The AI Leaderboard tracks 39 current, specialist, open-weight and migration-reference models across 11 providers. Claude Fable 5.1 leads the current top eight with an Overall score of 83.4. The remaining 31 models form a broader directory and are not assigned invented ranks.
Quick answer: Claude Fable 5.1 is the current overall leader. DeepSeek V4.1 Flash is the highest-ranked open-weights model. GPT-6 Astra is the highest-ranked OpenAI model, while Claude Opus 5.5 is the number-two overall model. Use the 39-model directory below to compare access, licensing, context and price before testing a shortlist.
September 2026 edition at a glance
Question | Current answer | Evidence boundary |
|---|---|---|
Best overall in the verified ranking | Claude Fable 5.1 | #1, Overall 83.4 |
Highest-ranked open weights | DeepSeek V4.1 Flash | #6, Overall 81.1 |
Highest-ranked OpenAI model | GPT-6 Astra | #4 in the current ranking |
Lower-priced top-three Anthropic option | Claude Opus 5.5 | #2 at a documented $4 input / $20 output per million tokens |
Large-context multimodal option | GPT-6 Astra | #4 with a 1,050,000-token context window |
Coding-specific OpenAI model | GPT-5.3-Codex | Specialist directory pick, not assigned an Overall rank here |
Restricted specialist models | Claude Mythos 5 and Claude Mythos 5.1 | Not generally purchasable |
Models covered | 39 | Eight ranked plus 31 broader directory entries |
What changed this month
Claude Fable 5.1 remains the current leader. Claude Opus 5.5 now ranks second, GPT-6 Astra ranks fourth, and DeepSeek V4.1 Flash is the highest-ranked open-weights model at number six. The directory now includes 39 dedicated guides.
- Added Claude Opus 5.5, GPT-6 Astra and DeepSeek V4.1 Flash to the complete model directory.
- Kept the verified ranking limited to eight models with current validated scores.
- Moved models that left the current top eight into the broader directory without deleting their guides.
- Preserved availability, licensing, context, pricing and provider navigation for every model.
- Synchronized this page with the same eight leaderboard records used by the homepage and llms.txt.
Current ranked top eight
These are the eight models in the current verified ranking. Overall scores are comparative benchmark scores, not accuracy percentages or production success rates.
Rank | Model | Provider | Overall | License | Context | Availability |
|---|---|---|---|---|---|---|
1 | Claude Fable 5.1 | Anthropic | 83.4 | Proprietary | 1M tokens | Available |
2 | Claude Opus 5.5 | Anthropic | 83.2 | Proprietary | 1 million tokens | Active on the Claude API and supported cloud platforms |
3 | Claude Fable 5 | Anthropic | 83.0 | Proprietary | 1M tokens | Available |
4 | GPT-6 Astra | OpenAI | 82.2 | Proprietary | 1,050,000 tokens | Rolling out through the OpenAI API and eligible ChatGPT plans |
5 | Muse Spark 1.3 | Meta | 81.6 | Proprietary hosted model | 1 million tokens | Available through Muse Code and Meta Model API; max reasoning was still pending additional safety testing at launch |
6 | DeepSeek V4.1 Flash | DeepSeek | 81.1 | MIT License for the released model weights and repository | 1 million tokens | Available through the DeepSeek API and downloadable open weights |
7 | GPT 5.6 Sol | OpenAI | 81.0 | Proprietary | 1.05M tokens | Available |
8 | GPT 5.5 | OpenAI | 80.2 | Proprietary | 1.05M tokens | Available |
Only 3.2 points separate first from eighth. Price, deployment control, modalities and workload fit can matter more than a small score gap.
Best AI model shortlists by need
These are starting shortlists, not additional benchmark rankings. They combine the verified top-eight order with documented model specialization, licensing and access.
