GPT-4o vs Llama 3.3 70B Instruct
Wins 3 of 6 areas
Coding · Facts · Long documents
Wins 1 of 6 areas
Following instructions
GPT-4o wins more areas, narrowly.Llama 3.3 70B Instruct is cheaper and better at following instructions.
Scores updated · 31 tests both models report · How we compare
Where each one wins
Tests won in each of the six 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.
- CodingWriting and fixing software20GPT-4o2 of 2 tests
- FactsGetting facts right instead of making them up21GPT-4o2 of 3 tests
- Long documentsFinding answers in very long texts10GPT-4o1 of 1 test
- Following instructionsDoing exactly what it is asked01Llama 3.3 70B Instruct1 of 1 test
- AgentsCarrying out multi-step tasks on its own00Even0 each · 1 tie
- ReasoningHard problems that need careful thinking11Even1 each · 1 tie
Agents, 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.
Llama 3.3 70B Instruct costs 93% less for the same work.
The biggest differences
The tests each model wins by the widest margin, up to three each. Scores are out of 100.
Where GPT-4o pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+52.3points ahead
- Reasons across sets of long documentsAA-LCR+40.3points ahead
- Code for real scientific research problemsSciCode+7.3points ahead
Where Llama 3.3 70B Instruct pulls ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+11.1points ahead
- Very hard expert questions across many subjectsHumanity's Last Exam+1.8points ahead
Every test, side by side
All 31 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingGPT-4o
- SciCodeGPT-4o by 7.333.326+7.3
- Terminal-Bench HardGPT-4o by 5.38.33+5.3
ReasoningEven
- GPQA DiamondGPT-4o by 2.852.649.8+2.8
- Humanity's Last ExamLlama 3.3 70B Instruct by 1.81.83.6+1.8
- CritPttie00tie
FactsGPT-4o
- AA-Omniscience · Non-hallucinationGPT-4o by 52.362.19.8+52.3
- Vectara HHEM hallucination ratelower is betterLlama 3.3 70B Instruct by 5.59.64.1+5.5
- AA-Omniscience · AccuracyGPT-4o by 4.723.718.9+4.7
Following instructionsLlama 3.3 70B Instruct
- IFBenchLlama 3.3 70B Instruct by 11.13647.1+11.1
Other results20 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-OmniscienceGPT-4o by 43.7-10.5-54.2+43.7
- vectara_avg_summary_lengthGPT-4o by 2286.664.6+22
- SimpleQAGPT-4o by 17.338.220.9+17.3
- Artificial Analysis Coding IndexGPT-4o by 12.324.211.9+12.3
- MMLULlama 3.3 70B Instruct by 12.273.886+12.2
- IFEvalLlama 3.3 70B Instruct by 11.18192.1+11.1
- MATHGPT-4o by 8.385.377+8.3
- DROPLlama 3.3 70B Instruct by 6.883.490.2+6.8
- naturalquestions_closedbookGPT-4o by 6.549.643.1+6.5
- MMLU-ProGPT-4o by 5.874.768.9+5.8
- vectara_answer_rateLlama 3.3 70B Instruct by 5.793.899.5+5.7
- vectara_factual_consistencyLlama 3.3 70B Instruct by 5.590.495.9+5.5
- LiveCodeBenchGPT-4o by 538.333.3+5
- OpenBookQAGPT-4o by 496.892.8+4
- GSM8KLlama 3.3 70B Instruct by 3.390.994.2+3.3
- τ²-Bench Telecom (AA run)GPT-4o by 2.328.926.6+2.3
- HumanEvalGPT-4o by 1.890.288.4+1.8
- MGSMtie90.591.1tie
- AA Intelligencetie7.37.7tie
- NarrativeQAtie79.579.1tie
Questions people ask
Which is better, GPT-4o or Llama 3.3 70B Instruct?
GPT-4o wins three of the six areas where both have results: coding, facts and long documents. Llama 3.3 70B Instruct wins following instructions, and costs 93% less. They are level on agents and reasoning.
Which is better for coding?
GPT-4o. It wins 2 of the 2 coding tests both models report; Llama 3.3 70B Instruct wins none.
Which is cheaper?
GPT-4o costs $5.00 per million input tokens and $15.00 per million output tokens; Llama 3.3 70B Instruct costs $0.71 and $0.72. That makes Llama 3.3 70B Instruct about 93% cheaper for the same work.
How do you compare the two?
We use the 31 benchmark tests both models have published scores on. The verdict counts the 11 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 20 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.