Claude Opus 4.8 vs GPT-5.6 Luna for research skills

The verdict

GPT-5.6 Luna wins, 96–90

Same 12 research skills tasks, marked blind by three rival labs. GPT-5.6 Luna took 4 tasks, Claude Opus 4.8 took 2, 6 tied. Tested 16 Aug 2026.

Where they differed most

The task — Structure the evidence

Turn these findings into JSON — reply with ONLY valid JSON, an array of objects with keys source, year, sample_size, finding (finding is one sentence): 'The Whitworth trial (2020) followed 850 runners and found calf strengthening halved recurrence of achilles injury. A 2017 survey by a running magazine of 3,200 readers found 61% had been injured that year. Meffert and Cole's 2023 review of 12 trials concluded the evidence that stretching prevents injury is weak.'

Claude Opus 4.85/10

```json
[
  {
    "source": "Whitworth trial",
    "year": 2020,
    "sample_size": 850,
    "finding": "Calf strengthening halved the recurrence of achilles injury among runners."
  },
  {
    "source": "Running magazine survey",
    "year": 2017,
    "sample_size": 3200,
    "finding": "61% of readers reported having been injured that year."
  },
  {
    "source": "Meffert and Cole review",
    "year": 2023,
    "sample_size": 12,
    "finding": "The evidence that stretching prevents injury is weak across the 12 trials reviewed."
  }
]
```

GPT-5.6 Luna10/10

[
  {
    "source": "Whitworth trial",
    "year": 2020,
    "sample_size": 850,
    "finding": "Calf strengthening halved recurrence of Achilles injury."
  },
  {
    "source": "Running magazine survey",
    "year": 2017,
    "sample_size": 3200,
    "finding": "Sixty-one percent of readers had been injured that year."
  },
  {
    "source": "Meffert and Cole review",
    "year": 2023,
    "sample_size": 12,
    "finding": "The evidence that stretching prevents injury is weak."
  }
]

Task by task

Task	Claude Opus 4.8	GPT-5.6 Luna
Make it answerable	10	10
Rank the sources	9	9
Synthesise the disagreement	9	9
Tear down the survey	9	10
Correlation for the neighbour	10	9
Plan the dig	9	10
What you cannot cite	10	10
Structure the evidence	5	10
Ask without leading	8	10
Two sources, head on	10	9
Primary or secondary	9	9
Refuse the invented bibliography	10	10

Full receipts: Claude Opus 4.8, GPT-5.6 Luna · judges gemini-3.1-pro-preview, gpt-5.6-terra, grok-4.5

Questions people ask

Which is better for research skills: Claude Opus 4.8 or GPT-5.6 Luna?

GPT-5.6 Luna — it scored 96/100 against 90/100 on our 12-task research skills suite, winning 4 tasks to 2 with 6 tied. Every answer was marked blind by three judges from three rival AI labs.

How was this tested?

Both models answered the identical published research skills tasks. Three AI judges from three different labs scored every answer blind against a fixed rubric; mechanically checkable rules (word limits, banned phrases) are enforced by the test harness in code. The raw outputs and judge verdicts are downloadable.

More research skills head-to-heads: DeepSeek V4 Pro vs GPT-5.6 Luna · Claude Opus 4.8 vs DeepSeek V4 Pro · GPT-5.6 Luna vs GPT-5.6 Sol · Claude Opus 4.8 vs GPT-5.6 Sol · GPT-5.5 vs GPT-5.6 Luna · Claude Opus 4.8 vs GPT-5.5

Full ranking: Best AI for research skills · model pages: Claude Opus 4.8, GPT-5.6 Luna