Here's a contradiction worth sitting with. Claire Vo ran five frontier models through her own benchmark and declared GPT-5.6 Sol the winner "by a meaningful margin." In the same breath, she named a different model for nearly every task: Terra for PRDs, Sonnet 5 for agentic voice, Sol for creative coding, Fable close behind on apps. So which is it — did Sol win, or did four different models win?
Both. And that's the whole story of the week.
Last week, the frame was still a bet. The 07/05 GLM-5.2 and Composer rotation was framed as a wager on which horse to back for the quarter. This week, four fresh frontier coding models landed in roughly seven days, and the bet quietly died. Routing isn't a hypothesis anymore. It's the stated practice.
Compress the calendar. Grok 4.5 dropped July 8, trained alongside Cursor, live on all Cursor plans day one. GPT-5.6 Sol landed July 9. Meta's Muse Spark 1.1 arrived late the same night, July 9, when Zuckerberg dusted off an X account that had sat idle for three years. Anthropic's Fable was already the incumbent to beat.
Each seized a crown. Then a rival took it back within a day.
Grok 4.5 claimed the top spot on Harvey's Legal Agent Bench on July 8. Muse Spark 1.1 dethroned it in under 24 hours — scoring a full 20.00 on that same legal benchmark against Grok's 12.92, before, as one writeup put it, "SpaceX AI's new crown had even had time to settle."
That's the tell. When a leaderboard crown has a half-life measured in hours, "winning" stops being something you can build a decision on.
So builders stopped trying. They started stitching.
When a leaderboard crown has a half-life measured in hours, winning stops being something you can build a decision on. So builders stopped trying. They started stitching.
Watch Peter Yang's actual weekend workflow, posted as three lines: "Use Fable to build plan.html with design guidelines. Use Claude Design to make components + screens. Get gpt 5.6 to build." That's not a man choosing a model. That's a man running an assembly line where each station is a different vendor.
Claire Vo's benchmark tells the same story from the evaluation side. Her split was 70% her own taste, 30% LLM-as-a-judge. Her takeaways: Sol her favorite for coding apps, Fable "not that far behind, just not as creative," Terra a surprise on PRD writing, and Sonnet "still the only human agentic voice of the bunch."
Four categories, four different answers. The rational response to that data isn't to pick one. It's to route each task to its per-task winner.
This is now the default posture, and it's genuinely the correct read of the evidence. The problem is what the industry is quietly pretending: that routing across four vendors is free.
It isn't. Routing means you now run four vendor relationships instead of one. Four billing models. Four rate-limit regimes, four safety-eval quirks, four sets of API changes that ship on nobody's schedule but their own.
And it means you re-test on every release. That's the part nobody has a budget line for.
Think about what Claire's benchmark actually is. It's a rig — five models, five task categories, a weighted index, side-by-side prototype comparisons. She built it to answer a question that was true the week she ran it. Grok, Sol, and Muse all landed inside that same window. Any team routing seriously has to rebuild that rig, or something like it, every time a frontier model ships.
In a normal year, that's an occasional chore. In a week where four models drop, it's a standing tax. And the tax scales with the exact thing everyone is celebrating: launch velocity. The faster the frontier moves, the more often your routing table goes stale, and the more often you pay to re-validate it.
There's no exit. You can't opt out by picking one model, because the whole week just proved that picking one leaves per-task performance on the table your competitor is routing to. And you can't set the router and forget it, because Muse took Grok's legal crown in a day — your Tuesday routing decision is a Wednesday liability.
This is the re-evaluation tax. It's permanent, it grows as the release cadence accelerates, and most product and eng teams have neither a line item for it nor a way to stop paying.
Now read the pricing through that lens, because the pricing only makes sense once you accept that nobody's trying to be your one model anymore.
GPT-5.6 Sol shipped at $5 per million input tokens and $30 output — explicitly, in OpenAI's framing, "midrange price, top-tier performance." Sol Ultra scored 91.9% on Terminal-Bench 2.1, standard Sol 88.8%, both ahead of Fable 5's 84.3%. Fable sits at $10 in, $50 out. Sol is the top-tier slot at half the flagship price.
Grok 4.5 came in cheaper still — $2 in, $6 out — and, crucially, coupled to Cursor as the default model on every plan. That's not a price. That's a distribution wedge into an editor developers already live in.
Muse Spark 1.1 undercut everyone: $1.25 in, $4.25 out. Roughly 10x cheaper than Fable overall — 8x on input, nearly 12x on output. It runs a Vals test in 388 seconds against 1000-plus for Fable, Opus, and Sonnet, at about $0.50 per run.
One observer nailed Meta's play: this "isn't so much about delivering groundbreaking new capabilities as it is about making Agent functionality extremely affordable." The 36Kr writeup put it harder — Meta "isn't competing on who can build the smartest model — it's competing on who can survive the pay-as-you-go billing race the longest."
Read those three moves together and the strategy is unmistakable. Sol's midrange slot, Grok's Cursor-coupled discount, Muse's 10x price cut — none of them are pitches to be your only model. They're pitches to be a permanent line in your router.
Sol wants to be the "build" step in workflows like Yang's. Grok wants to be the model that's already there when you open Cursor. Muse wants to own the high-volume, low-margin agentic work where a 10x price gap compounds across millions of tokens. Each is optimizing for a durable slot, not a monogamous commitment.
Which flips where lock-in lives. It used to sit at the model layer — you standardized on one provider, learned its quirks, wired your prompts to its behavior, and switching meant a migration. That's the world last week's bet-framing assumed.
Now the model layer is fluid by design. You're supposed to swap models per task, per release, per benchmark refresh. The switching cost there has been engineered close to zero, which is exactly why crowns change hands in a day.
So the lock-in migrated. It now lives at the routing layer — the accreted knowledge of which model wins which task, the billing plumbing across four vendors, the eval rig you rebuild every launch, the institutional muscle memory of your router's config. That's the thing you can't cheaply rip out. That's the thing that gets stickier every time you pay the re-evaluation tax.
The four-model week didn't crown a winner. It killed the premise that a winner was the point.
The builders already know this. Yang's three-line pipeline and Claire's four-way split are the same signal from two directions: the unit of decision is no longer the model, it's the route. Anthropic, OpenAI, xAI, and Meta all understand it too — that's why they priced and distributed for a slot in your router rather than a spot on your shortlist.
The only party still pretending otherwise is the org chart. There's a budget line for model spend. There's rarely a budget line, an owner, or an exit for the standing cost of deciding — again, next week — which model to spend it on.
That's the bill coming due. Not the tokens. The re-evaluation.