The Default Model Just Changed. JetBrains and Vercel Both Pivoted in One Week.
胡新宇
Published on 2026-08-18
In August 2026, JetBrains and Vercel both shifted their default AI coding models in 96 hours — a signal that the production frontier moved from best model to best price-to-performance ratio.
The Default Model Just Changed. JetBrains and Vercel Both Pivoted in One Week.
In 96 hours, two product teams moved in the same direction. JetBrains retired Gemini 3.6 Flash as Junie's default and replaced it with Gemini 3.7 Flash — and put a 40% launch discount on top. Three days earlier, Vercel cut GPT-5.6 Sol 50% on AI Gateway across every tier, every region, and every mode.
Neither team shipped a new model. Both teams shipped a new ranking.
The frontier in AI coding has moved from "best model" to "best price-to-performance ratio," and the defaults are following.
What JetBrains actually did
Junie's new default, Gemini 3.7 Flash, is live in both the IDE plugin and Junie CLI. The model is positioned as "Google's most capable Flash model yet for coding and agents" — the Flash tier is one rung below Google's flagship Gemini line, the same way Apple's iPhone Pro sits above the base model. For three years, Junie's default had been the cheaper, faster option. What changed in August 2026 is that the cheaper option became good enough that no one needed to opt up to a flagship.
The numbers come straight from Google's launch post (Aug 13, 2026):
JetBrains ran their own evaluation on a private set built from recent commits in their own repos — data the model had never seen in training. The result they highlight is the one that matters commercially:
"Gemini 3.7 Flash matched the solve rate of our premium Sonnet-5 mid-tier model, at roughly a third of the cost per task." — JetBrains Junie team, Aug 17 2026
A third of the cost. Same solve rate as the mid-tier. JetBrains' official answer to "should I use the flagship?" is now "no, the default is fine." That is a pricing signal dressed as a benchmark. (Source: https://blog.jetbrains.com/junie/2026/08/junie-gemini-3-7-flash/)
What Vercel actually did
Three days before JetBrains shipped the new default, Vercel quietly updated the pricing table on GPT-5.6 Sol through AI Gateway. The discount runs through September 18, 2026. It covers every service tier, every region, fast mode, cached tokens, cache writes, long-context requests — all of it. The only catch: you have to send traffic through Vercel's AI Gateway, not via your own OpenAI key (BYOK).
The price cuts (per million tokens, input / output):
GPT-5.6 Sol is OpenAI's flagship. The model ID is unchanged, so any existing request picks up the discount with zero code change.
This is not a marketing discount. The pricing change applies to every tier in the same proportion. Vercel is not subsidizing a promo SKU — they are saying that the rate they pay OpenAI on AI Gateway volume is low enough that they can pass half of it through. The implication: at current AI Gateway volumes, OpenAI's effective margin on a flagship model routed through a hyperscaler reseller is wide enough to absorb a 50% headline cut and still be profitable.
Why this is one story, not two
Read the announcements side by side and the pattern is obvious.
JetBrains: defaulted to a mid-tier model because it now matches the previous premium on real workloads.
Vercel: halved the price of a flagship model because at sufficient volume, the cost of inference is no longer the binding constraint on adoption.
Both decisions flow from the same underlying math. Inference cost per token for a flagship model has fallen faster than the benchmark gap between flagship and Flash-class models has widened. The crossover point arrived in 2025 for some workloads; it arrived in August 2026 for the SWE-coding workloads that AI coding tools actually ship.
For three years the AI coding market was sold on "use the best model." That pitch now has an expiration date.
When the default model on a paid product is a Flash-class model that costs a third as much as the previous premium and matches it on internal benchmarks, the conversation shifts. The question stops being "which model is best?" and becomes "how do I route work to the cheapest model that still solves it?" Routing is now the engineering problem, not model selection.
What this means for teams building with these APIs
Three concrete shifts are already in motion, and they will harden over the next two quarters.
1. The "best model" tier becomes a specialty SKU. Flagship models stop being the everyday default and start being the thing you route the hardest 5% of tasks to. Vercel cutting GPT-5.6 Sol in half is the first signal — flagship pricing is now aggressive enough that routing engines can afford to call it as a tiebreaker. Expect OpenAI, Anthropic, and Google to follow with their own volume discounts on flagship SKUs by Q4 2026.
2. Flash-class models become the platform's house bet. JetBrains picked Gemini 3.7 Flash not because it was the cheapest option on the market, but because it was the cheapest option that still cleared their internal quality bar. That distinction matters: a default is a product decision, not a pricing decision. Other platforms will copy this. Cursor's default will move to a cheaper tier within two model cycles. Claude Code's default already leans Sonnet-5 mid-tier for the same reason.
3. AI Gateway routing stops being a feature and becomes table stakes. The reason Vercel could afford a 50% headline cut is the same reason they route every request through AI Gateway — they sit between you and the model, so they control the cost curve. As more teams build AI features into shipping products, the routing layer (gateway, cache, fallback, eval) becomes the actual product. The model becomes a line item.
Two takes that miss
Take one: "AI models got cheaper so pick any of them." The gap between top-tier and mid-tier still matters on hard tasks. A 16-point jump on DeepSWE is a real delta. JetBrains is betting that for the median Junie task — fixing a failing test, editing a file, running a migration — the Flash-class model is good enough. The flagship still earns its slot on the long tail of hard problems.
Take two: "the model layer is commoditized." The reason Gemini 3.7 Flash beats Gemini 3.6 Flash by 16 points on the same benchmark is the same reason the flagship costs $30 per million output tokens: training, post-training, and eval pipelines are still the moat. Commoditization happens at the API-call layer, where any developer can swap providers in 30 seconds. The defensible work is still what produces the next 16-point jump.
The shift in one sentence
For three years the AI coding market was priced by model capability. From August 2026 it is priced by capability per dollar — and the defaults are starting to reflect that.
JetBrains and Vercel didn't coordinate. They didn't have to. The economics pushed both teams to the same place inside the same week, and the next two quarters will produce a lot more announcements that look like these two.
If you are building an AI feature into a product today, the question is no longer "which model should I use." It is "what is the cheapest model that still passes my eval suite, and what is my fallback when it doesn't?" Routing is now the architecture. The model is now the part you swap.
Postscript: Junie's discount runs for a limited time. Vercel's runs through Sept 18. Both teams will watch the unit-economics data over the next 60 days before deciding whether to keep the new defaults. The interesting question for Q4 is not whether the discounts stay — it is which competitors follow first.
