SCALE April 2026 Large‑Model SQL Capability Ranking Unveiled

Release Summary

Four new models are added to the SCALE benchmark: DeepSeek‑V4‑Pro , DeepSeek‑V4‑Flash , GPT‑5.5 and Claude Opus 4.7 . The evaluation continues to focus on three core dimensions—SQL Understanding, SQL Optimization, and Dialect Conversion—using a unified dataset for reproducibility.

GPT‑5.5 leads the SQL Understanding dimension with the highest execution‑accuracy score, making it suitable for complex semantic analysis.

Claude Opus 4.7 shows a balanced overall capability, ranking in the top‑10 for both SQL Optimization (score 88.7) and Dialect Conversion (Chinese database conversion score 100.0).

DeepSeek‑V4‑Pro and DeepSeek‑V4‑Flash excel at Chinese database conversion but lag in complex SQL understanding, deep optimization, and large‑SQL conversion.

Evaluation Methodology

SQL Understanding : measures depth of logical, intent and execution‑plan analysis. Metrics include execution accuracy, plan detection, and syntax‑error detection.

SQL Optimization : evaluates the ability to rewrite inefficient queries while preserving logical equivalence and syntax correctness. Metrics cover logical equivalence, optimization depth, syntax‑error detection, and index‑recommendation quality.

Dialect Conversion : assesses accurate migration between database dialects and complex procedural‑logic reconstruction. Metrics include large‑SQL conversion, Chinese database conversion, logical equivalence, and syntax‑error detection.

Model Evaluations

GPT‑5.5 (OpenAI)

Capability Positioning : strongest among the new models in SQL Understanding, with solid execution accuracy and syntax‑error detection; optimization and dialect conversion remain usable.

Figure 1: GPT‑5.5 capability dimension scores

Core Dimension Analysis

SQL Understanding : execution accuracy is strong, indicating good grasp of SQL semantics; plan detection is lower, showing room for improvement on complex execution paths.

SQL Optimization : logical equivalence is robust, but optimization depth scores 49.6 . Two typical failure modes observed:

Over‑simplification of multi‑level subqueries. Example: in a projection‑pushdown case the model removed the unused gender column but also dropped the IN filter, producing SELECT student_name FROM students; instead of preserving the filter.

Unstable ordering of composite index fields. For

SELECT * FROM products WHERE category_id = 5 AND is_imported = 1 AND stock < 10

the model suggested index order is_imported, category_id, stock rather than the optimal category_id, is_imported, stock.

Dialect Conversion : Chinese database conversion scores 97.4 , showing balanced syntax‑error detection and logical equivalence. Large‑SQL conversion scores only 45.2 . Example: converting an Oracle SP_BULK_UPDATE_INVENTORY procedure to PostgreSQL resulted in inconsistent handling of record types, batch structures, and default values.

Claude Opus 4.7 (Anthropic)

Capability Positioning : most balanced among the new models; both SQL Optimization and Dialect Conversion rank in the top‑10.

Figure 2: Claude Opus 4.7 capability dimension scores

Core Dimension Analysis

SQL Understanding : basic syntax and semantic checks are reliable; execution‑plan detection remains weaker.

SQL Optimization : syntax‑error detection scores 88.7 , indicating low‑risk, readable suggestions. Optimization depth scores 43.3 . Typical failure: simplifying a nested IN (SELECT student_id FROM (...)) query to SELECT student_name FROM students;, dropping the filter condition.

Dialect Conversion : Chinese database conversion reaches a perfect 100.0 . Other conversion metrics are balanced, but inconsistencies appear, e.g., mismatched END identifiers and differing TO_CHAR syntax when converting Oracle to OceanBase, or parameter‑binding errors when converting SQLServer to GaussDB.

DeepSeek‑V4‑Pro (DeepSeek)

Capability Positioning : thought‑oriented dialogue model (1.6 T total / 49 B active parameters). Performs better on SQL Optimization and Dialect Conversion than on SQL Understanding.

Figure 3: DeepSeek‑V4‑Pro capability dimension scores

Core Dimension Analysis

SQL Understanding : syntax‑error detection is decent, but execution‑plan detection scores 46.4 , indicating difficulty with complex result inference.

SQL Optimization : syntax‑error detection is solid; optimization depth scores 42.1 . Typical issue: multi‑layer subquery simplification that removes necessary filter conditions.

Dialect Conversion : Chinese database conversion is strong, but large‑SQL conversion scores 35.5 . Example: converting an Oracle bulk_delete_by_ids procedure to PostgreSQL loses the explicit COMMIT transaction semantics.

DeepSeek‑V4‑Flash (DeepSeek)

Capability Positioning : lightweight thought‑oriented model (284 B total / 13 B active parameters). Outperforms on SQL Understanding but lags behind in Optimization and Dialect Conversion.

Figure 4: DeepSeek‑V4‑Flash capability dimension scores

Core Dimension Analysis

SQL Understanding : reliable for basic SQL identification and result judgment; execution‑plan detection still needs improvement. Example: for an INSERT statement the model incorrectly filled Extra fields with SELECT‑style values.

SQL Optimization : syntax‑error detection and index suggestions are acceptable, but optimization depth scores 30.4 . Example: simplifying WHERE CONCAT("id_", student_id) >= "id_1000" to student_id >= 1000 drops the string‑concatenation semantics.

Dialect Conversion : Chinese database conversion scores 97.4 , but large‑SQL conversion scores 25.8 . Notable failures include incomplete query generation when converting Oracle procedures to PostgreSQL and malformed SELECT statements when converting SQLServer to GaussDB.

Overall Rankings

Top performers across the 33 evaluated models:

Gemini 3 Pro leads SQL Understanding with a score of 86.0 .

SQLFlash leads SQL Optimization with a score of 72.1 .

SQLShift leads Dialect Conversion with a score of 83.4 .

Conclusions

The four new models form a clear hierarchy:

GPT‑5.5 excels at deep SQL Understanding and is the primary choice for complex semantic analysis and execution‑result validation.

Claude Opus 4.7 offers balanced stability, making it suitable for reliable SQL Optimization suggestions and Chinese database migration (conversion score 100.0).

DeepSeek‑V4‑Pro is best used as an assistant for optimization drafts and migration tasks where logical equivalence is clear.

DeepSeek‑V4‑Flash fits low‑latency, cost‑sensitive batch processing, serving as an initial filter before higher‑capability models or manual review handle deep optimization or large‑SQL migration.

SCALE will continue to monitor large‑model advances, refine its evaluation framework, and provide objective capability references for users.