EVERY MAJOR AI LAB · ONE LIVE REFERENCE
Every frontier model, benchmarked and put side by side.
A single reference for every major AI product and model on the market — flagship LLMs, image and video generators, and coding agents — with model-to-model comparisons, abstract-reasoning (AGI-track) scores, and a dedicated cybersecurity posture index. Refreshed monthly.
20
models tracked across 8 labs
Claude Fable 5.1
current #1 on the Intelligence Index
95%
top public ARC-AGI-2 score (abstract reasoning)
OpenAI GPT-6 Astra released Sep 3, 2026
Google DeepMind Gemini 3.8 Flash released Sep 2, 2026
Meta Muse Spark 1.3 released Sep 2, 2026
Anthropic Claude Fable 5.1 released Sep 1, 2026
DeepSeek DeepSeek‑V4‑Pro released Aug 13, 2026
xAI Grok 4.6 released Aug 12, 2026
Alibaba Qwen3.8‑Max released Aug 3, 2026
Anthropic Claude Opus 5 released Jul 24, 2026
OpenAI GPT-5.6 Sol released Jul 9, 2026
OpenAI GPT-5.6 Terra released Jul 9, 2026
OpenAI GPT-6 Astra released Sep 3, 2026
Google DeepMind Gemini 3.8 Flash released Sep 2, 2026
Meta Muse Spark 1.3 released Sep 2, 2026
Anthropic Claude Fable 5.1 released Sep 1, 2026
DeepSeek DeepSeek‑V4‑Pro released Aug 13, 2026
xAI Grok 4.6 released Aug 12, 2026
Alibaba Qwen3.8‑Max released Aug 3, 2026
Anthropic Claude Opus 5 released Jul 24, 2026
OpenAI GPT-5.6 Sol released Jul 9, 2026
OpenAI GPT-5.6 Terra released Jul 9, 2026
01 — LABS & PRODUCTS
Who’s building what
Eight labs currently define the frontier. Each ships more than a chatbot — reasoning models, coding agents, image and video generators, and voice systems all sit under one roof.
AN
Anthropic
Claude Fable 5.1
Claude family — Haiku, Sonnet, Opus, plus the new Mythos-tier (Fable/Mythos 5.1) above Opus.
OP
OpenAI
GPT-6 Astra
GPT-6 Astra (Sept 3, 2026) tops the GPT-5.6 (Sol/Terra/Luna) family, with computer use and a Critical-tier cybersecurity designation, alongside Sora video and native image generation.
GO
Google DeepMind
Gemini 3.8 Flash
Gemini 3.x line spanning Flash to Pro/Deep Think, plus Veo video and Imagen.
XA
xAI
Grok 4.6
Grok 4.x line, now under SpaceXAI, with deep X/real-time data integration and long-running agentic workflows.
ME
Meta
Llama 4 Maverick
Open-weight Llama 4 line plus the newer Muse Spark consumer assistant models.
DE
DeepSeek
DeepSeek‑V4‑Pro
Cost-efficient reasoning and coding models, popular for API-heavy workloads.
MI
Mistral AI
Mistral Medium 3.5
EU-based models built for data residency, on-prem deployment and multilingual use.
AL
Alibaba
Qwen3.8‑Max
Qwen line — strong multilingual and coding performance, widely used open-weight base.
02 — PRODUCT CATEGORIES
Beyond the chatbot
The same labs compete across five other product categories. This isn’t a benchmarked ranking — just a map of who ships what today.
