OpenRouter
OpenRouter free model variants
One OpenAI-compatible API for a rotating pool of zero-price model variants.
Zero-price routesAccount
Shared free-tier limits; model availability changes.
A live, free-first catalogue that joins model availability, official access links and benchmark evidence — so you can choose the best service for your actual task.
Live directory
Free access firstCompare free routes across official gateways, web apps and local runtimes. Scores are shown only when a live benchmark source has them; a dash means “not measured”, not “bad”.
Free rows
0
across live feeds
Direct access
0
API, chat or local
Scored
0
with an intelligence index
Visible
0
of 0 free rows
Free does not mean unlimited. Provider rate limits, sign-in requirements, retention policies and model availability can change. Check the linked official service before sending sensitive or production data. Prices shown for paid context rows are approximate input cost per one million tokens, not a full task quote.
More ways to use AI
These are first-party destinations, not anonymous “free API” aggregators. Quotas and eligibility change, so the directory links to the provider’s current terms instead of promising a fixed limit.
OpenRouter free model variants
One OpenAI-compatible API for a rotating pool of zero-price model variants.
Shared free-tier limits; model availability changes.
OpenCode Zen free models
Official Zen gateway used by OpenCode with a small, changing free catalogue.
Gateway limits and eligibility are provider-controlled.
CommandCode free gateway models
OpenAI-compatible gateway with a published free-model list.
Gateway limits can change without notice.
Kilo gateway free models
Coding gateway that publishes the models currently available through its free route.
Gateway and provider rate limits apply.
Cline provider gateway
Cline’s provider endpoint; its catalogue can include free model routes.
Availability and limits are controlled by Cline and upstream providers.
Google AI Studio / Gemini API
First-party Gemini playground and developer API with a changing limited free tier.
Model- and region-specific rate limits; confirm the current quota in AI Studio.
GroqCloud free tier
Very fast hosted open-weight models through an OpenAI-compatible API.
Per-model requests and token limits; see the live limits page.
Cerebras Inference
Fast hosted open-weight inference with a limited developer tier.
Quota varies by account and model.
OpenRouter’s live catalogue is useful when you want one API key and several free routes.
Benchmark context
A capability signal for ranking models — not a guarantee of free availability or task quality.
Knowledge Work
25%
Real professional deliverables, business workflow automation and long-document reasoning.
Coding
25%
Changes to real repositories that a maintainer would merge, plus terminal-driven engineering.
Science
25%
Terminal-based research workflows and research-level physics reasoning.
General
25%
Factual recall, refusal to hallucinate, and an exceptionally hard knowledge/reasoning exam.
The index helps shortlist capability. The free-access links tell you where the model can actually be used.
Intelligence Index v5.0.1 blends 10 weighted benchmarks across 4 categories; the weights below sum to 100%. A model with no score is a new or specialised listing, not an automatic failure.
Treat the overall index as a starting point. Coding, tool use, vision, long-context reading and latency can disagree; use the dedicated sub-scores and capability filter for your workload.
Methodology
10 benchmarks feed the Intelligence Index; 14 standalone and 13 retired evaluations are shown for context only.
| Benchmark | Category | Scale | Weight | What it tests |
|---|---|---|---|---|
| AA-Briefcase v1.1 | Knowledge Work | 91 tasks · 4 scenarios | 15% | Multi-week knowledge-work projects graded by binary rubrics and three-judge pairwise comparisons on task success, analytical quality and presentation quality. |
| AutomationBench-AA | Knowledge Work | 657 SaaS workflow tasks | 5% | Realistic business workflows across six domains, driven through REST API tools; a task scores zero if the model breaks a guardrail. |
| GDP.pdf | Knowledge Work | 100 tasks · 4,592 PDF pages | 5% | Long professional documents answered in a single turn and graded all-pass against 1,275 rubric criteria. |
| FrontierCode-AA 2.0 | Coding | 132 repo tasks · 3 attempts | 15% | Changes to real open-source repositories, graded by held-out tests plus blocking and non-blocking rubric checks; failing any blocking check scores zero. |
| Terminal-Bench 4.0 | Coding | 66 terminal tasks · 3 repeats | 10% | Terminal work across software engineering, machine learning, science, operations, security and hardware — every task test must pass. |
| Terminal-Bench-Science 0.1 | Science | 70 tasks · 5 science domains | 15% | Research workflow tasks drawn from real scientific practice (life, physical, mathematical, engineering and earth sciences), graded by each task's own test suite. |
| CritPt v2 (Hard Subset) | Science | 40 of 70 physics problems | 10% | Research-level physics answered as Python functions, symbolic expressions or numeric values and graded by the official server; the hard subset drops the problems every frontier model already solved. |
