AI Benchmarks & Leaderboard — 2026-05-05

New Model Releases & Updates

GPT-5.5 by OpenAI

• Type: Closed-source; multiple effort tiers (xhigh, high)
• Key benchmarks: Intelligence Index score of 60 (xhigh) and 59 (high) — top two spots on Artificial Analysis leaderboard
• vs. Previous best: Outperforms GPT-5.4 (xhigh, score 57) and all Claude/Gemini variants in overall intelligence ranking
• What's notable: Two distinct compute tiers available; xhigh variant leads all models tracked by Artificial Analysis as of early May 2026

Claude Opus 4.7 by Anthropic

• Type: Closed-source; Adaptive Reasoning mode with Max Effort setting
• Key benchmarks: Intelligence Index score of 57 (Adaptive Reasoning, Max Effort) — tied for 3rd place with Gemini 3.1 Pro Preview
• vs. Previous best: Trails GPT-5.5 xhigh by 3 points but leads all non-OpenAI closed models at max effort
• What's notable: Anthropic stacked $50B in additional capital this week, signaling major continued investment; adaptive reasoning mode distinguishes it from prior Claude versions

Gemini 3.1 Pro Preview by Google

• Type: Closed-source
• Key benchmarks: Intelligence Index score of 57 — tied 3rd with Claude Opus 4.7; Flash-Lite Preview variant ranks among the fastest models at 371.0 tokens/sec
• vs. Previous best: Matches Claude Opus 4.7 on intelligence; Flash-Lite trades quality for speed
• What's notable: Google published a full April 2026 AI recap, confirming multiple Gemini model updates shipped across the month

DeepSeek V4 by DeepSeek (China)

• Type: Open-source/open-weight
• Key benchmarks: Cited as having "almost closed the gap" with frontier closed models on reasoning benchmarks; more efficient and performant than DeepSeek V3.2
• vs. Previous best: Surpasses DeepSeek V3.2 via architectural improvements; Forbes reported it is reshaping the open-source AI race alongside Qwen
• What's notable: DeepSeek again cut AI prices with V4, reigniting the AI price war roughly one year after the original DeepSeek disruption; fully open-weight release

Qwen 3.5 by Alibaba

• Type: Open-source; multiple parameter sizes
• Key benchmarks: Qwen3.5 0.8B (Reasoning) is the most affordable model tracked, at $0.02 per 1M tokens blended; 397B variant achieves 5.5+ tokens/sec locally on Apple M-series hardware
• vs. Previous best: Extends Qwen lineage; 0.8B variant undercuts all closed models on price while remaining competitive
• What's notable: Available across all parameter sizes from sub-1B to 397B; codersera.com ranked it among five frontier-class open-weight models shipping in 30 days

Leaderboard Snapshot

Frontier Models (Closed-Source)

Open-Source Leaders

Benchmark Deep Dive

State of AI — Cyber-Offense Benchmarks Double Every Four Months

The Air Street "State of AI: May 2026" report, published May 4, 2026, surfaced one of the most striking benchmark findings of the week: the UK's AI Safety Institute (AISI) reports that frontier models' cyber-offense capability is doubling approximately every four months. This finding comes from ongoing red-teaming evaluations that track how effectively frontier models can assist in or automate offensive cyber operations — a rapidly advancing and deeply concerning capability curve.

What makes this particularly significant for practitioners is the compounding nature of the trend. A doubling every four months means cyber-offense capability could be roughly 8× more potent by the end of 2026 compared to early 2025 baselines — a trajectory that has caught the attention of safety researchers and policymakers alike. The AISI measurement methodology tracks models across standardized offensive security tasks rather than general coding benchmarks, making it a distinct and specialized evaluation axis.

For AI practitioners and security teams, this finding reinforces the urgency of moving beyond general-purpose benchmarks (MMLU, GPQA, HumanEval) toward capability-specific evaluations. Standard leaderboards rank GPT-5.5 and Claude Opus 4.7 at the top, but those rankings don't capture the full safety-relevant picture. The Air Street report also noted that China's open-weight coding models have now reached Western frontier-level performance — meaning these accelerating cyber-offense capabilities are no longer confined to closed, auditable systems.

The practical implication: organizations relying on benchmark scores alone for deployment decisions need supplementary evaluation frameworks that directly probe for dual-use risks, particularly as open-weight models close the gap and become widely deployable without API gatekeeping.

1 sources

substackcdn.com

substackcdn.com

Analysis & Trends

• State of the art: OpenAI's GPT-5.5 (xhigh) holds the top intelligence ranking across all tracked models. For coding and reasoning, GPT-5.5 variants and Claude Opus 4.7 lead closed models; DeepSeek V4 and Qwen 3.5 are now competitive at the frontier for open-weight options.
• Open vs. Closed gap: The gap has narrowed dramatically. DeepSeek V4 has "almost closed" the gap on reasoning benchmarks per TechCrunch reporting; China's open-weight coding models are described as reaching "Western frontier" level in the Air Street May 2026 report. Five frontier-class open-weight models shipped in a single 30-day window (Llama 4, Qwen 3.5, DeepSeek V4, Gemma 4, Mistral Medium 3.5), an unprecedented pace.
• Cost-performance: Qwen3.5 0.8B at $0.02/1M tokens blended is the cheapest reasoning model tracked by Artificial Analysis. DeepSeek V4 has again cut prices, reigniting competition at the low end of the cost curve. Mercury 2 leads on raw inference speed at 902.3 tokens/sec for latency-sensitive workloads.
• Emerging patterns: The AI Safety Institute's finding that frontier cyber-offense capability doubles every ~4 months is the most urgent emerging signal this week. Separately, the volume and pace of open-weight frontier releases — five in 30 days — suggests the open-source ecosystem has entered a new phase of competitiveness with closed labs.

What to Watch Next

• Anthropic's deployment of Claude Opus 4.7 at scale: With a fresh $50B capital raise reported this week, Anthropic has the runway to aggressively expand Claude Opus 4.7 availability and push further on reasoning benchmarks. Watch for benchmark updates as the model rolls out more broadly.
• AISI cyber-offense benchmark methodology disclosure: The Air Street report surfaced the doubling-every-4-months finding, but the full AISI evaluation methodology has not been publicly released. If AISI publishes its framework, it could become a new standard axis for frontier model safety evaluation — shifting how leaderboards report risk alongside capability.
• DeepSeek V4 and Qwen 3.5 independent third-party evaluations: Forbes and codersera have reported on these models, but independent rigorous benchmarking (GPQA, MATH, HumanEval) from third parties like Artificial Analysis or Hugging Face's Open LLM Leaderboard is still emerging. Those evaluations will determine whether the "frontier parity" claims hold up under standardized testing.