Need | Start with | Also evaluate | Why |
|---|---|---|---|
Highest current Overall score | Claude Fable 5.1 | Claude Opus 5.5 and Claude Fable 5 | The current top three |
Coding and software agents | Claude Fable 5.1 | GPT-6 Astra, Claude Opus 5.5 and GPT-5.3-Codex | Overall leadership plus current agentic and coding specialists |
Open weights and deployment control | DeepSeek V4.1 Flash | Kimi K3, DeepSeek V4, MiniMax M3 and Qwen 3.8 | The ranked open-weight leader plus self-hosting candidates |
OpenAI ecosystem | GPT-6 Astra | GPT 5.6 Sol, GPT 5.5, GPT-5.4 and GPT-5.3-Codex | Current rank plus general and coding-focused alternatives |
Large multimodal inputs | GPT-6 Astra | Muse Spark 1.3, DeepSeek V4.1 Flash and Gemini 3.7 Flash | Current ranked options with documented mixed-media and long-context support |
Lower hosted token prices | Gemini 3.7 Flash | Grok 4.6 and GLM-5.3 | Start with documented rates, then measure cost per accepted result |
Restricted cyber or life-science work | Claude Mythos 5.1 | Claude Mythos 5 | Specialist safeguards and restricted access |
Local or private deployment experiments | DeepSeek V4.1 Flash | gpt-oss-120b, Gemma 4, Mistral Large 3 and Llama 4 | Downloadable weights with different infrastructure requirements |
Complete directory of 39 AI models
The table below is ordered by the current ranked top eight first, followed by the broader directory. Coverage group is a navigation label, not a hidden score or capability tier.
Model | Provider | Coverage group | Availability | License | Context | API price or deployment cost |
|---|---|---|---|---|---|---|
Anthropic | Current ranked top eight | Available | Proprietary | 1M tokens | $10 input / $50 output per million tokens; $0.25 cache reads | |
Anthropic | Current ranked top eight | Active on the Claude API and supported cloud platforms | Proprietary | 1 million tokens | $4 input and $20 output per million tokens | |
Anthropic | Current ranked top eight | Available | Proprietary | 1M tokens | $10 input / $50 output per million tokens | |
OpenAI | Current ranked top eight | Rolling out through the OpenAI API and eligible ChatGPT plans | Proprietary | 1,050,000 tokens | $10 input and $50 output per million tokens at Standard rates | |
Meta | Current ranked top eight | Available through Muse Code and Meta Model API; max reasoning was still pending additional safety testing at launch | Proprietary hosted model | 1 million tokens | Standard endpoint: $1.25 input, $0.15 cached input and $4.25 output per million tokens; contributor endpoint: $0.10 input, $0.002 cached input and $0.20 output | |
DeepSeek | Current ranked top eight | Available through the DeepSeek API and downloadable open weights | MIT License for the released model weights and repository | 1 million tokens | See the guide for current hosted pricing and self-hosting costs | |
OpenAI | Current ranked top eight | Available | Proprietary | 1.05M tokens | $4 input / $20 output per million tokens, promotional | |
OpenAI | Current ranked top eight | Available | Proprietary | 1.05M tokens | $5 input / $30 output per million tokens | |
Anthropic | Broader current directory | Available | Proprietary | 1M tokens | $5 input / $25 output per million tokens | |
Moonshot AI | Open weights | Available; open weights | Open weights | 1M tokens | $3 cache-miss input / $15 output per million tokens | |
Broader current directory | Available | Proprietary | 1,048,576 tokens | $0.75 input / $3.75 output per million tokens, introductory | ||
Anthropic | Broader current directory | Available | Proprietary | 200,000 tokens | $1 input and $5 output per million tokens | |
Anthropic | Restricted specialist | Restricted access | Proprietary | 1 million tokens | Shares Fable 5 pricing: $10 input and $50 output per million tokens | |
Anthropic | Restricted specialist | Restricted access | Proprietary | 1 million tokens | $10 input and $50 output per million tokens; access is limited to vetted organizations | |
Anthropic | Legacy or migration reference | Legacy but still documented | Proprietary | 1 million tokens | $5 input and $25 output per million tokens at launch | |
Anthropic | Broader current directory | Available | Proprietary | 1 million tokens | $5 input and $25 output per million tokens; fast mode uses separate pricing | |
Anthropic | Broader current directory | Available | Proprietary | 1 million tokens | $3 input and $15 output per million tokens | |
