GPT 5.5 Arrives, DeepSeek V4 Drops, and the Compute War Intensifies

TL;DR

GPT 5.5 and DeepSeek V4 launch simultaneously, each challenging the other on benchmarks, cost, and compute access in an intensifying AI race.

Key Points

• 1.GPT 5.5 shows jagged benchmark results, not a clean win. It underperforms Opus 4.7 by ~6% and Mythos Preview by ~20% on SWEBench Pro, but beats Mythos on Agentic Terminal Coding (82.7% vs 82.0%) and leads on ARC-AGI 2 — no API access yet means all scores are self-reported by OpenAI.
• 2.GPT 5.5 has a severe hallucination problem despite high accuracy. It answers 57% of obscure knowledge questions correctly vs Opus 4.7's 46%, but hallucinates 86% of wrong answers rather than admitting ignorance — compared to Opus 4.7's 36% hallucination rate on incorrect responses.
• 3.GPT 5.5 matches Mythos on cyber capabilities but with far less fanfare. The UK AI Security Institute ranks it strongest overall on narrow cyber tasks; it completed a 32-step corporate network attack in 1/10 attempts vs Mythos's 3/10, raising questions about OpenAI's quieter safety posture relative to Anthropic.
• 4.DeepSeek V4 Pro is a formidable, ultra-cheap rival launching amid compute constraints. It supports 1 million token context, has 1.6 trillion parameters (49B activated via MoE), costs roughly one-tenth of Opus 4.7, scored 61.2% on SimpleBench near Opus 4.7's level, and outperforms GPT 5.2 and Gemini 3 Pro on reasoning and coding benchmarks.
• 5.DeepSeek V4 dominates Chinese professional tasks, exposing language-domain specialization limits. In a blind evaluation across 30 advanced Chinese professional domains — finance, law, education, tech — V4 Pro Max beat Opus 4.6 Max significantly, suggesting specialized training data outperforms general intelligence scaling.
• 6.The compute war is intensifying, with OpenAI holding a major infrastructure advantage. Anthropic failed to anticipate its own success and faces a compute crunch; DeepSeek says V4 Pro capacity is extremely limited; Greg Brockman admitted OpenAI itself is entering an era of compute scarcity and rate limits are already visible.
• 7.Recursive self-improvement remains far off, despite GPT 5.5's bio and cyber capability jumps. OpenAI found 5.5 has no plausible path to high-threshold self-improvement — internal debugging tests show ~50% success at 8-hour tasks and ~6% at one-day tasks, similar to GPT 5.3 and 5.4, meaning autonomous AI research acceleration is not yet viable.