Claude 3.5 Haiku vs Qwen2.5 Instruct 72B
Wins 2 of 5 areas
Long documents · Following instructions
Wins 1 of 5 areas
Reasoning
Claude 3.5 Haiku wins more areas, narrowly.Qwen2.5 Instruct 72B is cheaper and better at reasoning.
Scores updated · 25 tests both models report · How we compare
Where each one wins
Tests won in each of the five areas where both have results. Each piece is one test, so a longer bar means more evidence; grey means the two scored within a point of each other.
- Long documentsFinding answers in very long texts10Claude 3.5 Haiku1 of 1 test
- Following instructionsDoing exactly what it is asked10Claude 3.5 Haiku1 of 1 test
- ReasoningHard problems that need careful thinking01Qwen2.5 Instruct 72B1 of 3 tests · 2 ties
- CodingWriting and fixing software11Even1 each · 1 tie
- FactsGetting facts right instead of making them up11Even1 each
Long documents and following instructions rest on a single test each.
What it costs
Prices per million tokens, roughly 750,000 words. The bars show the cost of a million tokens read plus a million written.
Qwen2.5 Instruct 72B costs 80% less for the same work.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where Claude 3.5 Haiku pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+44.4points ahead
- Fixes real GitHub issues in Python projectsSWE-bench Verified+16.8points ahead
- Reasons across sets of long documentsAA-LCR+7.3points ahead
Where Qwen2.5 Instruct 72B pulls ahead
- Graduate-level biology, physics and chemistry questionsGPQA Diamond+8.3points ahead
- Answers hard knowledge questions correctlyAA-Omniscience · Accuracy+4.3points ahead
- Hard command-line tasks in a real terminalTerminal-Bench Hard+2.2points ahead
Every test, side by side
All 25 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingEven
- SWE-bench VerifiedClaude 3.5 Haiku by 16.840.623.8+16.8
- Terminal-Bench HardQwen2.5 Instruct 72B by 2.22.34.5+2.2
- SciCodetie27.426.7tie
ReasoningQwen2.5 Instruct 72B
- GPQA DiamondQwen2.5 Instruct 72B by 8.340.849.1+8.3
- CritPttie00tie
- Humanity's Last Examtie3.63.6tie
FactsEven
- AA-Omniscience · Non-hallucinationClaude 3.5 Haiku by 44.458.914.5+44.4
- AA-Omniscience · AccuracyQwen2.5 Instruct 72B by 4.313.217.5+4.3
Following instructionsClaude 3.5 Haiku
- IFBenchClaude 3.5 Haiku by 5.942.836.9+5.9
Other results15 tests, not counted
Tests outside the eight areas. They are not counted above: several are summary scores built from other tests, or the same test under another name.
- AA-OmniscienceClaude 3.5 Haiku by 30.6-22.5-53.1+30.6
- GSM8KQwen2.5 Instruct 72B by 14.381.595.8+14.3
- OpenBookQAQwen2.5 Instruct 72B by 10.885.496.2+10.8
- MMLUQwen2.5 Instruct 72B by 9.967.177+9.9
- τ²-Bench Telecom (AA run)Qwen2.5 Instruct 72B by 9.924.634.5+9.9
- MGSMClaude 3.5 Haiku by 9.485.676.2+9.4
- DROPClaude 3.5 Haiku by 6.483.176.7+6.4
- MMLU-ProQwen2.5 Instruct 72B by 6.16571.1+6.1
- Artificial Analysis Coding IndexClaude 3.5 Haiku by 415.911.9+4
- MBPPQwen2.5 Instruct 72B by 2.685.688.2+2.6
- NarrativeQAClaude 3.5 Haiku by 1.876.374.5+1.8
- HumanEvalClaude 3.5 Haiku by 1.588.186.6+1.5
- naturalquestions_closedbookQwen2.5 Instruct 72B by 1.534.435.9+1.5
- AA IntelligenceClaude 3.5 Haiku by 1.28.97.7+1.2
- MATHQwen2.5 Instruct 72B by 1.287.288.4+1.2
Questions people ask
Which is better, Claude 3.5 Haiku or Qwen2.5 Instruct 72B?
Claude 3.5 Haiku wins two of the five areas where both have results: long documents and following instructions. Qwen2.5 Instruct 72B wins reasoning, and costs 80% less. They are level on coding and facts.
Which is better for coding?
Neither. They win 1 coding test each of the 3 both models report, and 1 is a tie.
Which is cheaper?
Claude 3.5 Haiku costs $0.80 per million input tokens and $4.00 per million output tokens; Qwen2.5 Instruct 72B costs $0.47 and $0.49. That makes Qwen2.5 Instruct 72B about 80% cheaper for the same work.
How do you compare the two?
We use the 25 benchmark tests both models have published scores on. The verdict counts the 10 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 15 are listed but not counted, because several are summary scores or repeat a test. Each score is the one shown on the model's own page.