Anthropic | Broader current directory | Available | Proprietary | 1 million tokens | $2 input and $10 output per million tokens; permanent pricing as of August 10, 2026 | |
OpenAI | Broader current directory | Available; previous flagship | Proprietary | 400,000 tokens | $1.75 input, $0.175 cached input and $14 output per million tokens | |
OpenAI | Broader current directory | Available | Proprietary | 400,000 tokens | Usage follows current Codex and API pricing; check the OpenAI model card for the active rate | |
OpenAI | Broader current directory | Available | Proprietary | 1,050,000 tokens | $2.50 input, $0.25 cached input and $15 output per million tokens; long-context multipliers apply above 272K input tokens | |
OpenAI | Open weights | Available; downloadable weights | Open weights under Apache 2.0 | 131,072 tokens | No OpenAI token fee for self-hosting; infrastructure and hosted-provider rates vary | |
Broader current directory | Available to Google AI Ultra subscribers | Proprietary | 1 million tokens | Included with eligible Google AI Ultra access; no standalone token price | ||
Preview | Available in selected products; API preview retired in favor of 3.1 Pro | Proprietary | 1,048,576 tokens | Historical API pricing was tiered; use Gemini 3.1 Pro pricing for new integrations | ||
Preview | Preview | Proprietary | 1,048,576 tokens | Tiered Gemini API pricing varies by prompt length and feature; verify the current Google pricing table | ||
Open weights | Available; downloadable model family | Open weights under the Gemma license | Up to 256,000 tokens depending on variant | No Google token fee for self-hosted weights; infrastructure and hosted-provider rates vary | ||
xAI | Broader current directory | Available | Proprietary | 1 million tokens | $1.25 input and $2.50 output per million tokens | |
xAI | Broader current directory | Available | Proprietary | 500,000 tokens | $2 input, $0.30 cached input and $6 output per million tokens below 200K prompt tokens; higher long-context rates apply | |
xAI | Broader current directory | Available; outside current top eight | Proprietary | 500K tokens | $2 input / $6 output per million tokens | |
Alibaba | Open weights | Available; open weights | Open weights | 1M tokens | $2 input / $6 output per million tokens | |
Alibaba | Preview | Available in preview | Proprietary hosted model | 1 million tokens | $2.50 input and $7.50 output per million tokens in the published QwenCloud tier | |
Moonshot AI | Open weights | Available | Open weights with hosted API access | 256,000 tokens | Self-hosting cost varies; check Kimi Platform for current hosted API rates | |
Z.ai | Open weights | Available; open weights | Open weights | 1 million tokens | Open-weight self-hosting plus Z.ai plans from the published subscription tier | |
Z.ai | Open weights | Available through Z.ai APIs and coding plans | Open weights | 1 million tokens | Plan and API pricing varies; Z.ai does not publish one universal rate on the model overview | |
Meta | Open weights | Available; open-weight Scout and Maverick | Open weights under the Llama license | Up to 10 million tokens for Scout; variant-specific for Maverick | No Meta token fee for weights; hosting and provider rates vary | |
Meta | Broader current directory | Available through Meta Model API access | Proprietary hosted model | 1 million tokens | Meta Model API pricing varies by access plan; verify the active console rate | |
DeepSeek | Open weights | Available; V4 Pro and V4 Flash | Open weights for the preview family; API variants available | 1 million tokens | V4 Pro peak rates: $0.044 cached input, $1.32 uncached input and $3.96 output per million tokens; off-peak is half | |
MiniMax | Open weights | Available; open-weight release announced | Open weights | Up to 1 million tokens; minimum 512K guaranteed by API documentation | Standard API up to 512K: $0.30 input, $0.06 cache read and $1.20 output per million tokens; rates double above 512K | |
Mistral AI | Open weights | Available; open weights | Open weights under Apache 2.0 | 256,000 tokens | $0.50 input and $1.50 output per million tokens on Mistral’s hosted API |
Pricing is quoted from each model guide as verified through September 28, 2026. Cache rates, batch discounts, long-context multipliers, tools, regional pricing and self-hosting costs can materially change the result.
