Provisional release

Camera-trap results

This is a provisional release. These results were privately scored and independently audited; there is an arXiv paper in preparation. Compare complete scored configurations on the same 790 expert-verified camera-trap events.

Model comparison

Compare saved results; exact values are in the table below.

Headline metric A taxonomic score (Adjusted TIS) that credits the most specific correct rank in each answer: species 5, genus 3, family 2, order 1, class 0.5. Five points are deducted if any stated rank is wrong, so an incorrect identification scores below a withheld answer (0). Averaged across events, from −5 to +5. The proportion of the 790 test events in which the model named the reference species exactly. Answers given only at a higher rank (such as genus), withheld answers and failed responses are all counted as incorrect. The F1 score averaged across species, giving each species equal weight. For a single species, F1 combines precision (the share of the model's identifications of that species that were correct) and recall (the share of that species' events the model identified). Ranges from 0 to 1.
Restricts the results to birds (40 species; 400 events), mammals (39 species; 390 events) or both.
Restricts the results to events recorded in daylight or at night (usually black-and-white infrared images), or to events rated easy, medium or hard to identify during expert curation. Species composition differs between subgroups, so differences between them are descriptive rather than causal.
Source continentRestricts the results by where the camera traps were deployed: Europe (Belgium and Luxembourg; 520 events) or South America (Ecuador; 270 events).
Reasoning levelThe amount of internal reasoning a Gemini model carries out before giving its answer, set from Minimal to High. Higher levels let the model work through more intermediate steps, which increases cost and response time but does not guarantee a more accurate identification. BioCLIP, SpeciesNet and Qwen have no equivalent setting.

The V1 default view shows High reasoning and specialist routes. Every configuration remains in the tables.

Private research preview on a development-overlapping V1 set. Saved candidate statistics; scientific adoption pending. The repaired Qwen primary has a dependent first-pass sensitivity.

Species accuracy is the exact-species match rate on a fixed 0–100% scale.

All results use the same 790-event V1 set.

Chart waiting to load. Exact values are available below.

Cost & performanceSpecies accuracy plotted against the estimated provider (API) charge for processing 100 events. BioCLIP and SpeciesNet run locally and incur no API charge; their hardware and energy costs were not measured.

Source continentRestricts the results by where the camera traps were deployed: Europe (Belgium and Luxembourg; 520 events) or South America (Ecuador; 270 events).

All 790 V1 events · API charges per 100 events. Local models have no API fee; compute costs are unmeasured.

USD display uses the saved 2026-10-16 daily ECB reference rate: £1 = $1.3335. This is a planning-price conversion, not an invoice.

Show estimated costs

Point shape shows reasoning: ● Minimal or other · ▲ Low · ■ Medium · ◆ High. Colour shows the model; labelled values follow.

Solid guides: median cost and performanceDashed links: reasoning levels within a model

    How to read this chart

    Lower API charges are to the right, on a linear scale including $0. BioCLIP and SpeciesNet run locally and have no hosted API fee; their hardware and electricity costs were not measured. Their zero-fee points do not mean zero total cost. Hosted charges use retained token prices or saved provider charges; USD converts GBP estimates at the stated reference rate. Other unknown costs are excluded. The exact tables keep unmeasured costs as unavailable. Shaded regions use the positive priced settings only and are descriptive, not quality ratings. BioCLIP has no comparable Adjusted TIS, so it appears only for species accuracy.

    ECB reference rate via Frankfurter · ECB source page. Daily reference rates are not live trading prices.

    Higher performance · higher costHigher performance · lower costLower performance · higher costLower performance · lower cost

    Confidence & correctnessEach general-purpose model rated its own confidence (low, medium or high) for every answer. The bars show how often answers at each level were correct; in a well-behaved model, accuracy rises with stated confidence.

    Source continentRestricts the results by where the camera traps were deployed: Europe (Belgium and Luxembourg; 520 events) or South America (Ecuador; 270 events).

    When a VLM reports Low, Medium or High confidence, how often is it correct? Confidence is a self-report, not a calibrated probability.

    V1 descriptive results · follows the animal group above; includes all lighting and difficulty levels.

