GPT‑5.

I've been running benchmarks on GPT‑5.6 for a couple of weeks now, and the number that matters most—cost per useful token—has genuinely dropped without much sacrifice. The model lands somewhere between GPT‑5 and GPT‑4o on raw accuracy but at about a third of GPT‑5's inference cost for large batches. That's not incremental, that's a different budget conversation.

The architecture shift is the story. They've gone with a sparse mixture-of-experts variant that only activates around 15% of parameters per forward pass. My latency numbers show first-token time dropped from 0.7s (GPT‑5) to 0.4s, and sustained throughput is up roughly 60% in my streaming setups. For the price tier I'm on:

• Input cost: $1.20 per million tokens (GPT‑5 was $3.00)
• Output cost: $4.80 per million tokens (GPT‑5 was $6.00)
• MMLU: 89.3% vs 90.1%
• GSM8K reasoning: 94.7% vs 95.2%
• HumanEval pass@1: 86.5% vs 87.1%

The delta in accuracy is under one percent on standard evals, which in practice means I can't tell the difference on RAG pipelines, summarization, or multi-turn dialogue. The big win is the cost: for my daily API bill, I'm getting about 2.5x more completions for the same spend.

What caught me off guard was the default speculative decoding mode. The API automatically pre-generates candidate tokens via a smaller drafter, so perceived latency in chat-like interactions is even lower than the raw numbers suggest. No config change needed—just hit the gpt-5.6 endpoint.

For anyone running cost-sensitive pipelines or real-time features, this is the first frontier model that doesn't make me recalculate margins every deployment. Migration was seamless from GPT‑5—same endpoint pattern, same parameters. The rate limits on Pro tier are also doubled, so batch jobs finish faster without throttling.

Worth spinning up a test run if you were sitting on the fence due to budget. The price-performance curve finally bends where it matters.