GPT-5.6 hit general availability on July 9. Opus 5 shipped July 24. That's a fifteen-day gap between two flagship model releases, and by the time I sat down to actually pick one for my agentic coding work, every comparison I found online was either a vendor's own benchmark page or a hot take with no sourcing at all.
So instead of running my own multi-week test suite (which would need at least a month to mean anything), I spent an evening doing something less exciting: checking which numbers about these two models are actually cross-referenced by a third party, and which ones are still just marketing copy.
Here's what held up.
BenchLM.ai, an independent leaderboard that ranks 215 models, has Opus 5 at #1 overall (85.88) and GPT-5.6 Sol at #4 (81.46). That's a third-party ranking, not a vendor claim, which is why I trust it more than the release blog posts. On coding specifically, Opus 5 posts 96.0% on SWE-bench Verified and 79.2% on SWE-bench Pro. GPT-5.6's tiers (Sol/Terra/Luna) land at 64.6% / 63.4% / 62.7% on SWE-bench Pro by the same measure, a real gap, not a rounding error.
Where it gets murkier: GPT-5.6 hasn't had SWE-bench Verified numbers published for its tiers at all, and OpenAI's own tier documentation for Terra and Luna was still incomplete as of this writing, so the Terminal-Bench figures floating around (88.8% for Sol, ~82-87% for the cheaper tiers) come from third-party sites cross-checking each other, not from OpenAI directly. I'm not calling that a red flag, new releases are always messy on the docs side, but it's a different confidence level than the SWE-bench numbers above.
Pricing is the other half of the decision and it's not close to a wash. Opus 5 is $5/$25 per million tokens, one tier, adaptive thinking baked in. GPT-5.6 Sol is $5/$30, but Luna drops to $1/$6, a sixth of Opus 5's output cost, if you don't need frontier-level coding performance for whatever you're routing to it. If GPT-5.6 has one genuinely distinct feature, it's "ultra mode," which farms a single hard task out to a subagent swarm instead of one long context window. Opus 5 doesn't have an equivalent; it just has a bigger, adaptive-thinking context.
My actual takeaway, for what it's worth: if the job is agentic coding against a real codebase, Opus 5's SWE-bench Pro number is a big enough gap that I switched my default for that use case. If the job is high-volume, lower-stakes generation where GPT-5.6 Luna's price makes ten runs cheaper than one Opus 5 run, I kept GPT-5.6 in the rotation. Neither model has been out long enough for anyone to have run it against a real production workload for more than a couple weeks, so treat all of this, mine included, as a snapshot, not a verdict.
I wrote up the full side-by-side, every tier, every published metric, sourcing notes, here: Claude Opus 5 vs GPT-5.6 comparison
Top comments (0)