Cloud API Throughput: Claude vs GPT-5.6

OpenAI's GPT-5.6 family (launched July 9, 2026) ships in three durable capability tiers named after celestial bodies: Sol (flagship), Terra (balanced mid-tier), Luna (fast, cheap). All three are available in ChatGPT, Codex, and the API, and generally available on Amazon Bedrock. Anthropic's current lineup for comparison: Claude Opus 5 (flagship) and Claude Sonnet 5 (mid-tier), both defaults in Claude Code.

Model	Typical throughput	Max-effort/benchmark throughput	Time to first token	Price ($/M tokens, in/out)
GPT-5.6 Sol	≈70 tok/s (OpenRouter P50)	74.3 tok/s (ArtificialAnalysis)	≈138.6s in max reasoning mode	$5 / $30
GPT-5.6 Terra	≈58 tok/s (OpenRouter P50)	n/a	≈2.48s	$2 / $12 (cut 20% from launch, effective July 30, 2026)
GPT-5.6 Luna	≈112 tok/s (OpenRouter P50)	140.7 tok/s (ArtificialAnalysis)	≈150.7s in max reasoning mode	$0.20 / $1.20 (cut 80% from launch, effective July 30, 2026)
Claude Sonnet 5	≈75.9 tok/s (ArtificialAnalysis, max effort)	n/a	not published	$2 / $10
Claude Opus 5	≈26 tok/s average (LLM-Benchmarks production telemetry, min 5.3, max 60.9)	n/a	≈2.87s	$5 / $25 (fast mode research preview: $10 / $50)

Claude Opus 5's ≈26 tok/s average comes from production telemetry across real calls, not a single benchmark preset, a materially lower number than the ≈55-80 tok/s figures reported for Opus 5 under lighter-effort presets on other trackers. Anthropic optimizes Opus 5 for reasoning depth, not raw streaming speed.

On specialized hardware (Cerebras wafer-scale systems, not a standard cloud GPU), Sol reaches roughly 750 tok/s, about 10x the typical cloud API rate. This confirms the throughput ceiling is set by deployment infrastructure and multi-tenant scheduling, not by the model's architecture alone.

Sources: OpenRouter Sol, OpenRouter Terra, OpenRouter Luna, ArtificialAnalysis Sol, ArtificialAnalysis Luna, ArtificialAnalysis Sonnet 5, LLM-Benchmarks Anthropic, OpenAI price-performance update, Cerebras.