The Best Free Model on OpenRouter Right Now

Start with our highest-priority model that clears the free-endpoint performance floor.

Complete-day usage not available
Free-model limits
Overall #1Our predicted leader

Inkling Small

Named · Benchmark evidence incomplete

Current catalog intelligence/coding/agentic signals of 25.7/52.9/23.5 keep it ahead of Inkling's 25/52.1/22.5 in our trial-priority order. Its free route clears all three performance floors and remains the first passing prediction. These signals are not mixed into the historical formal pool.

Daily tokens
Not available
Context
1.05M
Tool calling
Not listed
Input
text + image + audio
Latency P50 · 30min
1.22 s
Throughput P50 · 30min
123 token/s
Availability · 24h
97.90%
Download AgenaxyView on OpenRouter
thinkingmachines/inkling-small:free
Endpoint notes

The free endpoint lists tools but not tool_choice, so full tool-control support is not verified. JSON-schema and response-format controls are not listed. Complete same-version evidence for the retained formal pool is unavailable.

Observed rolling 24h availability: 97.90%. No shared complete UTC usage day is available; performance and availability use their own rolling windows.

Bring your OpenRouter connection to Agenaxy and try it on your own work. Free-model quotas and provider policies still apply.

The current rankings

How we choose

Up to three models that clear the performance floor, in predicted-strength order. Anonymous models can lead.

  1. 01Inkling SmallNamed · Benchmark evidence incompleteDaily tokensNot availableContext1.05M

    Current catalog intelligence/coding/agentic signals of 25.7/52.9/23.5 keep it ahead of Inkling's 25/52.1/22.5 in our trial-priority order. Its free route clears all three performance floors and remains the first passing prediction. These signals are not mixed into the historical formal pool.

    The free endpoint lists tools but not tool_choice, so full tool-control support is not verified. JSON-schema and response-format controls are not listed. Complete same-version evidence for the retained formal pool is unavailable.

    Input
    text + image + audio
    Tool calling
    Not listed
    Structured output
    Not listed
    Availability · rolling 24h
    97.90%
    Latency P50 · 30min
    1.22 s
    Throughput P50 · 30min
    123 token/s
    Latency P90 · 30min
    2.36 s
    Performance samples · 30min
    4,839
    Requests · complete UTC day
    Not available
    Provider
    Thinking Machines
    Max output · API limit
    262,144
    Free route expiration
    Not listed
    Availability · rolling 3d
    98.49%

    API sources: Performance · Availability · Daily activity · Endpoint. Observed 2026-10-19T23:02:48.489Z. Performance clearance: pass.

    thinkingmachines/inkling-small:free
    OpenRouter
  2. 02InklingNamed · Benchmark evidence incompleteDaily tokensNot availableContext1.05M

    Current catalog intelligence/coding/agentic signals of 25/52.1/22.5 support an earlier trial priority than Ultra. It clears all three floors and follows Inkling Small as the manual backup. Image and audio input with 1M context; this is editorial judgment, not a common formal score.

    P90 latency is 6.15 seconds versus a 1.919-second P50; long-tail waits remain substantially above the median. P90 is disclosed, not an extra gate. Current catalog indices lack verified same-version comparability with the retained formal pool. The endpoint lists tools but not tool_choice; JSON output controls are absent. Declared capabilities do not establish successful task execution.

    Input
    text + image + audio
    Tool calling
    Not listed
    Structured output
    Not listed
    Availability · rolling 24h
    99.27%
    Latency P50 · 30min
    1.92 s
    Throughput P50 · 30min
    32 token/s
    Latency P90 · 30min
    6.15 s
    Performance samples · 30min
    9,224
    Requests · complete UTC day
    Not available
    Provider
    Thinking Machines
    Max output · API limit
    262,144
    Free route expiration
    Not listed
    Availability · rolling 3d
    99.40%

    API sources: Performance · Availability · Daily activity · Endpoint. Observed 2026-10-19T23:02:48.489Z. Performance clearance: pass.

    thinkingmachines/inkling:free
    OpenRouter
  3. 03MiniMax M3.1 FlashNamed · Benchmark evidence incompleteDaily tokensNot availableContext1M

    A newly available MiniMax multimodal model with vendor-reported improvements over M3, 1M context, text/image/video input, 524K maximum output and declared tools/response-format controls. These support an aggressive early trial priority after the two benchmark-signaled Inkling candidates, not independently proven superior intelligence. Performance is an admission floor, not a strength score. Its exact free route clears all three performance floors and is the third admitted prediction.

    New free route with limited operating history and no verified complete same-version three-index evidence for the retained formal pool. Tool choice is advertised but the provider declares only auto, not forced function/required/none modes. Response format is listed; structured_outputs is not. Vendor claims and declared capabilities do not prove successful task execution.

    Input
    text + image + video
    Tool calling
    Supported
    Structured output
    Not listed
    Availability · rolling 24h
    99.84%
    Latency P50 · 30min
    1.28 s
    Throughput P50 · 30min
    103 token/s
    Latency P90 · 30min
    3.62 s
    Performance samples · 30min
    400,468
    Requests · complete UTC day
    Not available
    Provider
    Minimax
    Max output · API limit
    524,288
    Free route expiration
    Not listed
    Availability · rolling 3d
    99.86%

    API sources: Performance · Availability · Daily activity · Endpoint. Observed 2026-10-19T23:02:48.489Z. Performance clearance: pass.

    minimax/minimax-m3.1-flash:free
    OpenRouter

Complete-day usage is not available: no completed UTC day is shared across all candidates. Partial current-day counts are not substituted. This is a curated shortlist, not the full catalog. Expand a model for reasoning, limitations and sources.