| AA-Omniscience Accuracy | General | 3,000 questions · 42 topics | 10% | Factual recall on economically relevant questions, scored as the share of questions answered correctly. |
| AA-Omniscience Non-Hallucination | General | 3,000 questions · 42 topics | 5% | One minus the hallucination rate: abstaining stays neutral, a confident wrong answer costs points. |
| HLE-Diamond | General | 1,000 questions (500 reasoning / 500 knowledge) | 10% | A cleaned 1,000-question rebuild of Humanity's Last Exam, graded with the official judge prompt; 272 questions carry images. |
| Benchmark | Category | Scale | What it tests |
|---|---|---|---|
| GDPval-AA v2.1 | Knowledge Work | 220 tasks · 44 occupations | Economically valuable knowledge work across occupations and GDP sectors, Elo-ranked by a three-judge panel. Standalone since v5.0 — it was an index constituent up to v4.3.2. |
| Harvey LAB-AA v1.1 | Knowledge Work | 120 legal tasks | Agentic legal deliverable production, scored by a hallucination-gated all-pass rate: any material hallucination zeroes the task. |
| APEX-Agents-AA | Knowledge Work | 452 tasks | Long-horizon multi-step professional tasks in banking, consulting and law: plan, act, check, recover and deliver. |
| AA-AnalystAgent | Knowledge Work | 80 tasks · 14 domains | Agentic Python analysis with a free-form final answer, judged for binary correctness with a numeric pre-check that can override the judge. |
| EnterpriseOps-Gym-AA | Knowledge Work | 1,117 tasks · 8 domains | Multi-turn MCP tool use against resettable enterprise systems, verified by SQL state checks on the outcome. |
| MKA | Knowledge Work | 200 tasks | Agentic knowledge work over complex visual files, built with Google DeepMind and graded by a rubric-following judge. |
| ITBench-AA | Coding | 59 Kubernetes incidents | Root-cause diagnosis from offline Kubernetes incident snapshots, scored by entity matching at full recall. |
| Code2Video-AA | Coding | 114 tasks | HTML composition that is rendered to video and compared pairwise by a judge. |
| MLCR-AA | Science | 60 medical long-context questions | Medical long-context reasoning graded by a three-judge majority: complete, and no claim contradicting the record. |
| HealthBench / Professional | Science | 5,000 / 525 conversations | Free-text health replies scored against length-adjusted rubrics by a judge model. |
| CritPt v2 (full set) | Science | 70 problems · 5 attempts | The complete CritPt v2 physics set; the Intelligence Index uses the 40-problem hard subset of this evaluation. |
| IFBench | General | 294 tasks | Following verifiable formatting and constraint instructions precisely. |
| Global-MMLU-Lite | General | ~6,000 · 16 languages | Multilingual multiple-choice knowledge; the basis of the separate Artificial Analysis Multilingual Index. |
| MMMU Pro | General | 1,730 questions | College-level questions requiring image and text reasoning, reported separately from the index. |
| Benchmark | Category | Scale | What it tests |
|---|---|---|---|
| SciCode | Coding | 80 problems · 338 subproblems | Scientific code generation, retired from the index in v5.0 when FrontierCode-AA took the coding slot. |
| Terminal-Bench 2.1 | Coding | Earlier release | The previous Terminal-Bench release; Terminal-Bench 4.0 is the v5.0.1 component. |
| SWE-bench | Coding | 2,294 tasks | Real-world software-engineering tasks derived from GitHub issues across popular Python repositories. |
| LiveCodeBench | Coding | Continuous | Fresh contest problems from LeetCode, AtCoder and Codeforces designed to reduce contamination. |
| AA-LCR v1.1 | General | 10k–100k token documents | Extracting, reasoning about and synthesising information from long-form documents. Last in the index in v4.3. |
| Humanity's Last Exam (full) | General | 2,158 text-only questions | The full HLE text set; HLE-Diamond replaced it as the v5.0.1 component, so the two scores are not comparable. |
| GPQA Diamond | Science | Expert MCQ set | Expert-written Google-proof questions in biology, chemistry and physics. |
| τ³-Banking | Knowledge Work | Multi-turn | Tool use and knowledge retrieval in simulated banking customer-support flows; retired from the index in v4.3. |
| BrowseComp | General | 1,266 questions | Multi-hop hard web research and evidence gathering. |
| MATH-500 | Science | 500 competition problems | Competition mathematics; saturated by current frontier models. |
| AIME 2025 | Science | 30 problems | The 2025 American Invitational Mathematics Examination problem set. |
| MMLU-Pro | General | Multi-task MCQ set | Broad academic knowledge as ten-option multiple choice. |
| SimpleQA | General | Short factual questions | Short factual questions with stable ground-truth answers, designed to expose hallucination. |
Useful live catalogue views when you want one API key for several free routes. Always verify the current price marker before making a request.
Free · highest intelligence
Best first place to scan for general reasoning.
Free · highest coding
Start here for code generation and software work.
Free · highest agentic
Useful for tools, plans and multi-step tasks.
All free model routes
Check pricing and limits immediately before use.
Most popular overall
Useful context, but popularity is not a quality score.
Newest models
New models may appear before benchmark scores land.