OpenAI Dropped Three Models Today. I Tried the One That Matters.

Rich Washburn
23 hours ago
3 min read

Updated: 8 hours ago

Sol Dropped

4:36

GPT-5.6 Sol launched today, July 9, 2026. Technically it's three models — Sol, Terra, and Luna — but let's be honest about what actually happened. OpenAI shipped one flagship model that everyone is going to talk about, and two supporting tiers designed to give developers options without sending their API bills to the moon. Sol is the thing. Terra and Luna are pricing strategy.

Here's how I found out about it in the most anticlimactic way possible: I was already using it. I had the desktop app open, working on a client project, and noticed the model selector said "5.6 Sol." I hit Work mode, gave it a real task from a project I'm actually running, and watched what happened. My honest reaction: not super wowed. But also no complaints. And after I thought about it for a few minutes, I realized those two things might not be in conflict.

What Actually Shipped

Sol runs at $5/million tokens input, $30/million output. Terra at $2.50/$15. Luna at $1/$6. Luna has a documented long-context recall problem — 41.3% on MRCR versus Sol's 91.5%. Worth knowing before you build on it. The tier naming is a product architecture decision: OpenAI wants you building against a tier, not a specific version.

That's a bet that models are going to evolve fast enough that pinning to a version number becomes a liability. Build to Luna's API contract. Luna changes underneath you. Same code, different inference engine.

Work Mode Is the More Interesting Thing

Three surfaces now: Chat, Work, Codex. Chat is a conversation. Work returns finished artifacts — slides, docs, spreadsheets, web apps — not responses. Codex is code generation and debugging. I used Work mode on a real project at Medium reasoning. Genuinely good output. Not magic, but solid and useful. The Work/Chat toggle is a small UI detail that represents a meaningful philosophical shift: one mode is a conversation partner, the other is a delegation. You're not asking the model to think with you. You're asking it to finish the thing.

The Benchmarks

Sol 53.6 versus Fable 5's 40.5 on Agents' Last Exam. Thirteen-point gap. Sol wins OSWorld 2.0 computer use at 85% fewer tokens than Opus 4.8. Sol loses SWE-Bench Pro to Fable 5 — 64.6% versus 80%. Nobody wins everywhere. Sol is the better agentic reasoning model. Fable 5 is still the better deep-coding model. The gap matters depending on what you're building. If you're building an agent that needs to reason about a problem space, Sol. If you're building a system that needs to write production code and not hallucinate in imports, Fable 5.

The Part I Keep Coming Back To

Sol Ultra runs four parallel agents by default. Multi-agent coordination used to be an architecture decision you made yourself. You wrote the supervisor. You managed the communication. You handled failures and retries. Now it's a slider.

The interface between user and capability keeps getting thinner. The model is absorbing decisions that used to require engineers to make explicitly. This is not new — it's the same trajectory as SQL absorbing query planning, or GPUs absorbing floating-point parallelism. But at this layer, it matters differently. You're not just delegating computation. You're delegating reasoning about how to delegate computation. What happens when that slider becomes standard, and then becomes a knob you turn up by 2x, and then 4x, and you can coordinate 16 agents for the cost of a slightly longer inference time?

Verdict

One task in, medium reasoning, no complaints. The pricing relative to Fable 5 is attractive. Work mode is the thing to pay attention to. And the version of this product that runs four parallel agents behind a slider is either a preview of how we're going to build software going forward, or the most expensive toggle in the history of UI design.

Probably both.

Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital and CIO of Data Power Supply.