Claude Opus 4.8 vs GPT-5.5 vs Gemini 3.5 Flash: Who Leads the Frontier in May 2026?

* SWE-Bench Pro scores for Gemini 3.5 Flash, Grok 4.3, and DeepSeek V4 Pro are from a single third-party test — Google's own model card shows different numbers (see notes below).

Pricing & Performance Metrics

Key findings

Which model leads what?

• Claude Opus 4.8 leads on agentic coding (SWE-Bench Pro), knowledge work (GDPval-AA) at 1890, computer use (OSWorld) at 83.4%, multidisciplinary reasoning (Humanity's Last Exam), and financial analysis . It ranks #2 overall on provisional leaderboards with a 93/100 composite score .
• GPT-5.5 leads on terminal coding (Terminal-Bench 2.0/2.1) at 78.2–82.7%, ARC-AGI-2 at 85.0% (abstract visual reasoning), GPQA Diamond at 96.0%, FrontierMath, and the AA Intelligence Index at 60 .
• Gemini 3.5 Flash leads on tool-use orchestration (MCP Atlas) at 83.6% and is ~4x faster in output speed (~289 tokens/sec) than any other model here — at the lowest cost .
• DeepSeek V4 Pro leads on coding competitions with Codeforces ELO 3206 and LiveCodeBench at 93.5%, and matches Claude on SWE-bench Verified at 80.6% — at a fraction of the cost .
• Grok 4.3 is competitive in the mid-tier: AA Index 53, GPQA 90.1%, fast speed (159–207 tps), and very low pricing — but trails the leaders on most agentic benchmarks .
• Claude Opus 4.7 is the direct predecessor — still strong on SWE-bench Verified (87.6%) but now surpassed by its own successor on every key metric .

Important caveats

• Benchmark sources vary. The data comes from multiple vendors (Anthropic, OpenAI, Google DeepMind, independent testers) using different harnesses. Scores for the same benchmark can differ across sources (e.g., GPT-5.5 Terminal-Bench shows 78.2% in Google's model card vs. 82.7% in OpenAI's own report).
• SWE-Bench Pro vs. SWE-Bench Verified are different tests. Pro is harder (multi-file architectural changes); Verified is easier (single-issue fixes). Claude leads Pro; the field is tighter on Verified.
• DeepSeek, Grok, and Gemini 3.5 Flash are significantly cheaper per token than Claude Opus and GPT-5.5, making them better value for high-volume or speed-sensitive workloads.
• Gemini 3.5 Flash is a "Flash" (efficiency-optimized) tier model, not a direct frontier counterpart — yet it competes well on many agentic benchmarks while costing and running much faster.

What remains unclear

• No single cross-vendor benchmark suite tests all six models on the exact same harness with equal methodology. The composite picture is built from overlapping but not identical test sets.
• Grok 4.3 and DeepSeek V4 Pro have sparser public data on agentic and long-context benchmarks compared to the Big Three vendors.
• Claude Opus 4.8 just launched (May 28, 2026) — independent replication data is limited; most scores are vendor-reported.

Most trusted sources

1. Official vendor pages: Anthropic (claude/opus), OpenAI (introducing-gpt-5-5), Google DeepMind (gemini-3-5-flash model card) — primary source data.
2. NIST CAISI evaluation of DeepSeek V4 Pro — independent government evaluation .
3. Duke University analysis of Gemini 3.5 Flash — academic institution coverage .
4. Third-party aggregators: dev.to head-to-head test , BenchLM.ai , Artificial Analysis — useful for cross-model comparison but less authoritative than official data.