What about for the tasks of: data analysis, interpreting business requirements, and technical blog writing? How do the 2 set ups compare?
Worth stating up front: for these three tasks, neither harness is really the intended surface. Claude Code and Codex CLI are coding agents, and the knowledge-work benchmarks below are run through Artificial Analysis’s own harness (Stirrup), not through Claude Code or Codex. GDPval-AA v2 gives models sandbox filesystem shell access and web search through Stirrup — so the model results transfer, the harness results don’t. On the Anthropic side the intended surfaces are Cowork, Claude in Excel, and Claude Design; on the OpenAI side, ChatGPT Work and the Codex App. If you’re doing much of this, that changes the comparison more than the model does.
| Eval | What it measures | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|
| GDPval-AA v2 | Real professional deliverables (docs, sheets, slides), pairwise-graded | 1861 Elo, >100 pts ahead | ~1736 |
| AA-Briefcase | 4 multi-week projects, thousands of input files, 91 rubric-graded tasks | 1720 Elo | 2nd to Fable 5 at its launch; highest Presentation Elo of any model |
| BrowseComp | Agentic web research | 90.8% | 92.2% |
| Turn/token efficiency on Briefcase | Cost of getting there | ~103 turns, 2.0B input tokens | Not published; Sol runs materially leaner |
| Opus 5 / Claude Code | Sol / Codex | |
|---|---|---|
| Strength | Leads both agentic knowledge-work benchmarks, and AA-Briefcase is the closest public proxy for this work — spreadsheet-shaped tasks over large file sets | Best presentation quality of any model on Briefcase’s presentation dimension; leaner token use on long analytical runs |
| Strength | 1M context holds a large dataset plus schema docs plus prior analysis without a retrieval layer | Cheaper to iterate, which matters when you re-run an analysis six times |
| Weakness | Turn count and token burn are the worst in the comparison — roughly 4x some competitors on the same Briefcase work | Codex subagents run in cloud sandboxes with internet disabled, so pip installs and API pulls need setup |
| Weakness | Verbose intermediate output eats plan limits fast on exploratory analysis | Run-to-run variance is a real problem when you want a reproducible analysis, not a plausible one |
| Weakness | Reported weak on factual-knowledge/hallucination evals — one write-up flags a 50% hallucination rate as a hard limit for knowledge-intensive tasks | The METR reward-hacking finding applies directly: an agent that makes the numbers come out right is a specific hazard in analysis |
For your Python work this mostly means: Claude Code runs pandas/matplotlib locally against real data with full filesystem access, Codex sandboxes it. Local access is more convenient and less safe.
This is the one place the benchmark maps cleanly onto the task. AA-Briefcase’s rubric explicitly checks whether the model identified requirements hidden across source files, used the correct evidence, and reached the right conclusions — that is requirements interpretation, graded. Opus 5 leads it.
The qualitative reports point the same way and matter more than the Elo:
Net: Opus 5 is the better default, with a hard rule that it lists assumptions separately from conclusions.
The weakest evidence of the three. There’s no credible benchmark for prose quality, and Presentation Elo measures deliverable formatting — slide layout, document structure — not writing.
| Opus 5 | Sol | |
|---|---|---|
| Factual grounding | Weaker on hallucination evals; will state things confidently that need checking | Slight edge on BrowseComp, so marginally better at going and verifying rather than recalling |
| Prose | Longer, more explanatory by default — needs cutting, not expanding | Leaner output, which some people prefer and some find thin |
| Code examples in the post | Higher SWE-bench Verified means snippets are more likely to actually run | Fine, but the run-to-run variance shows up as inconsistent style across sections |
For a technical post, the failure that costs you is a wrong claim, not a clumsy sentence. Both need fact-checking; neither has an evidence-backed prose advantage. Do this one in chat or Cowork, not in a coding harness — a terminal agent has no reason to be in the loop.
| Task | Pick | Why |
|---|---|---|
| Exploratory data analysis, local data | Claude Code / Cowork | Local execution, best knowledge-work scores, 1M context — accept the token cost |
| Repeated analysis on a fixed spec | Codex | Cheaper per run; the spec constrains the drift problem |
| Turning a vague business ask into a spec | Opus 5, either surface | The AA-Briefcase result is the most directly relevant public evidence in this whole comparison |
| Charts/decks from the analysis | Sol has the presentation edge; Claude Design is the better surface on the Anthropic side | |
| Blog drafting | Either, in chat | No harness needed, and no benchmark separates them |
The honest summary: on published evidence Opus 5 has a real and reasonably well-measured lead on agentic knowledge work, and the lead comes with a cost and verbosity penalty you’ll feel on your plan limits. For the blog writing specifically, treat any claim of a winner — mine included — as unsupported.