Anthropic and OpenAI price their near-flagship models at $2 and $10 per million tokens while open-weight Qwen runs on a gaming GPU

Within one week, Anthropic and OpenAI each released a second-tier model that their own evaluations place within a few points of their flagship. Both set the same list price. Anthropic released Claude Sonnet 5.5 on September 28 at $2 per million input tokens and $10 per million output tokens, the same rates as Sonnet 5. The next day, at its DevDay developer event, OpenAI shipped GPT-6.1 Sol at $2 input, $0.10 cached input and $10 output. Those rates are one-fifth of what OpenAI charges for its flagship, GPT-6 Astra.

The gains Anthropic reports are large. According to The Decoder's summary of Anthropic's figures, Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0, up from 10.3% for Sonnet 5 and above Opus 5.5's 66.4%. On CursorBench 4.0 it scores 55.5%, against 34.1% for Sonnet 5 and 57.8% for Opus 5.5. It scores 1844 on GDPval-AA against Opus 5.5's 1846, and 80.1% on OSWorld 2.1 against 81.8%. These are vendor numbers. The only independent result in hand is the Vals Index, which ranks Sonnet 5.5 second of 43 models at 67.04%.

Opus 5.5 arrived on September 22. Anthropic says it performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5. List prices fell 20%, to $4 input and $20 output per million tokens, and cache reads fell 60%, to $0.20. One breakdown estimates that most of the 40% saving comes from cheaper cache reads in long agent sessions. By that estimate, a short chat with little cached context saves closer to 20%.

OpenAI reports on the same basis, cost per completed task. At maximum effort, GPT-6.1 Sol scores 57.0% on Terminal-Bench at $5.47 per task, against 27.6% at $12.18 for GPT-6 Sol. The two vendors report different Terminal-Bench configurations, so the Sol and Sonnet scores cannot be compared directly.

Open-weight models are also getting cheaper to run, which adds competitive pressure on these API prices. Alibaba's Qwen team released Qwen3.8-Flash-Next on August 26 as a preview of its Qwen4 architecture. It is a sparse MoE with 125B main parameters, plus 51B parameters of N-gram embeddings, and it activates about 6B parameters per token. It supports a 262K-token context. The Qwen team says training cost about one-ninth as much as Qwen3.7-Plus. An open-source engine called Strata keeps the most-used experts in GPU memory, holds the full expert set in system RAM and serves the N-gram table from SSD. Its developer reports about 100 tokens per second on a single RTX 4090, and other users report 40 to 60 tokens per second on an RTX 4080 Super. Unsloth's documentation says the model outperforms Claude 4.6 Opus. That comparison is with a model two Anthropic generations old, and no independent lab has reproduced it.