Claude Opus 5.5 vs GPT-6 Astra
Wins 4 of 8 areas
Coding · Reasoning · Images and charts · Long documents
Wins 3 of 8 areas
Facts · Math · Following instructions
Claude Opus 5.5 wins more areas, narrowly.GPT-6 Astra is better at facts.
Both rank among the ten best models we track in reasoning, agents, facts and math.
Scores updated · 81 tests both models report · How we compare
Where each one wins
Tests won in each of the eight areas we test. 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 software40Claude Opus 5.54 of 4 tests
- ReasoningHard problems that need careful thinking21Claude Opus 5.52 of 5 tests · 2 ties
- Images and chartsUnderstanding pictures, charts and video10Claude Opus 5.51 of 2 tests · 1 tie
- Long documentsFinding answers in very long texts10Claude Opus 5.51 of 1 test
- FactsGetting facts right instead of making them up13GPT-6 Astra3 of 4 tests
- MathCompetition and research-level math02GPT-6 Astra2 of 3 tests · 1 tie
- Following instructionsDoing exactly what it is asked01GPT-6 Astra1 of 1 test
- AgentsCarrying out multi-step tasks on its own11Even1 each · 2 ties
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.
Claude Opus 5.5 costs 60% 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 Opus 5.5 pulls ahead
- Real work tasks from 44 professionsGDPVal+14.7points ahead
- Fixes real GitHub issues, working as an agentLiveBench · Agentic Coding+14.4points ahead
- Code for real scientific research problemsSciCode+10.4points ahead
Where GPT-6 Astra pulls ahead
- Finds the root cause of IT system incidentsAA IT-Bench SRE+10.4points ahead
- Follows detailed instructions exactlyLiveBench · Instruction Following+9.9points ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+7.3points ahead
Every test, side by side
All 81 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingClaude Opus 5.5
- LiveBench · Agentic CodingClaude Opus 5.5 by 14.471.757.3+14.4
- SciCodeClaude Opus 5.5 by 10.466.956.5+10.4
- LiveBench · CodingClaude Opus 5.5 by 8.989.380.4+8.9
- LMArena · WebDevClaude Opus 5.5 by 51 rating points18151764+51 rating
AgentsEven
- GDPValClaude Opus 5.5 by 14.768.353.6+14.7
- AA IT-Bench SREGPT-6 Astra by 10.438.248.6+10.4
- BrowseComptie90.891.5tie
- Terminal-Bench 4.0tie59.659.1tie
ReasoningClaude Opus 5.5
- Humanity's Last ExamClaude Opus 5.5 by 6.761.454.7+6.7
- SimpleBenchClaude Opus 5.5 by 4.888.483.6+4.8
- ARC-AGI-2GPT-6 Astra by 3.391.795+3.3
- LiveBench · Reasoningtie92.292.7tie
- CritPttie31.731.7tie
FactsGPT-6 Astra
- AA-Omniscience · Non-hallucinationGPT-6 Astra by 7.341.448.7+7.3
- AA-Omniscience · AccuracyClaude Opus 5.5 by 3.666.262.6+3.6
- SimpleQA VerifiedGPT-6 Astra by 3.472.275.6+3.4
- Vectara HHEM hallucination ratelower is betterGPT-6 Astra by 1.29.98.7+1.2
Images and chartsClaude Opus 5.5
- LMArena · VisionClaude Opus 5.5 by 45 rating points13241279+45 rating
- MMMU-Protie87.786.9tie
MathGPT-6 Astra
- FrontierMath Tier 4GPT-6 Astra by 2.69597.6+2.6
- FrontierMath Tiers 1-3 (v2)GPT-6 Astra by 2.591.293.7+2.5
- LiveBench · Mathematicstie97.196.8tie
Following instructionsGPT-6 Astra
- LiveBench · Instruction FollowingGPT-6 Astra by 9.965.775.6+9.9
Other results57 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.