Image generationtext/image → image
GPT Image (OpenAI)Imagen (Google)Midjourney v7Flux (Black Forest Labs)
Video generationtext/image → video
Sora 2 (OpenAI)Veo 3 (Google)Runway Gen-4Kling / Seedream (ByteDance)
Voice & audiospeech, music, cloning
ElevenLabsGemini Live AudioOpenAI Realtime Voice
Coding agentsIDE / CLI / autonomous
Claude CodeGitHub Copilot / CodexGemini CLICursor
Search & research agentsmulti-step web research
Claude Cowork / ResearchChatGPT Deep ResearchGemini Deep ResearchGrok DeepSearch
03 — DIRECTORY
Every tracked model
Filter by lab or search by name. Context window and pricing shown per million tokens where publicly listed. Scores marked EST are our estimate where a lab hasn’t published a directly comparable figure.
| Model | Provider | Released | Context | Price in/out ($/M) | ARC‑AGI‑2 | GPQA | SWE‑bench | Best for |
|---|---|---|---|---|---|---|---|---|
| GPT-6 AstraNEWGATEDFlagship — Critical cyber tier | OpenAI | Sep 3, 2026 | 1.05M | $10 / $50 | 95% | 96% | 74.1% EST | Computer use, agentic software engineering, science |
| Gemini 3.8 FlashNEWLatest stable (Flash) | Google DeepMind | Sep 2, 2026 | 1M | $0.5 / $3 | 86% EST | 94% EST | 73.7% | Fast multimodal & agentic workflows |
| Muse Spark 1.3NEWConsumer assistant | Meta | Sep 2, 2026 | 256K | — / — | 42.5% | 55% EST | 75.4% | Consumer assistant experiences, edges frontier labs on some agentic-coding evals |
| Claude Fable 5.1NEWMythos-class flagship | Anthropic | Sep 1, 2026 | 1M | $15 / $75 | 90% | 92.6% | 81.2% | Frontier reasoning, long-running agents |
| DeepSeek‑V4‑ProFlagship | DeepSeek | Aug 13, 2026 | 128K | $0.5 / $1.5 | 55% EST | 78% EST | 52% EST | Cost-efficient coding & reasoning at scale |
| Grok 4.6NEWFlagship | xAI | Aug 12, 2026 | 500K | $2 / $6 | 58% EST | 89% EST | 65.9% | Chat + coding with real-time X data, long-running agents |
| Qwen3.8‑MaxFlagship | Alibaba | Aug 3, 2026 | 256K | $1.2 / $3.6 | 52% EST | 76% EST | 48% EST | Multilingual + open-weight coding |
| Claude Opus 5Flagship (public) | Anthropic | Jul 24, 2026 | 1M | $15 / $75 | 90.4% | 91.5% | 78% | Professional coding, enterprise work |
| GPT-5.6 SolFlagship reasoning | OpenAI | Jul 9, 2026 | 400K | $35 / $35 | 92.5% | 94.6% | 73% EST | Hardest mixed reasoning/business tasks |
| GPT-5.6 TerraBalanced workhorse | OpenAI | Jul 9, 2026 | 400K | $14 / $14 | 83.9% | 90.5% EST | 70% EST | Default general-purpose OpenAI model |
| GPT-5.6 LunaFast / cost-efficient | OpenAI | Jul 9, 2026 | 128K | $2 / $8 | 59.5% | 82% EST | 55% EST | High-throughput, cost-sensitive apps |
| Grok 4.5Prior flagship | xAI | Jul 8, 2026 | 500K | $2 / $6 | 52.6% | 87% EST | 75% | Multi-agent collaboration tasks |