Directory composition
Coverage group | Count | What it means |
|---|---|---|
Current ranked top eight | 8 | Models with a current verified Overall rank and score |
Open weights | 11 | Broader downloadable or self-hostable models outside the ranked eight |
Restricted specialist | 2 | Models limited to vetted programs or organizations |
Broader current or preview directory | 17 | Current proprietary or preview models with dedicated guides but no Overall rank here |
Legacy or migration reference | 1 | Older documented models retained for migration and historical comparison |
Provider | Models covered | Examples |
|---|---|---|
11 | Claude Fable 5.1, Claude Opus 5.5, Claude Fable 5, and more | |
7 | GPT-6 Astra, GPT 5.6 Sol, GPT 5.5, and more | |
5 | Gemini 3.7 Flash, Gemini 3 Deep Think, Gemini 3 Pro, and more | |
3 | Grok 4.3, Grok 4.5, Grok 4.6 | |
2 | Qwen 3.8, Qwen3.7-Max | |
2 | Kimi K3, Kimi K2.6 | |
2 | GLM-5.2, GLM-5.3 | |
3 | Muse Spark 1.3, Llama 4, Muse Spark 1.1 | |
2 | DeepSeek V4.1 Flash, DeepSeek V4 | |
1 | MiniMax M3 | |
1 | Mistral Large 3 |
Browse models by provider
Every model below has a dedicated guide with specifications, pricing, availability, limitations, official sources and an evaluation framework.
Anthropic: 11 models
Anthropic coverage includes Claude Fable 5.1, Claude Opus 5.5, Claude Fable 5, Claude Opus 5, Claude Haiku 4.5, Claude Mythos 5, Claude Mythos 5.1, Claude Opus 4.7, Claude Opus 4.8, Claude Sonnet 4.6, Claude Sonnet 5. Compare the models within this provider before comparing across providers, because access routes, naming, effort settings and pricing conventions often differ.
Claude Opus 5.5: Active on the Claude API and supported cloud platforms; Proprietary; 1 million tokens context. Best fit: Long-running agentic coding, computer use and knowledge work. Ranked #2 with an Overall score of 83.2.
Claude Fable 5.1: Available; Proprietary; 1M tokens context. Best fit: Long-running agents, coding, research and complex professional work. Ranked #1 with an Overall score of 83.4.
Claude Fable 5: Available; Proprietary; 1M tokens context. Best fit: Long-running coding, research and complex knowledge work. Ranked #3 with an Overall score of 83.0.
Claude Opus 5: Available; Proprietary; 1M tokens context. Best fit: Coding, computer use and analytical work. This model is covered but not assigned an Overall rank on this page.
Claude Haiku 4.5: Available; Proprietary; 200,000 tokens context. Best fit: high-volume classification, coding subagents, chat and low-latency workflows. This model is covered but not assigned an Overall rank on this page.
Claude Mythos 5: Restricted access; Proprietary; 1 million tokens context. Best fit: advanced cybersecurity and life-sciences research for approved organizations. This model is covered but not assigned an Overall rank on this page.
Claude Mythos 5.1: Restricted access; Proprietary; 1 million tokens context. Best fit: advanced defensive cybersecurity, protein design, computational biology and specialist scientific work. This model is covered but not assigned an Overall rank on this page.
Claude Opus 4.7: Legacy but still documented; Proprietary; 1 million tokens context. Best fit: reproducing older workflows, migration testing and historical Opus comparisons. This model is covered but not assigned an Overall rank on this page.
Claude Opus 4.8: Available; Proprietary; 1 million tokens context. Best fit: coding agents, professional analysis, computer use and long-running enterprise workflows. This model is covered but not assigned an Overall rank on this page.
Claude Sonnet 4.6: Available; Proprietary; 1 million tokens context. Best fit: coding, computer use, agent planning and professional knowledge work. This model is covered but not assigned an Overall rank on this page.
Claude Sonnet 5: Available; Proprietary; 1 million tokens context. Best fit: high-volume agents, coding, browser work and cost-conscious professional automation. This model is covered but not assigned an Overall rank on this page.
OpenAI: 7 models
OpenAI coverage includes GPT-6 Astra, GPT 5.6 Sol, GPT 5.5, GPT-5.2, GPT-5.3-Codex, GPT-5.4, gpt-oss-120b. Compare the models within this provider before comparing across providers, because access routes, naming, effort settings and pricing conventions often differ.
