Two frontier coding models shipped inside the same two-week window: OpenAI's GPT-5.6 Sol went public on July 9, and Anthropic's Claude Opus 5 landed on July 24. Both cost $5 per million input tokens. Both sit at the top of the coding leaderboards, separated by less than a point. If you're picking the backend for a coding agent this week, the benchmark chart won't decide it for you.
If you read one line: they're a capability tie, so pick the model whose harness you already build in — Claude Code and Claude Max default to Opus 5; Codex and ChatGPT Work default to the GPT-5.6 family — then let two cost levers (Opus's cheaper output vs Sol's 90% cache discount) break the tie on your own workload.
The benchmarks are a tie — stop treating them as the decision#
Here's the head-to-head on the numbers everyone cites:
- SWE-bench Verified: Opus 5 leads at 97.0%, GPT-5.6 Sol at 96.2% (Vellum; Vals AI). That's inside the run-to-run noise of a single benchmark.
- Terminal-Bench 2.1: Sol posts 88.8% in standard mode — and 91.9% only in the heavier Sol Ultra parallel-subagent mode that costs materially more per task and isn't the default (Artificial Analysis).
- Agentic knowledge work: Opus 5 is the new leader on AA-Briefcase; Sol (max) tops the Artificial Analysis Coding Agent Index.
Read together, they trade wins depending on which harness the benchmark runs. Neither model has a capability lead you'd feel on real work. The interesting fight is the one benchmark charts hide: what a completed task actually costs.
Cost: identical input, two different levers#
Input is a wash — $5 per million tokens on both. From there they diverge:
| Lever | Opus 5 | GPT-5.6 Sol |
|---|---|---|
| Output price (per 1M) | $25 | $30 |
| Caching | Prompt caching | 90% discount on cache reads |
Opus 5 is $5 cheaper on output, so an output-heavy run — long diffs, verbose plans — costs less on Opus. But a coding agent's token bill is rarely output-dominated. It's a loop: the same system prompt, tool schemas, and file context get resent on every step. That's almost all cache reads — exactly where GPT-5.6's 90% cache discount bites. For a cache-heavy agent, Sol can come out cheaper per completed task despite the higher sticker output rate.
Neither model wins this in the abstract. The tiebreaker is your loop's shape, which is why the only number that matters is cost-per-accepted-diff on your own repo, not cost-per-token on a pricing page — the same trap that makes one tokens-per-second figure lie to you.
The real tiebreaker: the harness and the route-down#
Both models live in a family, and that's where the practical decision is made.
- GPT-5.6 routes down in-family: Terra ($2.50 / $15) for everyday agent steps, Luna ($1 / $6) for bulk classification and extraction. If you already think in the Sol / Terra / Luna tier menu, Sol is the top of a ladder you're already on.
- Opus 5 pairs with Sonnet 5 and Haiku for the same cheap-step work, and it's the new default at the old Opus price — a drop-in upgrade if your agent is already on the Anthropic API, with a per-request effort toggle to dial reasoning compute down on easy calls.
The move almost nobody should make is switching harnesses to chase a sub-point benchmark gap. Reworking tool definitions, prompts, and your eval suite to move from Codex to Claude Code (or back) costs more than the 5–10% you'd gain. Let the ecosystem you already build in win the tie. Greenfield, or already unhappy with your current stack? Then the choice is open — and it comes down to which family's cheap tiers and caching you'd rather run, not which flagship tops one chart.
What a founder should actually do#
- Default to your harness. On Claude Code / Claude Max → Opus 5. On Codex / ChatGPT Work → GPT-5.6 Sol. This decides 80% of cases correctly on its own.
- Benchmark cost-per-accepted-diff, not per token, on a real slice of your repo. Run both for a day if you're genuinely undecided.
- Check your loop's cache ratio. Repeated-context loops favor Sol's 90% cache discount; output-heavy one-shots favor Opus's cheaper output.
- Route down aggressively. Whichever flagship you pick, most steps don't need it — put Terra/Luna or Sonnet/Haiku on the routine work and reserve the frontier model for planning and hard diffs. (See Kimi K3 vs Opus vs GPT-5.6 on cost if an open-weight tier is also on the table.)
Two frontier models, one price floor, a benchmark tie. In 2026 the model is no longer the hard part of the decision — the harness and the cost structure are.
