There were roughly 255 model releases in Q1 2026 alone — about three a day. As of June 2026, the frontier has fractured: there is no single best model, only a best model for your task. This guide compares every major family by benchmark, price, context window and license — and is built to be the source you can quote.

How to read this: all figures are a June 2026 snapshot. Vendor-reported numbers are marked *; everything else comes from official docs or independent trackers (see Sources). Prices are per 1M tokens (input / output).

Key Takeaways

• No model wins everything. Claude Opus 4.8 leads overall intelligence and agentic coding; GPT-5.5 leads terminal/CLI agents; Gemini 3.1 Pro leads value + multimodal reasoning.
• The top is tight. On the Artificial Analysis (AA) Intelligence Index: Opus 4.8 (61.4) → GPT-5.5 (60.2) → Gemini 3.1 Pro (57) → Grok 4.3 (53) — within ~8 points.
• Open weights caught up. DeepSeek V4-Pro and MiniMax M3 (AA Index ~44) post near-frontier coding scores at 5–12× lower price.
• Context exploded. Llama 4 Scout holds the largest publicly released window at 10M tokens; 1M is now the default.

How This Was Compiled

We tracked every actively marketed flagship per lab as of June 17, 2026, then cross-checked specs against official model cards / pricing pages and independent trackers (Artificial Analysis Intelligence Index v4.1, LLM-Stats). Where sources disagree — common in this market — we give the range and cite both. Treat every number as a dated snapshot, not a permanent fact.

The Master Table

Closed / proprietary frontier

Open-weight / open-source

Intelligence: the frontier is a cluster

Prefer a static version? Download the chart: AA Intelligence Index (PNG).

Context windows: 1M is the new default

Static version: Context windows (PNG).

Price: open weights reset the floor

Static version: Output price (PNG).

By Family

OpenAI — GPT

GPT-5.5 (Apr 23, 2026) unified the o-series + Codex into one flagship. It leads terminal-native agents (Terminal-Bench 2.0 ≈ 82.7%), is natively omnimodal, and ships an Instant default plus a higher-accuracy Pro. Pricing $5/$30; context 400K (Codex) to 1M (API). Lineage: GPT-5 (Aug 2025) → 5.1 → 5.2 → 5.4 → 5.5.

Anthropic — Claude

Opus 4.8 (May 28, 2026) is the June overall leader (AA 61.4): top agentic coding (SWE-V 88.6%), computer use (OSWorld 83.4%), knowledge work (GDPval-AA 1,890), best hallucination calibration, plus a Fast Mode. Tiers: Opus 4.8 ($5/$25) · Sonnet 4.6 ($3/$15) · Haiku 4.5 ($1/$5). The Mythos-class Fable 5 (Jun 9) set a coding ceiling (SWE-Pro 80.3%*) but its availability is currently restricted.

Google — Gemini

Gemini 3.1 Pro (Feb 2026, AA 57) is the value + multimodal leader (~$2 input, 1M ctx, GPQA ≈ 94.3%). 3.5 Flash (May 19, GA) is the fastest/cheapest tier yet ($1.50/$9, ~4× faster). 3.5 Pro (Deep Think, up to 2M ctx) is rolling out in June.

xAI — Grok

Grok 4.3 (Apr 30, 2026, AA 53): budget frontier — 1M context, function calling, structured outputs, $1.25/$2.50. Strongest features sit behind the $300/mo SuperGrok Heavy tier.

Meta — Llama + Muse

Llama 4 Scout (10M ctx) and Maverick (1M ctx) are MoE, natively multimodal, single-host friendly. Behemoth (~2T) was shelved. The Muse family now has three models: Muse Spark 1.2 (Aug 5, coding flagship, AA 54, $1.25/$4.25), Muse Glimmer (Aug 10, 30B Apache 2.0 for local agents), and the original Muse Spark (Apr 8, AA 52). Llama 5 slipped toward 2027.

DeepSeek (open weight, MIT)

V4 (Apr 24, 2026) made 1M context cheap. V4-Pro (1.6T/49B active) posts ~85% SWE-V; V4-Flash is the cheapest credible coder ($0.14/$0.28). R2 has not shipped — treat R2 specs as rumor.

Alibaba — Qwen

Qwen3.7-Max (May 19) is proprietary, API-only (1M ctx, $2.50/$7.50): the “Agent Frontier” with the highest knowledge scores in its peer set. Open weights continue lower: Qwen3.6-27B / 35B-A3B (Apache 2.0).

Chinese open-source wave

• Kimi K2.7 Code (Jun 12): 1T MoE, 256K ctx, leads MCP tool use (81.1%).
• GLM-5.2 (Jun 13): 1M ctx (largest open), MIT pending, ~$1.40/$4.40.
• MiniMax M3 (Jun 1): 1M ctx, native image/video, $0.30/$1.20, top open AA Index.

NVIDIA — Nemotron

Nemotron 3 Ultra (Jun 4): 550B, the most capable open model under a fully permissive license (no thresholds). Nemotron 3.5 Lightning (Aug 11): highest-efficiency model for long-running agentic workloads, ships with NeMo Switchyard multi-model routing library.

Mistral — Europe’s open standard

Mistral Large 3 (Dec 2, 2025): 41B active / 675B total MoE, 256K ctx, native vision, Apache 2.0 ($0.50/$1.50), shipped with edge-focused Ministral 3.