      Exact values and event counts

      Accuracy counts correct species across every event in each confidence group, including withheld answers. Adjusted TIS is the mean saved score, in points per event. Empty groups are unavailable; confidence that was not reported is listed separately. These descriptive groups have no newly estimated intervals.

      All · 790 V1 events per setting
      Model settingConfidenceEventsCorrect speciesSpecies accuracyAdjusted TIS
      Gemini 3.5 Flash-Lite - MinimalLow13410.75%+0.026
      Gemini 3.5 Flash-Lite - MinimalMedium2174420.28%+0.608
      Gemini 3.5 Flash-Lite - MinimalHigh43631371.79%+3.058
      Gemini 3.5 Flash-Lite - MinimalNot reported300.00%+0.000
      Gemini 3.5 Flash-Lite - LowLow15210.66%-0.398
      Gemini 3.5 Flash-Lite - LowMedium2426125.21%+0.696
      Gemini 3.5 Flash-Lite - LowHigh39530677.47%+3.415
      Gemini 3.5 Flash-Lite - LowNot reported100.00%+0.000
      Gemini 3.5 Flash-Lite - MediumLow7000.00%-0.164
      Gemini 3.5 Flash-Lite - MediumMedium32110231.78%+0.731
      Gemini 3.5 Flash-Lite - MediumHigh39729774.81%+3.370
      Gemini 3.5 Flash-Lite - MediumNot reported200.00%+0.000
      Gemini 3.5 Flash-Lite - HighLow4900.00%+0.061
      Gemini 3.5 Flash-Lite - HighMedium2878830.66%+0.362
      Gemini 3.5 Flash-Lite - HighHigh45131870.51%+3.045
      Gemini 3.5 Flash-Lite - HighNot reported300.00%+0.000
      Gemini 3.7 Flash - MediumLow4000.00%+0.175
      Gemini 3.7 Flash - MediumMedium1122421.43%+1.076
      Gemini 3.7 Flash - MediumHigh61141668.09%+3.318
      Gemini 3.7 Flash - MediumNot reported2700.00%+0.000
      Gemini 3.8 Flash - MediumLow3400.00%-0.074
      Gemini 3.8 Flash - MediumMedium1414229.79%+1.691
      Gemini 3.8 Flash - MediumHigh59140668.70%+3.448
      Gemini 3.8 Flash - MediumNot reported2400.00%+0.000
      Gemini 3.8 Flash - LowLow5400.00%-0.111
      Gemini 3.8 Flash - LowMedium1333627.07%+1.688
      Gemini 3.8 Flash - LowHigh57141472.50%+3.585
      Gemini 3.8 Flash - LowNot reported3200.00%+0.000
      Gemini 3.8 Flash - HighLow1900.00%-0.368
      Gemini 3.8 Flash - HighMedium1697343.20%+1.524
      Gemini 3.8 Flash - HighHigh58043374.66%+3.580
      Gemini 3.8 Flash - HighNot reported2200.00%+0.000
      Qwen3-VL-30BLow12100.00%-0.405
      Qwen3-VL-30BMedium2223314.86%-2.239
      Qwen3-VL-30BHigh44728764.21%+1.936
      Qwen3-VL-30BNot reported00Not availableNot available
      Birds · 400 V1 events per setting
      Model settingConfidenceEventsCorrect speciesSpecies accuracyAdjusted TIS
      Gemini 3.5 Flash-Lite - MinimalLow5911.69%-0.034
      Gemini 3.5 Flash-Lite - MinimalMedium912527.47%+0.813
      Gemini 3.5 Flash-Lite - MinimalHigh24817470.16%+2.933
      Gemini 3.5 Flash-Lite - MinimalNot reported200.00%+0.000
      Gemini 3.5 Flash-Lite - LowLow6500.00%-0.338
      Gemini 3.5 Flash-Lite - LowMedium1032726.21%+0.442
      Gemini 3.5 Flash-Lite - LowHigh23117676.19%+3.346
      Gemini 3.5 Flash-Lite - LowNot reported100.00%+0.000
      Gemini 3.5 Flash-Lite - MediumLow2700.00%-0.315
      Gemini 3.5 Flash-Lite - MediumMedium1465034.25%+0.507
      Gemini 3.5 Flash-Lite - MediumHigh22617075.22%+3.332
      Gemini 3.5 Flash-Lite - MediumNot reported100.00%+0.000
      Gemini 3.5 Flash-Lite - HighLow1300.00%-0.115
