: $1.25 per million input, $10 per million output
• Llama 4 Maverick (local): the cost of your hardware, amortized; effectively zero per token after that

For a single homeowner running a router-pattern assistant, $30 to $80 per month covers most personal usage. For a small business running it on every client interaction, $200 to $1,000 per month is more typical. For a luxury hospitality property, plan for $2,000 to $5,000 per month if AI is doing concierge work end-to-end.

Data residency, training, and what each company does with your inputs

• OpenAI: Free and Plus tiers may train on your inputs. API and Enterprise do not, by default. Opt-out requires settings change.
• Anthropic: Consumer Claude does not train on conversations by default. API does not.
• Google: Consumer Gemini may train. Workspace and Vertex AI do not, with the right toggles.
• xAI: Grok on X trains on tweets. Standalone usage policy varies.
• Meta Llama: Run it locally. The question goes away.
• Perplexity: Pro and Enterprise tiers do not train on your queries. Free tier may.
• Apple Intelligence: Does not train on user data. Period. This is the cleanest privacy story in the industry.

The lock-in question by lab

Picking a model is also picking a lab to depend on. The labs differ in ways that matter on a five-year horizon:

• OpenAI: the broadest ecosystem, the most product churn, the most legal exposure currently. Track record on deprecating older models is real; the GPT-3 and GPT-3.5 era APIs were sunset on relatively short notice.
• Anthropic: smaller surface area, more conservative product cadence, the strongest stated safety posture. Less churn, less variety.
• Google: the deepest distribution, the most history of sunsetting products. We deploy Gemini happily; we plan for the day a feature gets renamed or moved.
• Meta: open weights make lock-in structurally lower. Even if Meta changes direction, the weights you downloaded keep working.
• xAI: newest, fastest moving, most volatile product line. Capability swings are larger than peers.
• Mistral: European, sovereign, smaller. The right pick for clients with EU data residency requirements.
• Apple: integrated into hardware you already own. Lock-in is high (you cannot run Apple Intelligence on a Windows PC), but the privacy posture is the cleanest.

What we choose for clients (and why the answer is "it depends")

We do not pick one model and force it on every client. We pick a router. For a private residence with sensitive personal data, the default is local Llama plus Whisper plus Home Assistant, with optional cloud reach-out for non-sensitive tasks. For a hospitality property running guest concierge, the default is Claude or GPT for drafting, Perplexity for search, and a hard rule that no autonomous action happens without human confirmation. For a commercial board room, the default is whatever the corporate IT policy already approves, integrated through Crestron Sightline AI or Microsoft Teams Premium. The right model is downstream of the right architecture. We design the architecture first.

One more honest note: the right answer is the one we can support. We will not deploy a model we have not personally tested at length on the use case the client has in mind. We have a small library of internal evaluations we run on every new model release, and we update our recommendations when the evaluations move. Recommendations are perishable in this industry. Ours have an expiration date and we will tell you what it is.

Related

• Perplexity, Comet, and Perplexity Computer
• Build It Yourself: Local AI
• The AI Reality Check