DeepSeek V3.1 Terminus vs GPT-5

Summary of our 12-test suite (internal scores are ours; external math/coding tests are from Epoch AI): GPT-5 wins 7 categories, the rest are ties; DeepSeek wins none. Detailed comparison:

• Tool calling: GPT-5 5 vs DeepSeek 3 — GPT-5 tied for 1st of 54 models on tool calling; DeepSeek ranks 47 of 54. That means GPT-5 selects functions, sequences calls and fills arguments more reliably in our tests.
• Faithfulness: GPT-5 5 vs DeepSeek 3 — GPT-5 ties for 1st of 55; DeepSeek ranks 52 of 55. For source-faithful outputs (factual adherence), GPT-5 is substantially stronger in our testing.
• Classification: GPT-5 4 vs DeepSeek 3 — GPT-5 tied for 1st of 53; DeepSeek rank 31 of 53. GPT-5 is better at routing/categorization tasks in our suite.
• Agentic planning: GPT-5 5 vs DeepSeek 4 — GPT-5 tied for 1st of 54; DeepSeek rank 16. GPT-5 decomposes goals and recovers from failure more robustly in our scenarios.
• Constrained rewriting: GPT-5 4 vs DeepSeek 3 — GPT-5 rank 6 of 53; DeepSeek rank 31. GPT-5 performed better at tight-character-limit rewrites.
• Persona consistency: GPT-5 5 vs DeepSeek 4 — GPT-5 tied for 1st; DeepSeek rank 38. GPT-5 better resists persona injection in our tests.
• Safety calibration: GPT-5 2 vs DeepSeek 1 — GPT-5 ranks 12 of 55 vs DeepSeek 32; GPT-5 more often calibrated refusals vs allowed requests in our tests. Ties (both models scored the same): structured_output (5/5, tied for 1st), strategic_analysis (5/5, tied for 1st), creative_problem_solving (4/4), long_context (5/5, both tied for 1st), multilingual (5/5, tied for 1st). Those ties show both models handle long contexts, JSON/schema outputs, cross-lingual quality, and higher-level strategic reasoning well in our suite. External benchmarks (Epoch AI): GPT-5 scores 73.6% on SWE-bench Verified, 98.1% on MATH Level 5, and 91.4% on AIME 2025. These external results supplement our internal wins on tool calling/classification and indicate strong coding and advanced-math performance for GPT-5. DeepSeek has no external scores in the payload to compare.

Per-mTok prices from the payload: DeepSeek V3.1 Terminus — input $0.21, output $0.79; GPT-5 — input $1.25, output $10.00. Per-million-token math (1 mTok = 1,000 tokens):

• DeepSeek input-only: $210 / 1M tokens; output-only: $790 / 1M; balanced (50/50): $500 / 1M.
• GPT-5 input-only: $1,250 / 1M; output-only: $10,000 / 1M; balanced (50/50): $5,625 / 1M. At scale this gap multiplies: for balanced usage, DeepSeek = $500 / 1M, $5,000 / 10M, $50,000 / 100M; GPT-5 = $5,625 / 1M, $56,250 / 10M, $562,500 / 100M. Output-heavy workloads widen the gap further (GPT-5 output $10,000 / 1M vs DeepSeek $790 / 1M). Teams with high token volume (10M–100M+/mo), tight margins, or heavy output usage should care about the cost gap; occasional low-volume users may prefer GPT-5’s benchmark advantages despite higher spend.

Choose DeepSeek V3.1 Terminus if: you need long-context handling and structured-output at much lower cost (input $0.21 / mTok, output $0.79 / mTok), or you expect sustained high-volume usage (10M–100M tokens/month) and must control expenses. Choose GPT-5 if: you prioritize tool calling, faithfulness, classification, constrained rewriting, and agentic planning — its internal wins plus Epoch AI external scores (SWE-bench 73.6%, MATH Level 5 98.1%, AIME 2025 91.4%) make it the stronger choice for complex, correctness-sensitive workflows and math/coding tasks despite much higher per-mTok pricing.