      Gemini 3.5 Flash-Lite - HighMedium1253628.80%-0.084
      Gemini 3.5 Flash-Lite - HighHigh26118872.03%+3.031
      Gemini 3.5 Flash-Lite - HighNot reported100.00%+0.000
      Gemini 3.7 Flash - MediumLow1100.00%-0.227
      Gemini 3.7 Flash - MediumMedium55814.55%+0.600
      Gemini 3.7 Flash - MediumHigh32224275.16%+3.550
      Gemini 3.7 Flash - MediumNot reported1200.00%+0.000
      Gemini 3.8 Flash - MediumLow1100.00%-0.409
      Gemini 3.8 Flash - MediumMedium521325.00%+1.567
      Gemini 3.8 Flash - MediumHigh32625176.99%+3.643
      Gemini 3.8 Flash - MediumNot reported1100.00%+0.000
      Gemini 3.8 Flash - LowLow2500.00%-0.120
      Gemini 3.8 Flash - LowMedium481327.08%+1.604
      Gemini 3.8 Flash - LowHigh31425179.94%+3.771
      Gemini 3.8 Flash - LowNot reported1300.00%+0.000
      Gemini 3.8 Flash - HighLow600.00%-1.250
      Gemini 3.8 Flash - HighMedium592033.90%+1.042
      Gemini 3.8 Flash - HighHigh32526080.00%+3.758
      Gemini 3.8 Flash - HighNot reported1000.00%+0.000
      Qwen3-VL-30BLow5600.00%-0.482
      Qwen3-VL-30BMedium871314.94%-2.500
      Qwen3-VL-30BHigh25716965.76%+2.107
      Qwen3-VL-30BNot reported00Not availableNot available
      Mammals · 390 V1 events per setting
      Model settingConfidenceEventsCorrect speciesSpecies accuracyAdjusted TIS
      Gemini 3.5 Flash-Lite - MinimalLow7500.00%+0.073
      Gemini 3.5 Flash-Lite - MinimalMedium1261915.08%+0.460
      Gemini 3.5 Flash-Lite - MinimalHigh18813973.94%+3.223
      Gemini 3.5 Flash-Lite - MinimalNot reported100.00%+0.000
      Gemini 3.5 Flash-Lite - LowLow8711.15%-0.443
      Gemini 3.5 Flash-Lite - LowMedium1393424.46%+0.885
      Gemini 3.5 Flash-Lite - LowHigh16413079.27%+3.512
      Gemini 3.5 Flash-Lite - LowNot reported00Not availableNot available
      Gemini 3.5 Flash-Lite - MediumLow4300.00%-0.070
      Gemini 3.5 Flash-Lite - MediumMedium1755229.71%+0.917
      Gemini 3.5 Flash-Lite - MediumHigh17112774.27%+3.421
      Gemini 3.5 Flash-Lite - MediumNot reported100.00%+0.000
      Gemini 3.5 Flash-Lite - HighLow3600.00%+0.125
      Gemini 3.5 Flash-Lite - HighMedium1625232.10%+0.707
      Gemini 3.5 Flash-Lite - HighHigh19013068.42%+3.066
      Gemini 3.5 Flash-Lite - HighNot reported200.00%+0.000
      Gemini 3.7 Flash - MediumLow2900.00%+0.328
      Gemini 3.7 Flash - MediumMedium571628.07%+1.535
      Gemini 3.7 Flash - MediumHigh28917460.21%+3.061
      Gemini 3.7 Flash - MediumNot reported1500.00%+0.000
      Gemini 3.8 Flash - MediumLow2300.00%+0.087
      Gemini 3.8 Flash - MediumMedium892932.58%+1.764
      Gemini 3.8 Flash - MediumHigh26515558.49%+3.209
      Gemini 3.8 Flash - MediumNot reported1300.00%+0.000
      Gemini 3.8 Flash - LowLow2900.00%-0.103
      Gemini 3.8 Flash - LowMedium852327.06%+1.735
      Gemini 3.8 Flash - LowHigh25716363.42%+3.358
      Gemini 3.8 Flash - LowNot reported1900.00%+0.000
      Gemini 3.8 Flash - HighLow1300.00%+0.038
      Gemini 3.8 Flash - HighMedium1105348.18%+1.782
      Gemini 3.8 Flash - HighHigh25517367.84%+3.353
      Gemini 3.8 Flash - HighNot reported1200.00%+0.000
      Qwen3-VL-30BLow6500.00%-0.338
      Qwen3-VL-30BMedium1352014.81%-2.070
      Qwen3-VL-30BHigh19011862.11%+1.705
      Qwen3-VL-30BNot reported00Not availableNot available

