Same price, 3.5x the bill: GPT 6.1 Sol vs Claude Sonnet 5.5
Here's a weird one. GPT 6.1 Sol and Claude Sonnet 5.5 launched a day apart (Sept 28 vs 29). They have the exact same API price on OpenRouter: 10 per million output.
So, same sticker price. But when I ran both through the same real Agent tasks, Sol burned 10.9. That's a 3.5x difference. Same price, wildly different bill.
Here's what happened on the first two tasks.
Task 1: Build a 3D escape room
The job: a first-person Three.js escape room — 3 rooms, 6 interactive objects, all chained together (A unlocks B, B + C opens D, and so on). Plus: every click needs feedback, a "current objective" hint, beatable in 5 minutes, and the model has to play through its own game and report back.
Sol finished in 33:37. Its game, AFTERLIGHT, looks great — retro duty-room vibes, green walls, checkered floor, inventory, sound toggle, objective hint. All six objects wired into a proper dependency chain. But the reviewer said it felt like a checklist diary — every step ticked off — and the keycard looked like it was floating in mid-air, physically impossible.
Sonnet took 74:32, more than twice as long. Its version is a gray prison cell with less decoration, but the puzzle chain is complete and the key and keyhole actually feel real. It sets up real suspense.
So on the game: Sol is faster, Sonnet has better feel.
Task 2: Research a policy change
The job: investigate "AI renamed to SI" — US Executive Order 14434, the Kyoto Vision, first-hand sources — and write a report for Chinese tech workers. Five deliverables, every time-sensitive fact must have a source, and political wording must not be dressed up as a technical breakthrough.
Sonnet knocked it out in 5 minutes 11 seconds. 172 lines, ~23,000 bytes, 32 tool calls. Clean structure, and — the sharp part — it caught the discrepancy between how the White House and Chinese sources described "whether the rename was agreed to." That's a genuinely valuable catch.
Sol took 54 minutes 50 seconds — ten times longer. But it delivered a 14-file evidence package: the report, source and verification logs, 5 original documents (~600KB, hash-checked), and 15 verification judgments across 3 independent passes. This thing is built to be audited.
So on research: Sonnet draws the map fast, Sol builds the evidence trail slowly.
The numbers
| Task | GPT 6.1 Sol | Claude Sonnet 5.5 |
|---|---|---|
| 3D escape room | 33:37 | 74:32 |
| Policy research | 54:50 | 5:11 |
| GPT 6.1 Sol | Claude Sonnet 5.5 | |
|---|---|---|
| Total bill (3 tasks) | $38.5 | $10.9 |
Why the 3.5x gap
It's not the model price — that's identical. It's the path each model chooses:
- Sol keeps verifying. Multi-pass fact-checks, hash-checking the originals, re-running the browser, archiving everything. Every extra tool call and context window shows up on the bill.
- Sonnet takes the shortest path to a usable draft, then stops.
That's what I'd call zigzag intelligence: Sol is a deep, obsessive workaholic who double-checks everything; Sonnet is a fast, efficient daily operator.
The takeaway
The model only sets the ceiling. The toolchain, the verification strategy, and the stopping rule are what actually decide what you get — and what you pay.
Same sticker price, 3.5x the bill. If you're paying for Agent work, the number that matters isn't the per-token price. It's how much the model decides to do.
Worth it? Depends what you need. Fast first draft → Sonnet. Auditable, verified deliverable → Sol. Just know they cost very different money.