| Claude Sonnet 5Everyday flagship | Anthropic | Jun 30, 2026 | 1M | $3 / $15 | 60% EST | 89% EST | 72.2% | Day-to-day Claude deployments, chat + agents |
| Gemini 3.5 FlashMid-tier | Google DeepMind | Jun 20, 2026 | 1M | $0.4 / $2.5 | 72.1% | 88% EST | 37% | Balanced speed/cost multimodal tasks |
| Mistral Medium 3.5Flagship | Mistral AI | Apr 28, 2026 | 128K | $1 / $3 | 38% EST | 72% EST | 33% EST | EU data-residency, on-prem deployment |
| GPT-5.5Prior flagship | OpenAI | Apr 23, 2026 | 400K | $35 / $35 | 85% | 90% EST | 58.6% | General knowledge work |
| Gemini 3.1 ProFrontier value | Google DeepMind | Feb 19, 2026 | 1M–2M | $2 / $12 | 77.1% | 94.3% | 63.8% | Best reasoning-per-dollar, science QA, video |
| Claude Haiku 4.5Fast / low-cost | Anthropic | Oct 15, 2025 | 200K | $1 / $5 | 25% EST | 73% | 73.3% | High-volume, latency-sensitive tasks |
| Llama 4 MaverickOpen-weight flagship | Meta | Apr 5, 2025 | 1M | $0.2 / $0.6 | 18% EST | 48% EST | 28% EST | Self-hosted multimodal deployment |
| Llama 4 ScoutOpen-weight, long-context | Meta | Apr 5, 2025 | 10M | $0.15 / $0.4 | 14% EST | 44% EST | 22% EST | Extreme-long-context open workloads |
GPT-6 AstraNEWFlagship — Critical cyber tier
OpenAI
RELEASED
Sep 3, 2026
CONTEXT
1.05M
PRICE IN/OUT
$10 / $50
ARC-AGI-2
95%
GPQA
96%
SWE-BENCH
74.1% EST
BEST FOR
Computer use, agentic software engineering, science
Gemini 3.8 FlashNEWLatest stable (Flash)
Google DeepMind
RELEASED
Sep 2, 2026
CONTEXT
1M
PRICE IN/OUT
$0.5 / $3
ARC-AGI-2
86% EST
GPQA
94% EST
SWE-BENCH
73.7%
BEST FOR
Fast multimodal & agentic workflows
Muse Spark 1.3NEWConsumer assistant
Meta
RELEASED
Sep 2, 2026
CONTEXT
256K
PRICE IN/OUT
— / —
ARC-AGI-2
42.5%
GPQA
55% EST
SWE-BENCH
75.4%
BEST FOR
Consumer assistant experiences, edges frontier labs on some agentic-coding evals
Claude Fable 5.1NEWMythos-class flagship
Anthropic
RELEASED
Sep 1, 2026
CONTEXT
1M
PRICE IN/OUT
$15 / $75
ARC-AGI-2
90%
GPQA
92.6%
SWE-BENCH
81.2%
BEST FOR
Frontier reasoning, long-running agents
DeepSeek‑V4‑ProFlagship
DeepSeek
RELEASED
Aug 13, 2026
CONTEXT
128K
PRICE IN/OUT
$0.5 / $1.5
ARC-AGI-2
55% EST
GPQA
78% EST
SWE-BENCH
52% EST
BEST FOR
Cost-efficient coding & reasoning at scale
Grok 4.6NEWFlagship
xAI
RELEASED
Aug 12, 2026
CONTEXT
500K
PRICE IN/OUT
$2 / $6
ARC-AGI-2
58% EST
GPQA
89% EST
SWE-BENCH
65.9%
BEST FOR
Chat + coding with real-time X data, long-running agents
Qwen3.8‑MaxFlagship
Alibaba
RELEASED
Aug 3, 2026
CONTEXT
256K
PRICE IN/OUT
$1.2 / $3.6
ARC-AGI-2
52% EST
GPQA
76% EST
SWE-BENCH
48% EST
BEST FOR
Multilingual + open-weight coding
Claude Opus 5Flagship (public)
Anthropic
RELEASED
Jul 24, 2026
CONTEXT
1M
PRICE IN/OUT
$15 / $75
ARC-AGI-2