      Download confidence results

      Spider chartCompares two configurations across several measures at once, each scaled from 0 to 100, such as accuracy on day and night images. Each axis is read independently; the axes are not combined into an overall score.

      Source continentRestricts the results by where the camera traps were deployed: Europe (Belgium and Luxembourg; 520 events) or South America (Ecuador; 270 events).

      Compare two model settings using saved performance, cost and subgroup results.

      Axes use a 0–100 plotting scale; they are not combined into a score. Raw values and denominators are in the table below.

      Results table

      Source continentRestricts the results by where the camera traps were deployed: Europe (Belgium and Luxembourg; 520 events) or South America (Ecuador; 270 events).

      V1 results

      V1 results: 790 events; key results
      ConfigurationReasoningSpecies accuracyBenchmark scoreEstimated USD / 100 events
      BioCLIP 2.5 - No country contextNot applicable40.76%Not availableNot available: Local compute cost was not measured; zero is not assumed
      SpeciesNet v4.0.3b - No country contextNot applicable27.59%+0.6272 points/eventNot available: Local compute cost was not measured; zero is not assumed
      BioCLIP 2.5 - Country contextNot applicable43.92%Not availableNot available: Local compute cost was not measured; zero is not assumed
      SpeciesNet v4.0.3b - Country contextNot applicable28.35%+1.1753 points/eventNot available: Local compute cost was not measured; zero is not assumed
      Gemini 3.5 Flash-Lite - MinimalMinimal45.32%+1.8595 points/event$0.2330 estimated (n=790)
      Cost basisPriced selected response attempts; Saved cost entries for 779 of 790 events; 11 reconstructed from retained usage; 0 missing; 779 saved telemetry estimates plus corrected selected-response result tokens at retained prices; no provider invoice. Earlier replaced calls are separate.
      Gemini 3.5 Flash-Lite - LowLow46.58%+1.8443 points/event$0.2992 estimated (n=790)
      Cost basisPriced selected response attempts; Saved cost entries for 790 of 790 events; 0 reconstructed from retained usage; 0 missing; Saved selected-response usage at retained GBP price/FX buffers; reservations and earlier failed attempts are separate, not free.
      Gemini 3.5 Flash-Lite - MediumMedium50.51%+1.9759 points/event$0.4332 estimated (n=790)
      Cost basisPriced selected response attempts; Saved cost entries for 790 of 790 events; 0 reconstructed from retained usage; 0 missing; Saved selected-response usage at retained GBP price/FX buffers; reservations and earlier failed attempts are separate, not free.
      Gemini 3.5 Flash-Lite - HighHigh51.39%+1.8741 points/event$0.5421 estimated (n=790)
      Cost basisPriced selected response attempts; Saved cost entries for 790 of 790 events; 0 reconstructed from retained usage; 0 missing; Saved selected-response usage at retained GBP price/FX buffers; reservations and earlier failed attempts are separate, not free.
      Gemini 3.7 Flash - MediumMedium55.70%+2.7278 points/event$0.4296 estimated (n=790)
      Cost basisPriced selected response attempts; Saved cost entries for 790 of 790 events; 0 reconstructed from retained usage; 0 missing; Saved provider-token cost estimate for each selected response; not an invoice.
      Gemini 3.8 Flash - MediumMedium56.71%+2.8785 points/event$1.0087 estimated (n=790)
      Cost basisPriced selected response attempts; Saved cost entries for 0 of 790 events; 790 reconstructed from retained usage; 0 missing; Selected-response tokens repriced with the run's retained list prices and GBP buffers; no provider invoice. Earlier failed/uncertain attempts are separate.
      Gemini 3.8 Flash - LowLow56.96%+2.8677 points/event$0.8391 estimated (n=790)
      Cost basisPriced selected response attempts; Saved cost entries for 789 of 790 events; 1 reconstructed from retained usage; 0 missing; 789 saved reconciled selected-response costs plus one prior canary cost reconstructed from its hash-bound result tokens and retained price/FX buffer; earlier failed or uncertain attempts remain separate. Not an invoice.
      Gemini 3.8 Flash - HighHigh64.05%+2.9456 points/event$1.7527 estimated (n=790)
      Cost basisPriced selected response attempts; Saved cost entries for 789 of 790 events; 1 reconstructed from retained usage; 0 missing; 789 saved reconciled selected-response costs plus one prior canary cost reconstructed from its hash-bound result tokens and retained price/FX buffer; earlier failed or uncertain attempts remain separate. Not an invoice.
      Qwen3-VL-30BNot asserted40.51%+0.4044 points/event$0.2632 estimated (n=790)
      Cost basisPriced selected response attempts; Saved cost entries for 790 of 790 events; 0 reconstructed from retained usage; 0 missing; Selected evaluated response attempt only: 786 first-pass outputs, four successful repair outputs. Four superseded first-pass length charges are excluded from this comparison and retained in the audit below.