GPT-6 Astra: Rolling out through the OpenAI API and eligible ChatGPT plans; Proprietary; 1,050,000 tokens context. Best fit: Complex reasoning, coding, computer use, research and end-to-end professional work. Ranked #4 with an Overall score of 82.2.
GPT 5.6 Sol: Available; Proprietary; 1.05M tokens context. Best fit: Reasoning, tool-using agents, coding and large repositories. Ranked #7 with an Overall score of 81.0.
GPT 5.5: Available; Proprietary; 1.05M tokens context. Best fit: Professional work, coding, analysis and adjustable reasoning. Ranked #8 with an Overall score of 80.2.
GPT-5.2: Available; previous flagship; Proprietary; 400,000 tokens context. Best fit: professional work, reasoning, coding, vision and long-context analysis. This model is covered but not assigned an Overall rank on this page.
GPT-5.3-Codex: Available; Proprietary; 400,000 tokens context. Best fit: long-horizon agentic coding, repository work and technical professional tasks. This model is covered but not assigned an Overall rank on this page.
GPT-5.4: Available; Proprietary; 1,050,000 tokens context. Best fit: professional work, computer-use agents, coding and large-context analysis. This model is covered but not assigned an Overall rank on this page.
gpt-oss-120b: Available; downloadable weights; Open weights under Apache 2.0; 131,072 tokens context. Best fit: self-hosted reasoning, tool use and production workloads that fit one 80GB GPU. This model is covered but not assigned an Overall rank on this page.
Google: 5 models
Google coverage includes Gemini 3.7 Flash, Gemini 3 Deep Think, Gemini 3 Pro, Gemini 3.1 Pro, Gemma 4. Compare the models within this provider before comparing across providers, because access routes, naming, effort settings and pricing conventions often differ.
Gemini 3.7 Flash: Available; Proprietary; 1,048,576 tokens context. Best fit: Fast multimodal agents and high-volume applications. This model is covered but not assigned an Overall rank on this page.
Gemini 3 Deep Think: Available to Google AI Ultra subscribers; Proprietary; 1 million tokens context. Best fit: difficult science, mathematics, research and engineering reasoning. This model is covered but not assigned an Overall rank on this page.
Gemini 3 Pro: Available in selected products; API preview retired in favor of 3.1 Pro; Proprietary; 1,048,576 tokens context. Best fit: historical multimodal reasoning comparisons and migration testing. This model is covered but not assigned an Overall rank on this page.
Gemini 3.1 Pro: Preview; Proprietary; 1,048,576 tokens context. Best fit: multimodal reasoning, complex problem solving, software engineering and large mixed-media inputs. This model is covered but not assigned an Overall rank on this page.
Gemma 4: Available; downloadable model family; Open weights under the Gemma license; Up to 256,000 tokens depending on variant context. Best fit: open multimodal agents, on-device work and size-flexible deployment. This model is covered but not assigned an Overall rank on this page.
xAI: 3 models
xAI coverage includes Grok 4.3, Grok 4.5, Grok 4.6. Compare the models within this provider before comparing across providers, because access routes, naming, effort settings and pricing conventions often differ.
Grok 4.3: Available; Proprietary; 1 million tokens context. Best fit: cost-efficient reasoning, tool calling and large-context applications. This model is covered but not assigned an Overall rank on this page.
Grok 4.5: Available; Proprietary; 500,000 tokens context. Best fit: coding, office automation, technical research and tool-using agents. This model is covered but not assigned an Overall rank on this page.
Grok 4.6: Available; outside current top eight; Proprietary; 500K tokens context. Best fit: Cost-sensitive agents, coding and partner workflows. This model is covered but not assigned an Overall rank on this page.
Alibaba: 2 models
Alibaba coverage includes Qwen 3.8, Qwen3.7-Max. Compare the models within this provider before comparing across providers, because access routes, naming, effort settings and pricing conventions often differ.
Qwen 3.8: Available; open weights; Open weights; 1M tokens context. Best fit: Open-weight coding, multimodal agents and deployment control. This model is covered but not assigned an Overall rank on this page.