90.4%
GPQA
91.5%
SWE-BENCH
78%
BEST FOR
Professional coding, enterprise work
GPT-5.6 SolFlagship reasoning
OpenAI
RELEASED
Jul 9, 2026
CONTEXT
400K
PRICE IN/OUT
$35 / $35
ARC-AGI-2
92.5%
GPQA
94.6%
SWE-BENCH
73% EST
BEST FOR
Hardest mixed reasoning/business tasks
GPT-5.6 TerraBalanced workhorse
OpenAI
RELEASED
Jul 9, 2026
CONTEXT
400K
PRICE IN/OUT
$14 / $14
ARC-AGI-2
83.9%
GPQA
90.5% EST
SWE-BENCH
70% EST
BEST FOR
Default general-purpose OpenAI model
GPT-5.6 LunaFast / cost-efficient
OpenAI
RELEASED
Jul 9, 2026
CONTEXT
128K
PRICE IN/OUT
$2 / $8
ARC-AGI-2
59.5%
GPQA
82% EST
SWE-BENCH
55% EST
BEST FOR
High-throughput, cost-sensitive apps
Grok 4.5Prior flagship
xAI
RELEASED
Jul 8, 2026
CONTEXT
500K
PRICE IN/OUT
$2 / $6
ARC-AGI-2
52.6%
GPQA
87% EST
SWE-BENCH
75%
BEST FOR
Multi-agent collaboration tasks
Claude Sonnet 5Everyday flagship
Anthropic
RELEASED
Jun 30, 2026
CONTEXT
1M
PRICE IN/OUT
$3 / $15
ARC-AGI-2
60% EST
GPQA
89% EST
SWE-BENCH
72.2%
BEST FOR
Day-to-day Claude deployments, chat + agents
Gemini 3.5 FlashMid-tier
Google DeepMind
RELEASED
Jun 20, 2026
CONTEXT
1M
PRICE IN/OUT
$0.4 / $2.5
ARC-AGI-2
72.1%
GPQA
88% EST
SWE-BENCH
37%
BEST FOR
Balanced speed/cost multimodal tasks
Mistral Medium 3.5Flagship
Mistral AI
RELEASED
Apr 28, 2026
CONTEXT
128K
PRICE IN/OUT
$1 / $3
ARC-AGI-2
38% EST
GPQA
72% EST
SWE-BENCH
33% EST
BEST FOR
EU data-residency, on-prem deployment
GPT-5.5Prior flagship
OpenAI
RELEASED
Apr 23, 2026
CONTEXT
400K
PRICE IN/OUT
$35 / $35
ARC-AGI-2
85%
GPQA
90% EST
SWE-BENCH
58.6%
BEST FOR
General knowledge work
Gemini 3.1 ProFrontier value
Google DeepMind
RELEASED
Feb 19, 2026
CONTEXT
1M–2M
PRICE IN/OUT
$2 / $12
ARC-AGI-2
77.1%
GPQA
94.3%
SWE-BENCH
63.8%
BEST FOR
Best reasoning-per-dollar, science QA, video
Claude Haiku 4.5Fast / low-cost
Anthropic
RELEASED
Oct 15, 2025
CONTEXT
200K
PRICE IN/OUT
$1 / $5
ARC-AGI-2
25% EST
GPQA
73%
SWE-BENCH
73.3%
BEST FOR
High-volume, latency-sensitive tasks
Llama 4 MaverickOpen-weight flagship
Meta
RELEASED
Apr 5, 2025
CONTEXT
1M
PRICE IN/OUT
$0.2 / $0.6
ARC-AGI-2
18% EST
GPQA
48% EST
SWE-BENCH
28% EST
BEST FOR
Self-hosted multimodal deployment
Llama 4 ScoutOpen-weight, long-context
Meta
RELEASED
Apr 5, 2025
CONTEXT
10M
PRICE IN/OUT
$0.15 / $0.4
ARC-AGI-2
14% EST
GPQA
44% EST
SWE-BENCH
22% EST
BEST FOR
Extreme-long-context open workloads
04 — THIS MONTH
Latest flagship, per lab
The newest top-tier release from each major provider, right now.
GPT-6 Astra
OpenAI
Sep 3, 2026
Computer use, agentic software engineering, science. Flagship — Critical cyber tier.
95%
ARC-AGI-2
1.05M
CONTEXT
$10
$/M IN
Gemini 3.8 Flash
Google DeepMind
Sep 2, 2026
Fast multimodal & agentic workflows. Latest stable (Flash).
