Gemini 3.5 Flash hits 199 tok/s, GPT-5.4 offers half the price of GPT-5.5

What changed this week

Gemini 3.5 Flash (Google) now tops the throughput chart at 199 tok/s with a quality index of 50.2 at $3.38/M tokens. That's 2.5x the inference speed of Claude Sonnet 5 (Anthropic) at 78 tok/s, for 3.2 fewer quality points and 15% lower price. For batch pipelines where wall-clock time dominates, the trade is clear.

GPT-5.4 (OpenAI) is the quiet value play this week. It delivers quality 51.4 at $5.63/M and 153 tok/s. Compared to GPT-5.5 at $11.25/M, quality 54.8, and 87 tok/s, GPT-5.4 costs half as much, runs 1.76x faster, and gives up 3.4 quality points. If your pipeline tolerates a 51-quality baseline — and most retrieval-augmented generation does — GPT-5.4 is the stronger economic choice.

The throughput frontier

GLM 5.2 (Z AI) remains the open-source standout: quality 51.1 at $1.05/M and 177 tok/s. It beats Gemini 3.5 Flash on both quality and price while trailing by 22 tok/s. For teams that need self-hosting or want to avoid vendor lock-in, GLM 5.2 is the model to benchmark against.

Where the frontier models still win

Claude Sonnet 5 at quality 53.4 and $4/M is the cheapest model above the 53-quality line. Nothing in the speed tier matches it on raw quality. If your workload is quality-bound — agentic reasoning, complex code generation, multi-step planning — the 78 tok/s inference latency is the price you pay.

Claude Fable 5 at quality 59.9 and $20/M sits in a different category entirely. It's 9.7 quality points above Gemini 3.5 Flash at 5.9x the cost. Only relevant when output correctness justifies the spend.

Practical read

For throughput-bound batch jobs: Gemini 3.5 Flash or GLM 5.2. For quality-sensitive work under $6/M: Claude Sonnet 5 or GPT-5.4. The gap between the speed tier (50–51 quality) and the quality tier (53–54) is narrow enough that workload characteristics, not benchmark scores, should drive the decision.

What to watch

• Whether GPT-5.4's $5.63 price holds or OpenAI repositions it against GLM 5.2's $1.05.
• If Gemini 3.5 Flash gets a quality bump — at 199 tok/s, even 2 extra points would reshape the mid-tier.
• MiniMax M3 at $0.52/M and 104 tok/s — cheapest in the table, but quality 44.4 limits it to low-stakes classification and extraction.

Use the LLM Selector to match these trade-offs to your specific workload, or Explore the full leaderboard.