      Download full configuration fields as CSV · Download saved category values as JSON

      1 dependent sensitivity settings are retained in the full CSV.

      V1 results · Birds

      400 saved birds events. Pointwise intervals, where recorded, do not describe new species.

      V1 · Birds: 400 events; saved subgroup results
      ConfigurationReasoningSpecies accuracyBenchmark scoreEstimated USD / 100 events
      BioCLIP 2.5 - No country contextNot applicable38.75%Not availableNot available: Local compute cost was not measured; zero is not assumed
      Cost basisSaved cost entries for 0 of 400 events; 0 reconstructed from retained usage; 400 missing
      SpeciesNet v4.0.3b - No country contextNot applicable12.00%+0.1138 points/eventNot available: Local compute cost was not measured; zero is not assumed
      Cost basisSaved cost entries for 0 of 400 events; 0 reconstructed from retained usage; 400 missing
      BioCLIP 2.5 - Country contextNot applicable41.75%Not availableNot available: Local compute cost was not measured; zero is not assumed
      Cost basisSaved cost entries for 0 of 400 events; 0 reconstructed from retained usage; 400 missing
      SpeciesNet v4.0.3b - Country contextNot applicable12.25%+0.3150 points/eventNot available: Local compute cost was not measured; zero is not assumed
      Cost basisSaved cost entries for 0 of 400 events; 0 reconstructed from retained usage; 400 missing
      Gemini 3.5 Flash-Lite - MinimalMinimal50.00%+1.9988 points/event$0.2363 estimated (n=400)
      Cost basisPriced selected response attempts; Saved cost entries for 390 of 400 events; 10 reconstructed from retained usage; 0 missing; 779 saved telemetry estimates plus corrected selected-response result tokens at retained prices; no provider invoice. Earlier replaced calls are separate.
      Gemini 3.5 Flash-Lite - LowLow50.75%+1.9913 points/event$0.3024 estimated (n=400)
      Cost basisPriced selected response attempts; Saved cost entries for 400 of 400 events; 0 reconstructed from retained usage; 0 missing; Saved selected-response usage at retained GBP price/FX buffers; reservations and earlier failed attempts are separate, not free.
      Gemini 3.5 Flash-Lite - MediumMedium55.00%+2.0463 points/event$0.4325 estimated (n=400)
      Cost basisPriced selected response attempts; Saved cost entries for 400 of 400 events; 0 reconstructed from retained usage; 0 missing; Saved selected-response usage at retained GBP price/FX buffers; reservations and earlier failed attempts are separate, not free.
      Gemini 3.5 Flash-Lite - HighHigh56.00%+1.9475 points/event$0.5369 estimated (n=400)
      Cost basisPriced selected response attempts; Saved cost entries for 400 of 400 events; 0 reconstructed from retained usage; 0 missing; Saved selected-response usage at retained GBP price/FX buffers; reservations and earlier failed attempts are separate, not free.
      Gemini 3.7 Flash - MediumMedium62.50%+2.9337 points/event$0.4353 estimated (n=400)
      Cost basisPriced selected response attempts; Saved cost entries for 400 of 400 events; 0 reconstructed from retained usage; 0 missing; Saved provider-token cost estimate for each selected response; not an invoice.
      Gemini 3.8 Flash - MediumMedium66.00%+3.1612 points/event$1.0158 estimated (n=400)
      Cost basisPriced selected response attempts; Saved cost entries for 0 of 400 events; 400 reconstructed from retained usage; 0 missing; Selected-response tokens repriced with the run's retained list prices and GBP buffers; no provider invoice. Earlier failed/uncertain attempts are separate.
      Gemini 3.8 Flash - LowLow66.00%+3.1450 points/event$0.8573 estimated (n=400)
      Cost basisPriced selected response attempts; Saved cost entries for 399 of 400 events; 1 reconstructed from retained usage; 0 missing; 789 saved reconciled selected-response costs plus one prior canary cost reconstructed from its hash-bound result tokens and retained price/FX buffer; earlier failed or uncertain attempts remain separate. Not an invoice.
      Gemini 3.8 Flash - HighHigh70.00%+3.1888 points/event$1.6425 estimated (n=400)
      Cost basisPriced selected response attempts; Saved cost entries for 399 of 400 events; 1 reconstructed from retained usage; 0 missing; 789 saved reconciled selected-response costs plus one prior canary cost reconstructed from its hash-bound result tokens and retained price/FX buffer; earlier failed or uncertain attempts remain separate. Not an invoice.
      Qwen3-VL-30BNot asserted45.50%+0.7425 points/event$0.2914 estimated (n=400)
      Cost basisPriced selected response attempts; Saved cost entries for 400 of 400 events; 0 reconstructed from retained usage; 0 missing; Selected evaluated response attempt only: 786 first-pass outputs, four successful repair outputs. Four superseded first-pass length charges are excluded from this comparison and retained in the audit below.