Qwen3.7-Max: Available in preview; Proprietary hosted model; 1 million tokens context. Best fit: hosted multimodal reasoning, coding and long-context agents. This model is covered but not assigned an Overall rank on this page.
Moonshot AI: 2 models
Moonshot AI coverage includes Kimi K3, Kimi K2.6. Compare the models within this provider before comparing across providers, because access routes, naming, effort settings and pricing conventions often differ.
Kimi K3: Available; open weights; Open weights; 1M tokens context. Best fit: Open-weight frontier work, coding, research and self-hosting. This model is covered but not assigned an Overall rank on this page.
Kimi K2.6: Available; Open weights with hosted API access; 256,000 tokens context. Best fit: fast question answering, coding, multimodal understanding and agent tool use. This model is covered but not assigned an Overall rank on this page.
Z.ai: 2 models
Z.ai coverage includes GLM-5.2, GLM-5.3. Compare the models within this provider before comparing across providers, because access routes, naming, effort settings and pricing conventions often differ.
GLM-5.2: Available; open weights; Open weights; 1 million tokens context. Best fit: long-horizon coding, automated research and open deployment. This model is covered but not assigned an Overall rank on this page.
GLM-5.3: Available through Z.ai APIs and coding plans; Open weights; 1 million tokens context. Best fit: agentic coding, long-horizon software engineering, vulnerability discovery and text-based tool use. This model is covered but not assigned an Overall rank on this page.
Meta: 3 models
Meta coverage includes Muse Spark 1.3, Llama 4, Muse Spark 1.1. Compare the models within this provider before comparing across providers, because access routes, naming, effort settings and pricing conventions often differ.
Muse Spark 1.3: Available through Muse Code and Meta Model API; max reasoning was still pending additional safety testing at launch; Proprietary hosted model; 1 million tokens context. Best fit: long-horizon agentic workflows, coding, multimodal analysis and tool-assisted professional work. Ranked #5 with an Overall score of 81.6.
Llama 4: Available; open-weight Scout and Maverick; Open weights under the Llama license; Up to 10 million tokens for Scout; variant-specific for Maverick context. Best fit: open multimodal deployment, long-context research and custom private systems. This model is covered but not assigned an Overall rank on this page.
Muse Spark 1.1: Available through Meta Model API access; Proprietary hosted model; 1 million tokens context. Best fit: agentic coding, interface debugging and long-context multimodal work. This model is covered but not assigned an Overall rank on this page.
DeepSeek: 2 models
DeepSeek coverage includes DeepSeek V4.1 Flash, DeepSeek V4. Compare the models within this provider before comparing across providers, because access routes, naming, effort settings and pricing conventions often differ.
DeepSeek V4.1 Flash: Available through the DeepSeek API and downloadable open weights; MIT License for the released model weights and repository; 1 million tokens context. Best fit: Coding agents, long-context workflows, vision tasks and cost-sensitive inference. Ranked #6 with an Overall score of 81.1.
DeepSeek V4: Available; V4 Pro and V4 Flash; Open weights for the preview family; API variants available; 1 million tokens context. Best fit: low-cost long-context agents, coding, flexible reasoning and self-hosting experiments. This model is covered but not assigned an Overall rank on this page.
MiniMax: 1 model
MiniMax coverage includes MiniMax M3. Compare the models within this provider before comparing across providers, because access routes, naming, effort settings and pricing conventions often differ.
MiniMax M3: Available; open-weight release announced; Open weights; Up to 1 million tokens; minimum 512K guaranteed by API documentation context. Best fit: open-weight multimodal agents, long-context coding, browsing and private deployment experiments. This model is covered but not assigned an Overall rank on this page.
Mistral AI: 1 model
Mistral AI coverage includes Mistral Large 3. Compare the models within this provider before comparing across providers, because access routes, naming, effort settings and pricing conventions often differ.
Mistral Large 3: Available; open weights; Open weights under Apache 2.0; 256,000 tokens context. Best fit: cost-efficient open multimodal work, private deployment and general-purpose agents. This model is covered but not assigned an Overall rank on this page.