FAQ

What is the best LLM in June 2026? There is no single winner. Claude Opus 4.8 leads overall intelligence and agentic coding; GPT-5.5 leads terminal/CLI agents; Gemini 3.1 Pro leads value and multimodal reasoning; DeepSeek V4-Pro leads value-per-benchmark among open weights.

What is the best open-source LLM? MiniMax M3 and DeepSeek V4-Pro tie at the top of the AA Index among open weights (~44). Kimi K2.7 Code leads MCP tool use; Llama 4 Scout leads context length (10M); NVIDIA Nemotron 3 Ultra leads on permissive licensing.

Which model has the largest context window? Llama 4 Scout, at 10M tokens — the largest of any publicly released LLM as of June 2026.

What is the cheapest capable model? DeepSeek V4-Flash at $0.14 / $0.28 per 1M tokens, with a verified ~79% SWE-bench Verified score.

July 2026 Update

The landscape moved fast in July. Here’s what changed since the initial June snapshot:

New frontier models

New open-weight models

Updated takeaways

• Claude Opus 5 is the new overall leader, displacing Opus 4.8 at the same price. The gap to Fable 5 on SWE-bench Pro is now ~1 point (79.2% vs 80.3%).
• GPT-5.6 tiered pricing is live. Terra matches GPT-5.5 quality at half the price; Luna is the cheapest GPT-5.x at $1/$6.
• The open-weight ceiling jumped to 2.8T with Kimi K3, but the custom license and 64+ GPU requirement limit practical adoption. Inkling offers a US-origin Apache 2.0 alternative at 975B.
• Gemini 3.5 Pro is still missing. Google shipped 3.6 Flash instead. Pre-training for Gemini 4 has begun.

August 2026 Update

Early August brought a Chinese-model offensive and a video-generation surge; the week of August 5–11 added a Meta open-source pivot, NVIDIA’s agentic play, and significant GPT-5.6 updates:

New frontier models

New open-weight models

GPT-5.6 mid-August updates

GPT-5.6 received three significant changes:

• Sol — improved reasoning slider for granular compute allocation; now responds to “think more/less” in natural language.
• Luna — promoted to free default for all ChatGPT users (unlimited text, 25 image gens/day); added “Think” button for on-demand extended reasoning.
• GPT-5.6-Cyber — new security-hardened variant fine-tuned on adversarial datasets; 95% detection rate on prompt injection benchmarks. Available via API at Sol pricing.

Cross-modal expansion

The LLM landscape is increasingly entangled with video and multimodal models:

Updated takeaways

• Meta pivots to open source. Muse Glimmer (Apache 2.0) is Meta’s first fully permissive model — a strategic reversal after Llama’s restrictive community license. Muse Spark 1.2 (open weights announced) will follow. The Contributor tier ($0.10/$0.20 for frontier inference in exchange for training data) is a novel pricing model worth watching.
• Qwen 3.8-Max open weights arrive August 12. ModelScope shows a countdown. Combined with Kimi K3 (2.8T, July), the open-weight frontier is in a sprint. A 27B variant (Qwen3.8-27B) follows shortly after.
• Local agentic models are now a category. Muse Glimmer (30B, 24GB VRAM) and Nemotron 3.5 Lightning (NeMo Switchyard routing) both target always-on agents that run on local or enterprise hardware — a category that didn’t exist three months ago.
• GPT-5.6 gets wider, not taller. OpenAI is expanding horizontally (Luna free for all, Cyber for security) rather than pushing a new number. The reasoning slider on Sol and the Think button on Luna both address the “how much compute” question differently than any rival.
• Video input is the new frontier. Wan 3.0’s document-to-video and MAGI-2’s permissive video generation join FLUX 3 and MiniMax H3 — four new video models in one week, eclipsing text-to-image as the hot modality.
• The frontier pricing ladder is clearer: Fable 5 ($10/$50) for max capability → Opus 5 ($5/$25) for near-Fable at half price → Qwen 3.8-Max ($2/$6) for agentic workloads on a budget → Muse Spark 1.2 ($1.25/$4.25) for coding-focused value → Sonnet 5 ($2/$10) for workhorse tasks.

Sources

1. Anthropic — Claude models overview
2. OpenAI — Introducing GPT-5.5
3. Google — Gemini 3.5
4. xAI — Grok 4.3 docs
5. Meta — Llama 4 herd
6. DeepSeek — V4 release notes
7. Alibaba Cloud — Qwen3.7: The Agent Frontier
8. Mistral AI — Introducing Mistral 3
9. Artificial Analysis — Intelligence Index
10. Build Fast with AI — Best AI Models June 2026
11. Alibaba Cloud — Qwen3.8-Max: A New Bar for Coding and Cowork

Changelog: 2026-08-11 — August update 2: added Muse Spark 1.2, Muse Glimmer, Nemotron 3.5 Lightning; GPT-5.6 Sol/Luna/Cyber mid-August updates; Qwen 3.8-Max open-weight countdown; Grok Imagine Image 2.0, Wan 3.0, MAGI-2 to cross-modal expansion; updated Meta and Nemotron family sections. 2026-08-05 — August update: added Qwen 3.8-Max, cross-modal expansion section (MiniMax H3, FLUX 3 Video GA, Grok Imagine Video 1.5, SeedRealtime). 2026-07-29 — July update: added Opus 5, GPT-5.6 GA, Grok 4.5, Gemini 3.6 Flash, Kimi K3, Inkling. 2026-06-17 — initial publication.

Last updated: 2026-08-11