86% EST
ARC-AGI-2
1M
CONTEXT
$0.5
$/M IN
Muse Spark 1.3
Meta
Sep 2, 2026
Consumer assistant experiences, edges frontier labs on some agentic-coding evals. Consumer assistant.
42.5%
ARC-AGI-2
256K
CONTEXT
—
$/M IN
Claude Fable 5.1
Anthropic
Sep 1, 2026
Frontier reasoning, long-running agents. Mythos-class flagship.
90%
ARC-AGI-2
1M
CONTEXT
$15
$/M IN
Grok 4.6
xAI
Aug 12, 2026
Chat + coding with real-time X data, long-running agents. Flagship.
58% EST
ARC-AGI-2
500K
CONTEXT
$2
$/M IN
05 — HEAD TO HEAD
Compare up to three models
Pick any models from the dropdowns below to see full specs, pricing, and benchmark scores side by side.
OpenAI
GPT-6 Astra
TIER
Flagship — Critical cyber tier
RELEASED
Sep 3, 2026
CONTEXT WINDOW
1.05M
PRICE ($/M IN / OUT)
$10 / $50
ARC-AGI-2
95%
GPQA DIAMOND
96%
SWE-BENCH
74.1% EST
INTELLIGENCE INDEX
61
BEST FOR
Computer use, agentic software engineering, science
OpenAI
GPT-5.6 Sol
TIER
Flagship reasoning
RELEASED
Jul 9, 2026
CONTEXT WINDOW
400K
PRICE ($/M IN / OUT)
$35 / $35
ARC-AGI-2
92.5%
GPQA DIAMOND
94.6%
SWE-BENCH
73% EST
INTELLIGENCE INDEX
61
BEST FOR
Hardest mixed reasoning/business tasks
Google DeepMind
Gemini 3.8 Flash
TIER
Latest stable (Flash)
RELEASED
Sep 2, 2026
CONTEXT WINDOW
1M
PRICE ($/M IN / OUT)
$0.5 / $3
ARC-AGI-2
86% EST
GPQA DIAMOND
94% EST
SWE-BENCH
73.7%
INTELLIGENCE INDEX
58 EST
BEST FOR
Fast multimodal & agentic workflows
06 — AGI TRACK
Abstract reasoning & general intelligence
No model has passed a real AGI test — there isn’t a certified one. These are the closest public proxies: novel visual puzzles, graduate-level science questions, and aggregate “intelligence index” scores that combine dozens of evals.
New this update: GPT-6 Astra (Sept 3, 2026) reports up to 99.9% on ARC-AGI-3 — a harder, newer generation of the puzzle set below, measured on OpenAI’s own stateful adapter harness — and 97.6% on FrontierMath Tier 4. Independent runs on the standard neutral harness put the same ARC-AGI-3 figure closer to 62.7%, and ARC Prize’s own standard-harness ARC-AGI-2 score for Astra is 95.0%, which is on the same scale as the bars below and is now included there. On the independent Artificial Analysis Intelligence Index, Astra scores ~61 — roughly level with GPT-5.6 Sol and behind Claude Fable 5.1 and Meta’s Muse Spark 1.3.
ARC‑AGI‑2 — novel abstract reasoning
higher = better · % tasks solvedGrid-puzzle tasks designed to resist memorization; average untrained human scores ~66%. Considered the hardest widely-used public reasoning benchmark. Source: arcprize.org public leaderboard.
GPQA Diamond — graduate-level science reasoning
higher = better · % correctPhD-level, Google-proof multiple-choice questions across biology, chemistry and physics.
Artificial-Analysis-style Intelligence Index
composite of ~10 public evals, normalized 0–100A weighted blend of reasoning, knowledge, coding and math benchmarks used as a rough single-number stand-in for general capability.
07 — CYBERSECURITY TRACK
Cybersecurity posture index
How current models behave as coding assistants and how they resist misuse for offensive cyber tasks.
