Seven foundation models now carry the weight of the global enterprise AI stack. This brief ranks each across eleven decision vectors — from raw code generation to a twelve-month strategic prognosis — for leaders who must commit capital, vendors, and architecture before the next compression cycle.
The frontier has fragmented. No single model dominates across capability, cost, and distribution — and that fact is itself the most important strategic signal of 2026. The right architecture for an international group is a portfolio, not a vendor.
Six frontier models now sit within 1.3 points of one another on SWE-Bench Verified. Claude Opus 4.7 leads coding at 87.6%, Gemini 3.1 Pro leads pure reasoning at 94.3% on GPQA Diamond, and Grok 4 leads raw LiveCodeBench. The variance you feel between them is workflow fit, not raw intellect.
DeepSeek V3.2 delivers GPT-5-class reasoning at $0.14 / $0.28 per million tokens — roughly 96% below proprietary equivalents — and the open-weight stack (Llama 4, DeepSeek, Qwen, Gemma) now runs within a single benchmark point of closed frontier on most tasks. Pricing-as-moat is dead.
Microsoft Copilot's 100M+ monthly users, Google's Workspace footprint, and xAI's embedded loop with X give incumbent platforms a strategic advantage independent of model quality. For enterprises, "the best model" increasingly means "the model already in your knowledge surface."
Each model is a bet on a different future — frontier reasoning, multimodal saturation, cost disruption, real-time data, search-grounded research, open-weight infrastructure, or productivity distribution.
Best-in-class coding agent, exceptional instruction-following, deep enterprise trust. The model behind Cursor and Windsurf, and the new reference for long-horizon agentic work.
Native text, image (Nano Banana 2), and video (Veo 3.1). 2M-token context — the largest of any commercial model. Leads GPQA Diamond at 94.3% and ARC-AGI-2 at 77.1%.
Open-weight (MIT), GPT-5-class reasoning, gold-medal IMO/IOI. Cheapest serious frontier model on Earth — and the single most important pricing event in AI since the GPT-4 launch.
Tight integration with the X social graph, native voice, image and video generation via Grok Imagine. Strong on agentic benchmarks and live information; weaker on long-form reasoning.
Search-grounded answers with citations as a first-class output. Comet browser, Deep Research mode, Sonar API. 45M users, ~1B queries per month — the research-vertical winner.
Mixture-of-experts at 400B / 17B active. Scout offers a 10M-token context window — the largest open model on the market. Beats GPT-4o on most benchmarks; free to self-host.
Multi-model orchestration: GPT-5.4 Thinking, GPT-5.3 Instant, plus Anthropic Claude integration ("Critique" and "Council" modes). Sora 2 video. The default AI surface for 100M+ Office users.
Capability scores are calibrated 1–5 against the current frontier (May 2026). Cost reflects per-million-token API pricing for the flagship tier. On desktop, click any column header to re-rank the field.
| Model▼ | Programming▼ | Research▼ | Images▼ | Video▼ | Accuracy▼ | Speed▼ | Cost / 1M tok▼ | Tech▼ | Strategic Position | 12-Month Prognosis |
|---|---|---|---|---|---|---|---|---|---|---|
|
Claude
Opus 4.7
Anthropic · Apr 2026
|
$5 / $25 | Enterprise leader: 40% of LLM spend, $30B run-rate. Powers Cursor, Windsurf, Claude Code. Reference for agentic, regulated, and code-heavy workloads. |
Continued enterprise dominance. Mythos preview surfaces late‑2026 will widen its lead on hard reasoning. Risk: pricing pressure from DeepSeek tier. |
|||||||
|
Gemini
3.1 Pro
Google DeepMind · Feb 2026
|
$2 / $12 | The only fully-multimodal frontier stack: LLM, image, video, audio under one API and one bill. Deep integration with Workspace, Android, YouTube, Vertex AI. |
Multimodal leadership consolidates. Gemini 4 expected H2 2026. Ecosystem distribution becomes its primary moat — coding remains the open flank. |
|||||||
|
Copilot
M365 · 2026
Microsoft · multi-model
|
$20–$30 / seat | Distribution moat at 100M+ MAU. Multi-model routing (OpenAI + Anthropic + MAI/Phi) makes it AI-vendor-agnostic at the workflow layer. Azure OpenAI underneath. |
Wins productivity by default. Agent 365 rollout converts seats into autonomous workers. Real risk: customer fatigue with monthly add-ons. |
|||||||
|
DeepSeek
V3.2 / R1
DeepSeek · China · Dec 2025
|
$0.14 / $0.28 | Cost disruptor. Open-weight under MIT; can be self-hosted to zero marginal cost. China's strategic AI champion; pricing forces every Western lab into compression. |
V4 / R2 narrow the closed-source gap further. Geopolitical risk: U.S. export controls or hosted-API restrictions. Open weights insulate against most adverse outcomes. |
|||||||
|
Grok
4.3
xAI · Apr 2026
|
$1.25 / $2.50 | Only model with native real-time access to the X social graph plus integrated text/image/video/voice. Memphis 1M-GPU build out — most aggressive compute trajectory in the industry. |
Grok 5 (6T params) targets Jan 2026 launch. Niche dominance in real-time and creative — enterprise adoption remains capped by brand and trust positioning. |
|||||||
|
Perplexity
Sonar Pro
Perplexity · 2025–26
|
$3 / $15 incl. search | Search-vertical winner. Comet browser (free), Deep Research, Model Council. 45M users, ~1B queries/month. Disrupts Google search at the AI-answer layer. |
Agentic Pro Search and Comet adoption deepen the moat. Risk: Google's AI Overviews close the gap as Gemini matures. M&A target candidate for 2026–27. |
|||||||
|
Meta Llama
4 Maverick / Scout
Meta · Apr 2025+
|
$0 self-host | Open-weight ecosystem champion. Llama runs on every inference stack (Ollama, vLLM, Bedrock, NIM). Scout's 10M context window is unmatched in the open. License has 700M MAU clause. |
Llama 4 Behemoth (2T) GA expected H2 2026; will close more of the closed-source gap. Strategic value compounds with Meta's in-house chip rollout. |
Capability scores summarise; strategy explains. Below, each contender's operating thesis, structural advantage, and the single risk a CEO must price into a multi-year contract.
