Technology

OpenAI's GPT-5.6 Sol scores highest on the standardized coding and agentic benchmarks all three labs report, including SWE-Bench Verified at 80.0% versus Gemini 3 Pro's 76.2%, while Anthropic declined to publish that benchmark for Claude Opus 4.7 at all. On price, Opus 4.7 undercuts Sol by 17% on output tokens ($25 versus $30 per million), making it the cheaper option for agentic workloads where output dominates cost. None of the three vendors disclose parameter counts, training compute, or architecture details, so the entire capability comparison rests on selectively chosen benchmarks that each lab controls.

Meta's third-generation in-house AI accelerator, Iris, will begin mass production in September 2026 on TSMC's 3nm process, with Samsung negotiating a separate $6.5 billion contract for roughly 40 percent of wafer volume. The chip is designed exclusively for generative AI inference and ranking workloads, where its per-token cost is estimated at half that of an Nvidia H200, while Nvidia GPUs retain all frontier LLM training. Meta simultaneously continues expanding Nvidia and AMD procurement, meaning Iris functions as a cost-control and supplier-diversification tool within a broader $145 billion 2026 infrastructure buildout rather than a displacement strategy.