Top models for all 26 LLL.bz tasks.

For broad conversational quality across everyday prompts.

Note: The primary ranking is a blind-preference arena; confidence intervals overlap.

For Russian-language conversation and writing.

Note: Fresh primary data exists, but three independent Russian-specific sources reproducing the full top three were not found.

For difficult analytical and professional knowledge tasks.

Note: The top three intervals overlap, so the exact order is not statistically decisive.

For hard calculations, proofs and multi-step maths.

Note: All three reported confidence intervals overlap materially.

For prompts with detailed constraints and required formats.

Note: First and second are practically tied within the reported intervals.

For persuasive, creative and polished marketing text.

Note: The source measures blind writing preference, not measured conversion performance.

For web search, evidence gathering and sourced answers.

Note: The dedicated search board updates slower than the ten-day rebuild cycle.

For understanding long documents, PDFs and mixed document content.

Note: The document board updates slower than the ten-day cycle; duplicate Claude Opus 4.6 configurations are collapsed into one entry.

For translating text accurately between languages.

Note: Measured by model family rather than exact version: the evaluation reports Gemini, Claude and GPT as families, so read this as a family-level ranking.

For repository work, terminal coding and software engineering.

Note: Terminal-Bench measures agentic terminal coding; SWE-bench and Code Arena measure different coding workloads.

For building and editing interactive web interfaces.

Note: The ranking measures blind preference over generated web applications.

For autonomous multi-step work with tools and external environments.

Metric: Higher net improvement is better.

Note: Claude Opus 5 Max is omitted because it is another configuration of the same base model.

For tables, CSV analysis, SQL generation and data-driven conclusions.

Note: The fresh source measures data-oriented web applications, not normalized text-to-SQL execution accuracy. BIRD and Spider2 did not yield a clean current raw-model top three that could be reconciled with this board.

For reliable function calls, JSON arguments and structured tool use.

Note: Scores come from the Berkeley function-calling benchmark. Tool schemas differ between providers, so narrow gaps here mean little in practice.

For reasoning about photos, diagrams and visual content.

Note: Second and third are practically tied within the reported intervals.

For reading text embedded in images and scanned content.

Note: OCR Arena is preference-based rather than a character-error-rate benchmark.

For generating images from natural-language prompts.

Note: Second and third differ by one displayed Elo point.

For editing existing images from text instructions.

Note: The two major public image-edit arenas currently disagree materially.

For generating complete videos from text prompts.

Note: This ranking uses the with-audio table; the no-audio table has a different order.

For animating a still image into a generated video.

Note: The selected ranking uses the no-audio table to isolate visual animation quality.

For instruction-driven edits and transformations of existing video.

Note: The two public video-edit arenas currently disagree on first place.

For transcribing spoken audio into text.

Metric: Lower WER is better.

For generating natural spoken audio from text.

Note: Second and third have the same displayed Elo score.

For real-time spoken conversation with an interactive AI agent.

Note: Duplicate effort settings of Gemini 3.1 Flash Live were collapsed to the best configuration.

For generating complete instrumental or vocal music.

Note: Task-level models mirror the instrumental variant. Only one specialized evaluator covers this area, so treat the order as indicative.

For reproducing a speaker's voice from a short reference sample.

Metric: Lower rank is better.

Note: The order comes from a single controlled-voice comparison, so only the ranking is stored, not exact scores.