Choose another model only if…

  • You need image input.Inkling Small is our first admitted image-capable choice.
  • Your first choice is unavailable.Switch manually to Inkling. Observed availability is not a guarantee.
  • Your data policy requires it.Check the exact provider's policy. Free does not mean private; your requirements take priority.

Check required API features before connecting any model; declared tool or JSON-schema support does not prove correct task execution.

How we choose

Pick the model we predict is strongest overall, and use it by default. Only split the recommendation for multimodal input. Keep a backup for availability, not a separate specialist for every task.

Recent adoption, usable context and capabilities can justify a first try before a complete benchmark set exists. This is a forward-looking editorial judgment, not proof that a model is smarter.

Performance floor: availability ≥95% over 24 hours, latency P50 ≤10 seconds and throughput P50 ≥20 token/s over 30 minutes. All three must pass. Check 5–10 candidates in strength order and keep the first three that pass; no speed-based reordering or weighted score. Missing data awaits verification. If fewer qualify, recommend fewer.

Candidate performance screening
  • Inkling Small: Admitted — 97.90% availability; 1.22 s latency; 123 token/s throughput.
  • Inkling: Admitted — 99.27% availability; 1.92 s latency; 32 token/s throughput.
  • MiniMax M3.1 Flash: Admitted — 99.84% availability; 1.28 s latency; 103 token/s throughput.
  • NVIDIA Nemotron 3.5 Super VL: Below performance floor — 89.88% availability; 1.6 s latency; 92 token/s throughput.
  • NVIDIA Nemotron 3 Ultra: Passes; later candidate — 97.68% availability; 1.73 s latency; 57 token/s throughput.
  • Dots 3 Note Preview: Passes; later candidate — 97.86% availability; 0.95 s latency; 68 token/s throughput.
  • Laguna S 2.1: Passes; later candidate — 99.75% availability; 1.67 s latency; 42 token/s throughput.
  • Gemma 4 31B: Passes; later candidate — 97.08% availability; 1.72 s latency; 21 token/s throughput.
  • Gemma 4 26B A4B: Passes; later candidate — 98.17% availability; 0.98 s latency; 40 token/s throughput.
  • NVIDIA Nemotron 3 Super: Passes; later candidate — 99.28% availability; 0.93 s latency; 46 token/s throughput.

P90 and sample count are disclosed, not extra thresholds. Formal benchmark membership is independent of this screen.

Daily signals, anonymous models & availability

Daily heat uses the latest complete UTC day available at check time, not a rolling 24-hour period. Token volume is an adoption signal, not request count or measured intelligence. Missing observations stay unknown. Context and supported parameters refer to the exact endpoint, not every deployment of a model family.

Anonymous models can rank first overall. The observation list separates identity from evidence: named models with incomplete comparable benchmarks belong here too. Endpoint features such as structured-output controls are disclosed as compatibility facts, not used to invent task-specialist recommendations.

Availability is OpenRouter's observed proportion of time inference was successfully served. The rolling 24-hour values were collected at 2026-10-19T23:02:48.489Z; exact bucket bounds are in the snapshot. They are not provider uptime, intelligence or a forecast. Capacity, load, rate limits and provider errors can affect them. The floor screens the current free route, not the underlying model's intelligence.

OX Alpha was revealed as GLM 5.3 Flash. That historical reveal is not an active anonymous recommendation in this snapshot.

Formal benchmark formula & complete calculation

The retained 2026-09-09 snapshot uses Artificial Analysis v4.1.1. The formula below applies only to the formal benchmark pool. It does not assign a score to Ultra, Dots or other models without complete comparable evidence.

General intelligence percentile40%
Agentic percentile25%
Coding percentile10%
Current OpenRouter :free-ID context15%
Input modalities10%

For each benchmark, percentile = values strictly below the model ÷ (pool size − 1) × 100. Ties share a percentile. Round each weighted benchmark component before summing. Context contributes 6 / 9 / 12 / 15 points at 128K / 256K / 512K / 1M. Images add six points; audio or video adds four. Equal totals sort by exact model ID. Data policy remains a separate filter.

ModelIntelligenceAgenticCodingContextInputTotal
Gemma 4 31B29.7 → P100 → 4014.4 → P100 → 2543.4 → P100 → 10262K → 9text + image + video → 1094
Gemma 4 26B A4B26.1 → P50 → 2011 → P50 → 1339.3 → P50 → 5262K → 9text + image + video → 1057
Nemotron 3 Super25.7 → P0 → 08.8 → P0 → 037.7 → P0 → 0262K → 9text → 09

Benchmark cells: raw index → pool percentile → weighted points. A low relative score is not a universal quality rating.

Formal-pool eligibility

  • Listed as a current :free model in the OpenRouter catalog snapshot
  • Supports tools and tool choice
  • Supports response_format or structured_outputs
  • Provides at least 128K of free-endpoint context
  • Includes all three published Artificial Analysis indices used by the formula

These are the retained formal pool's rules, not a ban on other overall picks. New benchmark versions must not be mixed into this comparison.

Sources & dated snapshots

Routine collection uses JSON APIs, not model webpages. Each expanded row links its exact API sources. Public frontend stats APIs are not a compatibility guarantee; invalid or missing evidence is not treated as zero or silently published as a member removal.

Formal inputs retrieved 2026-09-09T03:46:16Z. Models SHA-256: 83d4c142bffd7ca1119a5ddc46cef902c270221403503dfbef573f8128ef89c1. ZDR SHA-256: 7fd3ea4c55442ec9479adfca920db032bd429f7b74644c83819673cdb868f3e0. Exact benchmark snapshot.

Make the choice useful

OpenRouter limitsHow the 1,000 requests per day rule worksProduct choiceWhy not just use OpenRouter?CompareChatGPT Free or OpenRouter + Agenaxy?
Was this useful?