DeepSeek V4.1-Flash: When "Too Cheap to Meter" Stops Being a Slogan

DeepSeek just shipped a frontier-class model that costs roughly one percent of GPT6 Astra while landing at about 98 percent of its performance, which turns the "intelligence too cheap to meter" line from an AI-CEO talking point into an actual price sheet. The video walks through the September 10, 2026 launch of DeepSeek-V4.1-Flash and argues that when a model this good costs less than a cent per million input tokens, the economics behind expensive proprietary models start to look shaky.

The Pricing Is the Story

The video pulls up DeepSeek's own pricing table, and the numbers are the whole point. For 1M input tokens on a cache hit, off-peak pricing is $0.003 and peak-hour is $0.006. Cache misses run $0.15 off-peak and $0.30 at peak. Output tokens are $0.60 off-peak and $1.20 at peak.

That cache-hit input price is the eye-catching figure. Less than a cent per million tokens, in both off-peak and peak windows, is what the creator calls "diabolical." The design choice here is deliberate: DeepSeek optimized hard around KV cache reuse, so workloads that hit the cache pay almost nothing. The new pricing took effect at 4:00 UTC on September 10, 2026, exactly as the on-screen post states.

98 Percent of the Frontier at 1 Percent of the Cost

The benchmark table in the video compares V4.1-Flash against DeepSeek V4-Pro, GLM 5.3, Kimi K3, GPT 5.6-Sol, and Claude Opus 5. V4.1-Flash posts a Codeforces rating of 3471, the highest in the table, and leads on Terminal-Bench 2.1 (90.6), DeepSWE v1.1 (74.2), and CyberGym (88.1). It sits at 90.9 on GPQA Diamond against GPT's 94.1 and Opus 5's 93.4.

The pattern is consistent: a small, cheap model trading within a few points of the flagship proprietary systems on most benchmarks, and beating them on several coding and agentic tasks. That is the "98 percent performance" claim in concrete terms. The gaps that remain, like HLE where V4.1-Flash scores 36.8 against Opus 5's 56.3, show it is not a clean sweep, but the price-to-capability ratio is the argument.

Efficiency by Design: The KV Cache Curve

One frame shows a chart titled "Global KV Cache Per Token (Bytes)" tracing DeepSeek's own models over time. DeepSeek-V1 used 389,120 bytes per token in late 2023. V3.2 cut that to 48,068 (8.1x smaller). V4-Flash dropped to 3,514 (13.7x smaller), and V4.1-Flash reaches 890 bytes per token (another 3.9x smaller) as of September 2026.

This is the mechanism behind the pricing. Smaller KV cache per token means cheaper memory during inference, higher throughput, and lower serving cost. The cheap API price is not a loss-leader promotion, it is a downstream effect of years of architecture work on cache efficiency. The model is also natively multimodal, with visual understanding built into the base architecture rather than bolted on.

Open Weights and the Competitive Pressure

The creator's framing, "you can just straight up go ahead and cancel these IPOs," is hyperbole, but it points at a real dynamic. Open-weight models from Chinese labs keep matching proprietary frontier performance and then undercutting it on price by an order of magnitude or more. When a model that runs near the top of the benchmark tables also ships open and prices input at fractions of a cent, the moat for closed models narrows to whatever is left after capability and cost are commoditized.

The joke at the end, "you could run ads," gets at the same thing. At these token prices, the marginal cost of serving a query approaches zero, which changes what business models around AI inference can look like.

Key Takeaways

DeepSeek-V4.1-Flash launched September 10, 2026, with new pricing effective 4:00 UTC that day.
Cache-hit input pricing is under a cent per million tokens ($0.003 off-peak, $0.006 peak); output is $0.60 to $1.20 per million.
It trades within a few points of GPT 5.6-Sol and Claude Opus 5 on most benchmarks and leads on several coding and agentic tests.
The low price is enabled by a huge reduction in KV cache per token, down to 890 bytes from 389,120 in V1.
The model is natively multimodal and open, and it now serves legacy deepseek-v4-flash and (from September 14) deepseek-v4-pro requests.