# Fm · Frontier Model — element 1 of 58

> The reasoning core of everything. Turns any question into working judgment.

- **Group:** 1 · Intelligence
- **Necessity:** Essential
- **Price band:** $$ · $30–150/mo
- **Maturity:** Stable
- **Edition:** v2026.Q3 · verified 2026-09-13

## Leading tools (v2026.Q3)

- **Claude** — deep reasoning, enterprise leader
- **ChatGPT / GPT-6 Astra** — the ecosystem default
- **Gemini** — multimodal + google surface
- **Grok** — real-time x, cheapest flagship
- **Meta AI (Muse Spark)** — free frontier, no api

## Our take

Pick one primary model and go deep. Model-hopping is procrastination with extra steps; the switching cost is your prompts, not the subscription.

## Combines with

Ca, Re, Ag
- Guide: [Best frontier model for startups (2026)](https://elems.ai/best/best-frontier-model-for-startups.html)

## The top 5 — deep dossier (verified 2026-09-13)

Claude, and in Aug 2026 it is no longer close on the measures a founder should care about: Anthropic holds the #1 Artificial Analysis Intelligence Index score (Opus 5, 61) plus three of the top five slots, 40% of enterprise LLM API spend (54% in coding), and a $47B revenue run-rate that passed OpenAI's in May 2026. ChatGPT/GPT-5.6 is the pick when consumer distribution, ecosystem breadth, or cheap volume decide — Luna at $0.20/$1.20 per MTok is the lowest price any frontier lab has ever printed. Gemini wins when you live on Google surfaces, need 1M-context multimodal at mid-tier prices, or want the strongest bundled-subscription value. Grok when real-time X data or the cheapest frontier-flagship API matters. Meta AI when free frontier-class answers beat having an API at all.

1. **Claude** (Anthropic) — Free · Pro $17–20/mo · Max from $100/mo · Team $20–25/seat · API: Opus 5 $5/$25, Fable 5 $10/$50 per MTok. Best for: The primary model a founder goes deep on — hardest reasoning, long-horizon agentic work, and the whole Claude Code/Cowork surface on one subscription. Why: Owns the top of every serious capability index: Opus 5 leads the Artificial Analysis Intelligence Index at 61 and Anthropic holds three of the top five slots (Aug 2026); Fable 5 (Jun 9, 2026) was the first model past 90% on core analytics benchmarks with a 1M-token window. The market agrees — 40% of enterprise LLM API spend (54% in coding, Menlo Ventures) and a $47B run-rate by May 2026, up from $9B at end-2025, now the largest of any AI lab. Fable 5.1 and Mythos 5.1 shipped Sep 1, 2026, with a 75% cost reduction on Fable cache reads — the cheapest the top of this index has been. Watch: Consumer distribution is tiny — ~2.1% of chatbot web traffic vs ChatGPT's ~57–67%. Fable 5 shipped, was suspended three days later (Jun 12), and only redeployed Jul 1 — frontier-tier access can move under you. Fable 5's $10/$50 pricing is the most expensive mainstream API on the market. [https://claude.com](https://claude.com)
2. **ChatGPT / GPT-6 Astra** (OpenAI) — Free · Go ~$4.50/mo · Plus $20/mo · Pro $200/mo · Team $25/seat · API: Sol $5/$30, Terra $2/$12, Luna $0.20/$1.20 per MTok (after Jul 30 cuts). Best for: Broadest ecosystem and cheapest high-quality volume — the default when your users, integrations, and hires already live in ChatGPT. Why: GPT-6 Astra (Sep 3, 2026) replaced the GPT-5.6 Sol/Terra/Luna line as the flagship, with an Astra Pro tier at $10/$50 per million tokens. GPT-5.6 (Jul 9, 2026) split the flagship into durable Sol/Terra/Luna tiers, and the Jul 30 price cuts made Luna ($0.20/$1.20) the cheapest near-frontier API ever sold. Distribution is unmatched: ~900M weekly ChatGPT users (Feb 2026), 50M paying subscribers, $25B annualized revenue — the largest consumer AI footprint on earth. Watch: Astra is five weeks old at the time of writing and the pricing above is the launch pricing, not a settled one. The enterprise story is going the other way — API share fell from 50% (2023) to 27% (Menlo), and Sol (59) sits below Opus 5 (61) on the Intelligence Index. Still deeply unprofitable (~$14B/yr losses reported) despite a record $122B raise at $852B (Mar 2026). Pricing has four compounding multipliers (fast mode, long-context 2x over 272k, caching, residency) — worst-case Sol bills run ~4x list. [https://openai.com](https://openai.com)
3. **Gemini** (Google (DeepMind)) — Free · AI Plus $4.99/mo · AI Pro $19.99/mo · AI Ultra $99.99 (5x) / $199.99 (20x) · API: 3.1 Pro $2/$12 per MTok (≤200k ctx; higher above). Best for: Multimodal work, 1M-token context at mid-tier prices, and founders already living on Google Workspace/Cloud — the best bundled-subscription value in AI. Why: Gemini 3.1 Pro (GA Feb 19, 2026) is the abstract-reasoning standout — 94.3% GPQA Diamond, 77.1% ARC-AGI-2, 80.6% SWE-bench Verified — at $2/$12, a third of Sol's output price. Distribution is compounding: 900M Gemini app MAU (doubled YoY) and 1B+ AI Mode users announced at I/O 2026, with the subscription bundling TB-scale storage, Flow credits, and the Jules coding agent. Watch: The divergence has widened: Google shipped four Flash models between May and September 2026 (3.6, 3.7, 3.8 and 3.8 Cyber) while the Pro flagship has not moved since Gemini 3.1 Pro in February. Treating "Gemini" as one moving target mis-states the product — the cheap tier is racing and the flagship is parked. The flagship succession slipped: Gemini 3.5 Pro was announced May 19 at I/O with a 'next month' target and still hadn't shipped by early Aug — only 3.5 Flash made it out. Current 3.1 Pro sits below the Anthropic/OpenAI leaders on the Intelligence Index, and the Plus/Pro/Ultra×Flow-credits tiering is genuinely confusing to buy. [https://gemini.google.com/](https://gemini.google.com/)
4. **Grok** (xAI (SpaceX, SPCX)) — Free · X Premium $8/mo · SuperGrok $30/mo · SuperGrok Plus $100 · Heavy $300 · API: Grok 4.5 $2/$6, Grok 4.3 $1.25/$2.50 per MTok. Best for: Real-time X/web data baked into the model, native video generation, and the cheapest frontier-flagship API on the market. Why: Grok 4.5 (Jul 8, 2026) prices a frontier flagship at $2/$6 with a 500k context — a fraction of Fable 5 or Sol — and is the only frontier model with live X firehose grounding plus native 2K video generation. Backed by the Colossus buildout (~555k GPUs, ~2GW) and now inside the public SpaceX entity after the Feb 2, 2026 all-stock absorption and Jun 12 SPCX IPO. Watch: Commercially it is a distant fourth: ~$500M annualized run-rate mid-2026 vs Anthropic's $47B and OpenAI's $25B, and 117M MAU vs Gemini's 900M. June 2026 moved all paid tiers to one shared weekly usage pool and retired budget models with two months' notice — plan churn is real. Moderation controversies keep it off many enterprise shortlists. [https://x.ai](https://x.ai)
5. **Meta AI (Muse Spark)** (Meta Superintelligence Labs) — Free on meta.ai and Meta AI app · API in private preview, no public pricing. Best for: Free frontier-class multimodal answers at consumer scale — and a bet that Meta's closed-model pivot produces a real fifth frontier seat. Why: Muse Spark (Apr 8, 2026) is Meta's first frontier model and its first without open weights — top-5 on the Artificial Analysis Intelligence Index at launch (52) and 50.2% on Humanity's Last Exam, with strong multimodal and visual-STEM performance. Distribution through Meta's app family gives it the largest free frontier audience anywhere. Watch: Not really buyable yet: the API is private-preview with no published pricing, and it trails rivals on coding and long-horizon agentic work (1,427 vs Claude Sonnet's 1,648 on agent tasks). Strategy whiplash — open-weights champion to closed model in one generation, with leadership saying future versions may reopen — makes it hard to build on. Enterprise share fell 16% → 8% (Menlo). [https://meta.ai](https://meta.ai)

