The Flash multimodal tier arrives — Gemini 3.5 Flash and Seedance 2.0 redefine what 'cheap' delivers
Gemini 3.5 Flash hits 76.2% Terminal-Bench at Flash pricing. Seedance 2.0 takes the #1 spot on the Artificial Analysis video leaderboard. Two different labs, two different modalities, same architectural move: the cheap tier now ships frontier capability.
The two data points
Gemini 3.5 Flash: 76.2% Terminal-Bench 2.1, 1656 Elo GDPval-AA, 83.6% MCP Atlas — at Flash pricing.
Seedance 2.0: #1 on Artificial Analysis Video Arena leaderboard, Elo 1351 image-to-video, ahead of Veo, Kling, and Sora 2.
Both releases hit similar architectural patterns: the cheaper, faster product tier now delivers capability that was previously reserved for the flagship tier. Multimodal isn't an exception to the Flash-tier rotation — it's a more dramatic example of it.
Why video is the harder case
Video generation is the modality where compute economics dominate user experience. A 60-second 4K clip needs orders of magnitude more compute than a 1000-token text generation. If Flash-tier video models hit frontier quality at Flash pricing, the unit economics of the entire creative-production workflow shift.
Seedance 2.0 taking the leaderboard with Elo 1351 image-to-video isn't a marginal benchmark improvement. It's a re-rank of what 'best' means in the video tier. The standard three-majors framing (Veo 3.1, Sora 2, Kling 3.0) now has a fourth that leads on prompt-following.
What this means for the bifurcation
The unified-vs-pipeline bifurcation we wrote about earlier today gets sharper. Consumer-tier converges on unified multimodal models (Gemini Omni, GPT-Omni); production-tier converges on pipeline orchestration that routes between best-in-class single-modality specialists (Seedance for image-to-video, Veo for cinematic landscapes, Kling for multi-shot continuity).
The Flash tier is the new frontier across modalities. The Pro tier is increasingly a defensive moat against a tail of premium workloads.
The procurement implication
Production-creative buyers running 2026 H2 procurement should explicitly include Seedance 2.0 in the routing layer. The leaderboard data is unambiguous; the licensing terms are competitive; the latency-and-cost profile is operational. Buyers that haven't tested Seedance against their workflows are running on outdated capability assumptions.
For text-and-reasoning buyers, Gemini 3.5 Flash should be the default routing target for 80–90% of traffic. Pro tier (Gemini 3.5 Pro, Claude family, GPT-5.5 family) gets reserved for the residual workloads that the Flash benchmarks confirm Flash can't handle yet.
CNBC — Gemini 3.5 Flash → · AIMLAPI — Seedance 2.0 → · Pixflow — best AI video generator 2026 →