      Download full configuration fields as CSV · Download saved category values as JSON

      V1 results · Mammals

      390 saved mammals events. Pointwise intervals, where recorded, do not describe new species.

      V1 · Mammals: 390 events; saved subgroup results
      ConfigurationReasoningSpecies accuracyBenchmark scoreEstimated USD / 100 events
      BioCLIP 2.5 - No country contextNot applicable42.82%Not availableNot available: Local compute cost was not measured; zero is not assumed
      Cost basisSaved cost entries for 0 of 390 events; 0 reconstructed from retained usage; 390 missing
      SpeciesNet v4.0.3b - No country contextNot applicable43.59%+1.1538 points/eventNot available: Local compute cost was not measured; zero is not assumed
      Cost basisSaved cost entries for 0 of 390 events; 0 reconstructed from retained usage; 390 missing
      BioCLIP 2.5 - Country contextNot applicable46.15%Not availableNot available: Local compute cost was not measured; zero is not assumed
      Cost basisSaved cost entries for 0 of 390 events; 0 reconstructed from retained usage; 390 missing
      SpeciesNet v4.0.3b - Country contextNot applicable44.87%+2.0577 points/eventNot available: Local compute cost was not measured; zero is not assumed
      Cost basisSaved cost entries for 0 of 390 events; 0 reconstructed from retained usage; 390 missing
      Gemini 3.5 Flash-Lite - MinimalMinimal40.51%+1.7167 points/event$0.2296 estimated (n=390)
      Cost basisPriced selected response attempts; Saved cost entries for 389 of 390 events; 1 reconstructed from retained usage; 0 missing; 779 saved telemetry estimates plus corrected selected-response result tokens at retained prices; no provider invoice. Earlier replaced calls are separate.
      Gemini 3.5 Flash-Lite - LowLow42.31%+1.6936 points/event$0.2958 estimated (n=390)
      Cost basisPriced selected response attempts; Saved cost entries for 390 of 390 events; 0 reconstructed from retained usage; 0 missing; Saved selected-response usage at retained GBP price/FX buffers; reservations and earlier failed attempts are separate, not free.
      Gemini 3.5 Flash-Lite - MediumMedium45.90%+1.9038 points/event$0.4339 estimated (n=390)
      Cost basisPriced selected response attempts; Saved cost entries for 390 of 390 events; 0 reconstructed from retained usage; 0 missing; Saved selected-response usage at retained GBP price/FX buffers; reservations and earlier failed attempts are separate, not free.
      Gemini 3.5 Flash-Lite - HighHigh46.67%+1.7987 points/event$0.5475 estimated (n=390)
      Cost basisPriced selected response attempts; Saved cost entries for 390 of 390 events; 0 reconstructed from retained usage; 0 missing; Saved selected-response usage at retained GBP price/FX buffers; reservations and earlier failed attempts are separate, not free.