Reading this table: there is no single certified “AI cybersecurity leaderboard” — labs publish results using different internal red-team suites and methodologies (Meta’s CyberSecEval, SecBench, and lab-specific system-card evaluations). Ratings below are a directional synthesis of published safety/system-card disclosures and third-party security research, not a precise head-to-head score. Industry-wide baseline from CyberSecEval: LLMs suggest insecure code in roughly 1 of every 3 completions on average, and comply with cyberattack-assistance requests roughly half the time absent added safeguards — frontier labs now layer additional filtering on top of the base model to push these numbers down.
| Model | Secure-coding tendency | Prompt-injection resistance | Cyberattack-request refusal | Uplift gating | Notes |
|---|---|---|---|---|---|
| GPT-6 Astra | High | Medium | High | GATED | First model OpenAI has designated “Critical” (its highest tier) under the Preparedness Framework for cyber risk. Public rollout ships refusing advanced offensive-cyber tasks; full capability limited to vetted defenders in OpenAI’s Daybreak program. Scored 100% on OpenAI’s internal ExploitBench. |
| Claude Opus 5 / Fable 5.1 | High | High | High | GATED | Anthropic’s Responsible Scaling Policy applies extra cyber-uplift safeguards at this capability tier; Mythos variant relaxes some restrictions for vetted orgs only. |
| GPT-5.6 Sol / Terra | High | Medium | High | GATED | OpenAI’s Preparedness Framework gates high-uplift cyber capability; strong on secure-code suggestion benchmarks. |
| Gemini 3.1 Pro / 3.8 Flash | Medium | Medium | High | GATED | Google’s Frontier Safety Framework applies dangerous-capability evaluations pre-release; a dedicated “Flash Cyber” variant ships alongside the base 3.8 Flash model. |
| Grok 4.6 | Medium | Medium | Medium | ungated | Lighter published safety-framework detail than the other three frontier labs; independent red-team coverage is thinner. |
| Llama 4 (open-weight) | Medium | Lower | Lower | ungated | Open weights mean any safety fine-tuning can be stripped by a downstream deployer — protections aren’t guaranteed at inference time. |
| DeepSeek‑V4‑Pro | Medium | Lower | Lower | ungated | Limited independent third-party security red-teaming publicly available as of this update. |
Secure-coding tendency
How often the model’s generated code avoids known insecure patterns (CWE-mapped), per CyberSecEval-style static analysis.
Prompt-injection resistance
Resistance to hidden instructions embedded in tool outputs, documents, or web content hijacking the model’s behavior.
Offensive-cyber uplift gating
Whether the lab applies extra restrictions/monitoring to outputs that could materially assist real-world cyberattacks, per its published safety framework.
08 — METHODOLOGY
How this page is built
01
Public sources only
Every figure traces to a lab’s own release notes/system cards, or a recognized public leaderboard (ARC Prize, GPQA, SWE-bench, Artificial Analysis).
02
Refreshed monthly
New releases, price changes and re-benchmarked scores are folded in on a monthly pass — labs currently ship a new model every few weeks.
03
No single “best”
Rankings shift by task. This page favors side-by-side specs over crowning one universal winner.
04
Estimated fields, marked
Smaller or less-benchmarked providers don’t always publish ARC-AGI-2/GPQA/SWE-bench figures directly. Where we’ve interpolated from a closely related published eval, the directory marks the model with an EST badge.
· arcprize.org — ARC-AGI-2 public leaderboard· Artificial Analysis — Intelligence Index· OpenAI — GPT-6 Astra system card & launch benchmarks· xAI — Grok 4.6 announcement· Meta AI — CyberSecEval / Purple Llama benchmark suite· Provider system cards & release notes (OpenAI, Anthropic, Google DeepMind, xAI, Meta)· SWE-bench Verified / SWE-bench Pro public leaderboards· GPQA Diamond public leaderboard· Independent trackers (DataLearner, BenchLM, BenchmarkList) for cross-model comparison