- Program BenchClaude Opus 5.5 by 41.291.250+41.2
- GDPval-AA 2.1Claude Opus 5.5 by 304 rating points18461542+304 rating
- GDPval-AA v2.1 (max effort)Claude Opus 5.5 by 282 rating points18661584+282 rating
- HLE (with tools)Claude Opus 5.5 by 10.567.757.2+10.5
- OSWorld 2.0Claude Opus 5.5 by 9.281.872.6+9.2
- Terminal-Bench 4.0Claude Opus 5.5 by 8.766.457.7+8.7
- Agentic coding (Terminal-bench 4.0)Claude Opus 5.5 by 8.266.458.2+8.2
- GraphWalks (Up to 128k, BFS (F1))GPT-6 Astra by 8.190.698.7+8.1
- OSWorld 2.1 (Offline set, partial)Claude Opus 5.5 by 7.787.279.5+7.7
- Computer use (OSWorld-2.1)Claude Opus 5.5 by 6.387.280.9+6.3
- HiL-Bench (Tools-allowed)GPT-6 Astra by 6.35965.3+6.3
- HealthBench ProfessionalGPT-6 Astra by 6.159.966+6.1
- General capabilities (Vals Index v2.2)Claude Opus 5.5 by 5.270.265+5.2
- GraphWalks (256k to 1M, BFS (F1))GPT-6 Astra by 566.871.8+5
- PostTrainBenchClaude Opus 5.5 by 549.344.3+5
- AA IntelligenceClaude Opus 5.5 by 4.957.652.7+4.9
- GDP (Surge AI)GPT-6 Astra by 4.826.231+4.8
- HealthBench Professional length-adjustedGPT-6 Astra by 4.859.964.7+4.8
- Science and math (Terminal-Bench Science 0.1)GPT-6 Astra by 4.863.368.1+4.8
- vectara_avg_summary_lengthClaude Opus 5.5 by 43.6 rating points192.1148.5+43.6 rating
- Science and math (Terminal-Bench-Science 0.1 (AA))GPT-6 Astra by 4.35963.3+4.3
- Terminal-Bench-Science 0.1GPT-6 Astra by 4.35963.3+4.3
- HLE-DiamondClaude Opus 5.5 by 4.167.963.8+4.1
- Agent's Last Exam (Pass rate)Claude Opus 5.5 by 438.234.2+4
- Computer use (Agent's Last Exam)Claude Opus 5.5 by 438.234.2+4
- Knowledge work (Vals Index)Claude Opus 5.5 by 3.96763.1+3.9
- Vals IndexClaude Opus 5.5 by 3.96763.1+3.9
- LVBenchGPT-6 Astra by 3.883.787.5+3.8
- Multimodal understanding (LVBench)GPT-6 Astra by 3.883.787.5+3.8
- SWE Marathon v1.1Claude Opus 5.5 by 3.846.342.5+3.8
- livebench_languageGPT-6 Astra by 3.186.389.4+3.1
- AA-OmniscienceClaude Opus 5.5 by 346.443.4+3
- ZeroBench_main w/ python (pass@5)Claude Opus 5.5 by 36158+3
- livebench_data_analysisGPT-6 Astra by 2.780.383+2.7
- CritPt v2 HardGPT-6 Astra by 2.56870.5+2.5
- RiemannBenchGPT-6 Astra by 2.469.672+2.4
- Science and math (RiemannBench)GPT-6 Astra by 2.469.672+2.4
- ChartoGraphy w/ pythonGPT-6 Astra by 28991+2
- mmmu_pro_with_pythonClaude Opus 5.5 by 289.887.8+2
- PostTrain Bench v1.2Claude Opus 5.5 by 1.943.841.9+1.9
- Computer use (Agents' Last Exam v1.1)GPT-6 Astra by 1.571.673.1+1.5
- HealthBench length-adjustedGPT-6 Astra by 1.35758.3+1.3
- HealthBenchGPT-6 Astra by 1.25758.2+1.2
- vectara_factual_consistencyGPT-6 Astra by 1.290.191.3+1.2
- AutomationBench (Score)Claude Opus 5.5 by 1.142.541.4+1.1
- baby_vision_with_pythonClaude Opus 5.5 by 1.197.796.6+1.1
- Knowledge work (AutomationBench)Claude Opus 5.5 by 1.142.541.4+1.1
- charxiv_rq_with_pythonClaude Opus 5.5 by 193.892.8+1
- Cybersecurity (CWE-bench v1)GPT-6 Astra by 16768+1
- Cybersecurity (CWE-bench)GPT-6 Astra by 16768+1
- General capabilities (AA Index v5.0)Claude Opus 5.5 by 15857+1
- Agentic coding (Terminal-Bench 4.0 (AA))tie59.659.1tie
- Remote Labor Index (RLI)tie21.320.8tie
- BenchCAD w/ toolstie96.295.9tie
- vectara_answer_ratetie99.7100tie
- DeepSWE 1.1tie74.274.1tie
- ARC-AGI-1tie97.597.5tie
Questions people ask
Which is better, Claude Opus 5.5 or GPT-6 Astra?
Claude Opus 5.5 wins four of the eight areas we test: coding, reasoning, images and charts and long documents. GPT-6 Astra wins facts, math and following instructions. They are level on agents.
Which is better for coding?
Claude Opus 5.5. It wins 4 of the 4 coding tests both models report; GPT-6 Astra wins none.
Which is cheaper?
Claude Opus 5.5 costs $4.00 per million input tokens and $20.00 per million output tokens; GPT-6 Astra costs $10.00 and $50.00. That makes Claude Opus 5.5 about 60% cheaper for the same work.
How do you compare the two?
We use the 81 benchmark tests both models have published scores on. The verdict counts the 24 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 57 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.