      Gemini 3.7 Flash - MediumMedium48.72%+2.5167 points/event$0.4237 estimated (n=390)
      Cost basisPriced selected response attempts; Saved cost entries for 390 of 390 events; 0 reconstructed from retained usage; 0 missing; Saved provider-token cost estimate for each selected response; not an invoice.
      Gemini 3.8 Flash - MediumMedium47.18%+2.5885 points/event$1.0014 estimated (n=390)
      Cost basisPriced selected response attempts; Saved cost entries for 0 of 390 events; 390 reconstructed from retained usage; 0 missing; Selected-response tokens repriced with the run's retained list prices and GBP buffers; no provider invoice. Earlier failed/uncertain attempts are separate.
      Gemini 3.8 Flash - LowLow47.69%+2.5833 points/event$0.8205 estimated (n=390)
      Cost basisPriced selected response attempts; Saved cost entries for 390 of 390 events; 0 reconstructed from retained usage; 0 missing; 789 saved reconciled selected-response costs plus one prior canary cost reconstructed from its hash-bound result tokens and retained price/FX buffer; earlier failed or uncertain attempts remain separate. Not an invoice.
      Gemini 3.8 Flash - HighHigh57.95%+2.6962 points/event$1.8657 estimated (n=390)
      Cost basisPriced selected response attempts; Saved cost entries for 390 of 390 events; 0 reconstructed from retained usage; 0 missing; 789 saved reconciled selected-response costs plus one prior canary cost reconstructed from its hash-bound result tokens and retained price/FX buffer; earlier failed or uncertain attempts remain separate. Not an invoice.
      Qwen3-VL-30BNot asserted35.38%+0.0577 points/event$0.2344 estimated (n=390)
      Cost basisPriced selected response attempts; Saved cost entries for 390 of 390 events; 0 reconstructed from retained usage; 0 missing; Selected evaluated response attempt only: 786 first-pass outputs, four successful repair outputs. Four superseded first-pass length charges are excluded from this comparison and retained in the audit below.

      Download full configuration fields as CSV · Download saved category values as JSON

      Evidence and uncertainty

      Download complete configurations, subgroup data, intervals and paired analyses. Pointwise intervals describe sampled events; they do not establish field performance.

      95% species-accuracy intervals as a chart
      Source continentRestricts the results by where the camera traps were deployed: Europe (Belgium and Luxembourg; 520 events) or South America (Ecuador; 270 events).

      Dots show exact-species accuracy; bars show saved pointwise 95% intervals for the selected V1 animal group. The resampling method varies by group; exact methods and endpoints are in the downloadable evidence.

      Open this section to draw the interval chart. Exact endpoints are in the downloads.