Available, restricted, preview and legacy labels
Label | Meaning | What to verify |
|---|---|---|
Available | The provider documents a current access route | Region, plan, endpoint, snapshot and quotas |
Restricted | Access is limited to vetted users or programs | Eligibility, safeguards and data-handling terms |
Preview | The model or endpoint can change before general availability | Version pinning, deprecation policy and regression tests |
Open weights | Weights can be downloaded or deployed through compatible infrastructure | Exact license, checkpoint, quantization, hardware and serving stack |
Legacy or reference | Retained for migration, reproduction or historical comparison | Retirement date, replacement model and compatibility changes |
How to compare models fairly
Choose a shortlist of two to five models that fit your access, licensing and modality requirements. Then test them with the same tasks, tool permissions, context, retry limits and acceptance rubric. Record the exact model identifier and reasoning settings.
Measure | Record | Why it matters |
|---|---|---|
Task completion | Pass rate across representative and adversarial tasks | A benchmark score is not your production success rate |
Reliability | Repeated-run variance and recovery after errors | One strong run can hide instability |
Quality | Factuality, instruction adherence and reviewer acceptance | Long or confident output is not necessarily correct |
Agent behavior | Tool errors, unsafe actions, retries and approval requests | Base-model quality depends on the harness and permissions |
Latency | Time to first token and end-to-end completion time | Fast generation can still produce slow multi-step work |
Total cost | Tokens, caching, tools, retries, infrastructure and human review | Token price alone can reverse the real value order |
Governance | Retention, training policy, regional processing and auditability | The best technical score may not fit the deployment requirements |
Ranking and evidence methodology
The ranked top eight comes from The AI Leaderboard’s current validated dataset. Overall gives equal weight to seven evaluation-category averages. Model names use canonical family labels, while the exact tested variant and effort setting are retained during source validation.
The broader 31-model directory does not receive invented scores. Its specifications come from each published model guide, which uses first-party provider documentation and labels vendor-reported benchmark evidence. Inclusion means the model has a verified page, not that it outranks a model in the top eight.
Page element | Evidence used | Limitation |
|---|---|---|
Top-eight rank and Overall | Current validated leaderboard records | Not a production success rate or universal task ranking |
Specifications and pricing | First-party documentation summarized in each guide | Can change after the verification date |
Availability and license | Provider model cards, launch pages and access documentation | Terms can differ by region and distribution partner |
Shortlists by need | Editorial synthesis of verified rank, specialization, access and licensing | Must be tested on the reader’s own workload |
Broader directory order | Ranked eight first, then provider and model name | No hidden score is implied |
Read how we rank for the site-wide methodology and evidence policy.
Last verified September 28, 2026. The page should be refreshed whenever the ranked dataset, provider specifications, pricing, access or model inventory changes.
Frequently asked questions
What is the best AI model right now?
Claude Fable 5.1 is number one in the current verified ranking with an Overall score of 83.4. That makes it the benchmark leader in this dataset, not the automatic winner for every workload.
How many AI models are included?
The directory includes 39 published model guides across 11 providers. Eight have current verified Overall ranks, while 31 are broader current, specialist, open-weight, preview or migration-reference entries.
Why are only eight models ranked?
Only the eight validated records in the current leaderboard dataset receive an Overall rank. The broader directory is intentionally unranked until comparable source data is available.
What is the best open-weights AI model?
DeepSeek V4.1 Flash is the highest-ranked open-weights model in the current top eight, at number 6 with an Overall score of 81.1. Kimi K3, DeepSeek V4, MiniMax M3, Qwen 3.8, Llama 4, gpt-oss-120b, Gemma 4 and Mistral Large 3 are also covered for different deployment needs.
Are Overall scores percentages?
No. An Overall score of 83.4 does not mean the model is correct 83.4 percent of the time. It is a comparative score produced from the current evaluation categories.
How should I choose between proprietary and open weights?
Use a proprietary API when managed infrastructure and fast integration matter most. Evaluate open weights when deployment control, version stability, data locality or customization justify the operational burden. Compare total cost, not only token or download price.
How often is this page updated?
The ranked section follows the leaderboard refresh cycle. The broader directory should also be updated whenever a provider changes a model’s availability, pricing, license or technical limits.