### How to choose
- If You're picking your one primary model — the brief's own advice — and your work is judgment-heavy → Claude. Start on Opus 5 ($5/$25, Intelligence Index #1) and escalate to Fable 5 only for the hardest long-horizon work; a Pro/Max subscription carries Claude Code and Cowork with it.
- If You're shipping AI inside a consumer product, or burning tokens at volume on routine tasks → GPT-5.6 Luna/Terra — $0.20/$1.20 and $2/$12 after the Jul 30 cuts undercut every frontier rival; keep Sol for the hard 5% of calls.
- If Your workload is multimodal (video, audio, images, PDFs) or lives on Google Workspace/Cloud → Gemini 3.1 Pro at $2/$12 with 1M context, and the AI Pro bundle at $19.99 — the strongest per-dollar consumer subscription in AI.
- If You need real-time social/news signal or the cheapest frontier-flagship API → Grok 4.5 at $2/$6 — accepting the plan-churn and the fact that xAI's commercial traction is a rounding error next to the top three.
- If Your bill, privacy, or latency argument beats the capability argument → Don't buy a bigger frontier subscription — go to element Ow: DeepSeek V4-Flash (Index 50 at $0.14/$0.28) and Kimi K3 (Index 57, open) have closed most of the gap at ~1/10 the price.

### The field (16 more)

Claude Mythos 5, Microsoft MAI, Amazon Nova 2, Gemini Spark, DeepSeek V4, Kimi K3, Mistral / Le Chat (Vibe), GLM-5.2, Qwen, Llama frontier line (fading), Cohere (fading), AI21 Labs (fading), Reka (fading), Inflection AI (Pi) (dead), xAI (standalone) (acquired), Grok 3 / 4.1 Fast (dead)

Full dossier data: https://elems.ai/e/fm.json

---
Source: [elems.ai](https://elems.ai/e/fm.html) — the periodic table of the AI-led startup. Data: https://elems.ai/elements.json (CC BY 4.0, cite elems.ai).