      Day and night exact subgroup values
      Saved V1 species accuracy by lighting; non-causal subgroup descriptions
      ConfigurationDayNight / infrared
      BioCLIP 2.5 - No country context43.4% (95% interval 38.6–48.2%; n=429)37.7% (95% interval 32.4–43.0%; n=361)
      SpeciesNet v4.0.3b - No country context20.3% (95% interval 16.2–24.5%; n=429)36.3% (95% interval 31.0–41.7%; n=361)
      Gemini 3.5 Flash-Lite - Minimal53.8% (95% interval 48.9–58.7%; n=429)35.2% (95% interval 30.1–40.3%; n=361)
      Gemini 3.5 Flash-Lite - Low55.0% (95% interval 50.0–59.8%; n=429)36.6% (95% interval 31.3–41.8%; n=361)
      Gemini 3.5 Flash-Lite - Medium60.1% (95% interval 55.1–65.0%; n=429)39.1% (95% interval 33.9–44.4%; n=361)
      Gemini 3.5 Flash-Lite - High60.6% (95% interval 55.8–65.4%; n=429)40.4% (95% interval 35.2–45.7%; n=361)
      Gemini 3.7 Flash - Medium65.0% (95% interval 60.2–69.8%; n=429)44.6% (95% interval 39.3–49.9%; n=361)
      Gemini 3.8 Flash - Medium67.8% (95% interval 63.1–72.5%; n=429)43.5% (95% interval 38.1–48.9%; n=361)
      Gemini 3.8 Flash - Low68.8% (95% interval 64.0–73.3%; n=429)42.9% (95% interval 37.7–48.3%; n=361)
      Gemini 3.8 Flash - High73.2% (95% interval 68.8–77.4%; n=429)53.2% (95% interval 47.8–58.5%; n=361)
      Qwen3-VL-30B49.0% (95% interval 44.0–54.0%; n=429)30.5% (95% interval 25.5–35.6%; n=361)
      Spider chart values and intervals
      Selected V1 profile configurations; six disjoint lighting and difficulty groups
      Configurationeasy / dayeasy / night_irmedium / daymedium / night_irhard / dayhard / night_ir
      Gemini 3.5 Flash-Lite - High75.5% (95% interval 69.6–81.2%; n=245)56.9% (95% interval 48.4–65.5%; n=130)57.9% (95% interval 48.5–67.0%; n=107)39.1% (95% interval 30.8–47.6%; n=133)15.8% (95% interval 7.8–24.4%; n=76)20.4% (95% interval 12.6–28.9%; n=98)
      Gemini 3.8 Flash - High86.5% (95% interval 81.9–90.8%; n=245)71.5% (95% interval 63.6–79.5%; n=130)69.2% (95% interval 60.3–77.6%; n=107)54.1% (95% interval 45.4–62.8%; n=133)35.5% (95% interval 25.0–46.4%; n=76)27.6% (95% interval 18.8–37.0%; n=98)

      One event has unset difficulty; complete subgroup values are in the downloadable analysis.

      Download detailed analysis tables

      CSV files contain all rows. JSON files retain units, definitions and analysis notes.

      • Headline uncertainty · 1140 aggregate rows CSV · JSON
      • Paired changes · 213 aggregate rows CSV · JSON
      • Paired tests and multiplicity · 18 aggregate rows CSV · JSON
      • Lighting and difficulty intervals · 4320 aggregate rows CSV · JSON
      • Subgroup composition · 36 aggregate rows CSV · JSON
      • Verbal confidence and answer reliability · 58 aggregate rows CSV · JSON
      • Coverage and selective accuracy · 60 aggregate rows CSV · JSON
      • Specialist native score reliability · 20 aggregate rows CSV · JSON
      • Per-species descriptive results · 1106 aggregate rows CSV · JSON
      • Ecological gaps · 78 aggregate rows CSV · JSON
      • Exploratory time interactions · 9 aggregate rows CSV · JSON
      • Paired prediction transitions · 192 aggregate rows CSV · JSON
      • Taxonomic rank confidence · 250 aggregate rows CSV · JSON
      • Specialist output states · 4 aggregate rows CSV · JSON
      • Specialist paired tests · 3 aggregate rows CSV · JSON

      How to read these results

      Species accuracy counts exact species identifications across all events, including withheld answers and technical nonresponses. Benchmark score is eligible Adjusted TIS on a signed −5 to +5 points-per-event scale. BioCLIP has no comparable Adjusted TIS. Missing values and missing costs are not zero.

      Compare cost and performance · Methods summary (full version coming soon) · Inspect five saved examples

      Source code, the full image dataset and a citable release are coming soon.