Anthropic has converted "Constitutional AI" from a research talking point into the dominant enterprise distribution strategy. Opus 4.7 is the technical artefact; the business story is 40% of enterprise LLM spend.
Claude becomes the default substrate for agentic and regulated workloads through 2027. The "Mythos" preview signals further reasoning gains. Pricing remains the single largest commercial risk.
No competitor delivers comparable depth across text, image, video, and audio under a single API surface. Veo 3.1 is the most consequential video model in commercial use; Nano Banana 2 is its image counterpart.
The most likely 2026 platform winner. Multimodal saturation, integrated tooling, and Vertex AI distribution compound into a near-monopoly in cross-modal workflows.
The most consequential pricing event in the industry since the launch of GPT-4. DeepSeek's sparse attention architecture, MIT-licensed weights, and gold-medal mathematical reasoning have re-baselined what "expensive AI" means.
DeepSeek becomes the cost benchmark every CFO measures against. The realistic deployment path is self-hosted V3.2 / V4 for non-sensitive workloads alongside a Western model for client-facing surfaces.
A 100× training-compute jump versus the prior generation. Grok's structural advantage is real-time integration with the X graph — for media, market intelligence, and live narrative tracking, no other model has equivalent ground-truth.
Grok 5 (Jan 2026 launch window, 6T-parameter target) is the watch event. If it executes, xAI moves from edge-case to default for media, finance, and intelligence verticals.
Perplexity is not competing on raw model quality — it routes between frontier providers — but on the workflow of grounded research. Comet (free), Deep Research, and citation-native answers redefine how analysts consume the web.
Wins the research vertical, struggles to break into general-purpose AI. Most likely outcome: a $20–50B acquisition or a vertical SaaS pivot toward regulated industries.
Meta has converted open weights into a strategic moat. Llama 4 Scout's 10M-token context window — unmatched by any closed model — paired with Meta's in-house chip programme, makes Llama the de facto open infrastructure standard.
The default "private deployment" stack for regulated, sovereign, and cost-sensitive workloads. Behemoth GA (H2 2026) closes the frontier gap meaningfully — possibly fully.
Copilot is no longer a model — it is a workflow. Microsoft has integrated GPT-5.4 Thinking, GPT-5.3 Instant, Anthropic Claude (via "Critique" and "Council"), and Microsoft's own MAI / Phi family into a single addressable surface across Word, Excel, Outlook, Teams, and GitHub.
Wins enterprise productivity by default. Real strategic story: every Microsoft 365 seat becomes an AI seat. Risk is OpenAI-relationship volatility, not model quality.
Capability convergence is now the base case. The differentiators that determine winners — distribution, governance, sovereignty, and ecosystem lock-in — are increasingly non-technical. Plan for a portfolio architecture, not a single vendor.
Expect frontier-class output pricing to fall a further 40–60% as DeepSeek V4, Llama 4 Behemoth, and Qwen 3.5 close the closed-source gap. CFOs will reset budgets on a quarterly cadence rather than annual.
The agent layer (Claude Code, GPT-5 Codex, Agent 365, Gemini Antigravity) replaces chat as the dominant interaction mode. Headcount-equivalent ROI calculations move from marketing claims to procurement reality.
EU AI Act GPAI enforcement (Aug 2026), U.S. export-control tightening, and Mercosur data-residency rules force every multinational to architect for jurisdictional model routing. Open-weight stacks gain decisive deployment advantage.
For 80% of enterprise workloads, "best model" becomes "best model already in the productivity surface." Microsoft, Google, and Anthropic widen the gap on customers; xAI, Meta, and DeepSeek consolidate vertical and infrastructure positions.
The five-point capability rating is calibrated against the current frontier, refreshed monthly. Inputs combine published benchmarks, vendor API documentation, independent evaluators, and observed real-world performance.
Quantitative inputs include SWE-Bench Verified, GPQA Diamond, ARC-AGI-2, LiveCodeBench Pro, LMArena Elo, AIME, MATH-500, and Terminal-Bench. Latency and throughput are sourced from Artificial Analysis and direct API instrumentation. Cost reflects flagship-tier per-million-token rates as of 22 May 2026. Strategic positioning combines public revenue disclosures, enterprise adoption data (Menlo Ventures), valuation events, and qualitative ecosystem analysis.
This brief is informational. It is not investment advice. Model capabilities, pricing, and availability change weekly — verify against vendor documentation before contracting.