[V] = vendor self-reported.

[I] = independently measured (Vals AI, Artificial Analysis, Open LLM Leaderboard).

MMLU and GSM8K no longer distinguish between top models. GPQA Diamond and SWE-bench show the real differences.

Model	MMLU	MMLU-Pro	GPQA (D)	HumanEval	MATH/GSM8K	Tier
Proprietary API models
Claude Opus 5	N/A	91.6 [I]	93.4 [I]	N/A	N/A	frontier
GPT-5.6 Sol	N/A	N/A	95.2 [I]	N/A	N/A	frontier
GPT-4o	88.7 [V]	74.7 [I]	53.6 [V]	90.2 [V]	76.6/90.5	legacy
GPT-4o Mini	82.0 [V]	64.0 [I]	40.0 [I]	87.2 [V]	68.0/88.0	budget
Gemini 2.5 Flash	~88 [I]	80 [I]	81 [V]	N/A	AIME 73.3	balanced
Haiku 4.5	N/A	~78 [I]	~66 [I]	N/A	N/A	budget
Open-weight models
Kimi K3 (2.8T MoE)	N/A	N/A	93.5 [V]	N/A	N/A	~Opus 4.8 [V]
DeepSeek R1	90.8 [V]	84.0 [V]	71.5 [V]	N/A	97.3/--	~Flash reasoning
DeepSeek V3	88.5 [V]	75.9 [V]	59.1 [V]	82.6 [I]	90.2/--	~GPT-4o
Llama 3.1 405B	87.3 [V]	73.3 [I]	50.7 [V]	89.0 [V]	70.3/96.8	~GPT-4o
Qwen 2.5 72B	86.1 [V]	64 [V]	49.1 [I]	86 [V]	83.1/95.8	>GPT-4o Mini
Qwen 3 32B	~85 [V]	71.9 [I]	54.6 [I]	-- [V]	62/--	~4o Mini/Flash
Qwen 3 14B	84.7 [V]	70.5 [V]	51.4 [V]	-- [V]	--/--	>GPT-4o Mini
Llama 3.1 70B	86.0 [V]	66.4 [I]	41 [I]	80.5 [V]	68/95	~GPT-4o knowl.
Phi-3 Medium	78.0 [V]	~40 est	low	55 [V]	53.1/91	~Mini (math)
Gemma 2 27B	75.2 [V]	49.4 [V]	26.3 [V]	51.2 [V]	42.1/74.6	~Mini (pref.)
Gemma 2 9B	71.3 [V]	43.7 [V]	24.8 [V]	40.2 [V]	36.4/68.6	~Haiku (chat)
Llama 3.1 8B	69.4 [V]	38 [I]	29 [I]	72.6 [V]	--/84.5	~Haiku/Mini
Phi-3 Mini	68.8 [V]	~30 est	low	58 [V]	41.3/82.5	~Haiku (math)
Mixtral 8x7B	70.5 [V]	~40 [I]	~30 est	40 [V]	28.4/74	GPT-3.5
Mistral 7B	62.5 [V]	~30 est	~28 est	40 [V]	--/45	GPT-3.5

Disclaimer: This is an estimation tool, not financial or technical advice. All prices, benchmarks, and performance figures are approximate and change frequently. Verify current pricing directly with each provider before making decisions.

If you spot an error or outdated number, please let us know at contact@cirqwit.com.