LLM pricing briefing: March 23, 2026

The cheapest models worth using keep getting cheaper. This week, the sub-$2/M tier gained real quality contenders, and the premium tier is starting to look hard to justify for most workloads.

The value tier is no longer a compromise

MiniMax M2.7 (MiniMax) sits at $0.52/M input tokens with a 49.6 quality index. MiniMax M2.7 scores within striking distance of GPT-5.2 (51.3 quality) at roughly one-ninth the price. The 43 tok/s inference speed is the trade-off — batch-heavy workloads won't mind, but interactive use will feel sluggish.

Z.ai: GLM 5 (Z AI) lands at $1.11/M with 49.8 quality and 89 tok/s. GLM 5 is open source, which matters for self-hosting economics. At that quality-to-price ratio, it's the strongest open-source option on the board right now.

OpenAI: GPT-5.4 Mini (OpenAI) at $1.69/M delivers 48.1 quality and a blistering 237 tok/s. GPT-5.4 Mini is the speed leader by a wide margin — nearly double the next fastest model in this set. For latency-sensitive applications where you need fast iteration, nothing else comes close.

Model	Quality	Price/1M	Speed
MiniMax M2.7	49.6	$0.52	43 tok/s
GLM 5	49.8	$1.11	89 tok/s
GPT-5.4 Mini	48.1	$1.69	237 tok/s
GPT-5.4	57.2	$5.63	85 tok/s
Claude Opus 4.6	53.0	$10.00	51 tok/s

Premium models: tied at the top, split on price

Gemini 3.1 Pro Preview (Google) and GPT-5.4 (OpenAI) share an identical 57.2 quality index. Google undercuts OpenAI by $1.13/M ($4.50 vs. $5.63) and runs faster at 117 tok/s versus 85 tok/s. For general-purpose premium work, Gemini 3.1 Pro is the better deal on both axes.

Anthropic's adaptive reasoning models remain the most expensive options at $10.00/M. Claude Opus 4.6 hits 53.0 quality — 4 points below the leaders. That's a tough sell at nearly double the price unless your workflow specifically benefits from Anthropic's reasoning approach.