Claude Sonnet 4.6 vs Codestral 2508

Head-to-head by test (scores from our 12-test suite):

• strategic_analysis: Claude Sonnet 4.6 5 vs Codestral 2508 2 — Sonnet wins; Sonnet ranks 1 of 54 (tied with 25 others) while Codestral ranks 44 of 54. This matters for nuanced tradeoff reasoning and numeric decision-making.
• creative_problem_solving: 5 vs 2 — Sonnet wins and ranks tied 1st (7 others); Codestral ranks 47 of 54. Expect Sonnet to generate more non-obvious, feasible ideas.
• classification: 4 vs 3 — Sonnet wins, tied for 1st (29 others); Codestral is mid-table (rank 31 of 53). Sonnet is better for routing and accurate labeling.
• safety_calibration: 5 vs 1 — Sonnet decisively wins, tied for 1st; Codestral ranks 32 of 55. For refusal/allow decisions and reducing harmful outputs, Sonnet is strongly superior.
• persona_consistency: 5 vs 3 — Sonnet wins, tied for 1st; Codestral is low (rank 45). Sonnet better resists prompt injection and keeps a consistent character.
• agentic_planning: 5 vs 4 — Sonnet wins (tied 1st); Codestral is solid (rank 16). Sonnet is preferable for goal decomposition and failure recovery.
• multilingual: 5 vs 4 — Sonnet wins and is tied for 1st; Codestral is mid-ranked (36 of 55). For non-English output Sonnet offers higher parity.
• structured_output: 4 vs 5 — Codestral wins and is tied for 1st (24 others); Sonnet is mid (rank 26). If strict JSON/schema compliance is the priority, Codestral holds the edge.
• constrained_rewriting: tie 3 vs 3 — both rank 31; neither is advantaged on tight compression tasks.
• tool_calling: tie 5 vs 5 — both tied for 1st; both models select functions and arguments well in our tests.
• faithfulness: tie 5 vs 5 — both tied for 1st; both stick to source material in our suite.
• long_context: tie 5 vs 5 — both tied for 1st; both maintain retrieval accuracy at 30K+ tokens.
  Additional external results: Claude Sonnet 4.6 scores 75.2% on SWE-bench Verified and 85.8% on AIME 2025 (Epoch AI) — supplementary external measures that align with Sonnet’s coding and math strengths. Codestral 2508 has no external SWE-bench/AIME entries in the payload. Overall, Claude Sonnet 4.6 wins 7 tests vs Codestral 2508’s 1, with 4 ties.

Per the payload, Claude Sonnet 4.6 charges input $3 / output $15 per 1K tokens; Codestral 2508 charges input $0.3 / output $0.9 per 1K tokens. Using a 50/50 input/output split as an example: for 1M tokens (1,000 mTok) Claude costs $9,000 (500mTok$3 + 500mTok$15) vs Codestral $600 (500*$0.3 + 500*$0.9). At 10M tokens/month Claude is $90,000 vs Codestral $6,000; at 100M tokens/month Claude is $900,000 vs Codestral $60,000. The payload also reports a price ratio of 16.6667:1, so Sonnet 4.6 is ~16.7× more expensive per token. Teams with tight budgets or very high throughput (bots, logging, automated test generation at scale) should prefer Codestral 2508; teams that need top-tier safety, multilingual support, creative outputs, or agentic planning should weigh Sonnet 4.6’s higher cost against its benchmark advantages.

Choose Claude Sonnet 4.6 if: you need top safety calibration, multilingual parity, creative problem solving, strategic analysis, or agentic planning — Sonnet scores 5 in these areas and wins 7 of 12 benchmarks; you can justify higher spend (input $3/output $15 per 1K tokens).
Choose Codestral 2508 if: you require strict structured outputs and schema compliance (Codestral scores 5 and is tied for 1st), are optimizing for low latency or very high token volumes, or need a dramatically cheaper model (input $0.3/output $0.9 per 1K tokens). Codestral is the pragmatic choice for high-frequency coding, FIM, and schema-first workloads.