DeepSeek V4 Flash Hits 8 Trillion Daily Tokens, Tops Trending Charts at 1/105th the Cost of Claude Fable 5

8 trillion tokens. One day. One company's consumption.

OpenCode posted on X: "DeepSeek Flash did 8T tokens on August 1st. 5T of free usage + 3T on OpenCode Go." 798,000 views, and with it, #DeepSeek consumed 8 trillion in a day# trended on Weibo — and note, this is just OpenCode's consumption alone.

OpenCode is an AI coding tool from Anomaly, priced at $10/month. Users pay $10 and get unlimited access to DeepSeek Flash for coding. As one netizen put it: "No matter how hard you try, you can't use it all up." The 5T came from free tier usage, while 3T came from paying subscribers on the Go plan. A Weibo user commented: "Just bought it yesterday — you get $5 for referrals too."

Everlier on X said: "Two years ago this would sound like a sci fi, crazy." Ashish was more direct: "Opencode is in hurry to close Anthropic asap."

What the 8T token figure actually means is worth unpacking.

Let's rewind the timeline two days. On July 31, DeepSeek released V4 Flash 0731. Artificial Analysis's benchmark data: Intelligence Index score of 50, ranking 3rd out of 101 models. Ahead of it: Claude Opus 5 (61 points, $2.34/task), Claude Fable 5 (60 points), GPT-5.6 Sol (59 points, $1.86/task), Kimi K3 (57 points, $0.86/task), Grok 4.5 (54 points, $0.69/task).

V4 Flash pricing: $0.03/task.

Opus 5 scores 11 points higher but costs 78x more. GPT-5.6 Sol scores 9 points higher but costs 62x more. The only model that can compete on price is DeepSeek's own V4 Pro ($0.05/task), but Flash matches it with a 50 Intelligence Index score — beating the Pro outright.

Nothing smarter than DeepSeek V4 Flash is cheaper. Nothing cheaper is — well, there's nothing cheaper.

The architecture hasn't changed: MoE with 284B total parameters, 13B activated parameters. DeepSeek purely relied on re-post-training to pull the score up. Cache hit pricing is $0.003/million tokens, down 98%. MIT open-source license, weights available on Hugging Face.

That's the backdrop. Within 48 hours of V4 Flash 0731's release, OpenCode alone hit 8T tokens in daily consumption. What does 8T tokens actually mean? V4 Flash inference pricing is $0.12/million input tokens and $0.48/million output tokens.

Even using a blended rate of roughly $0.20/million tokens, 8T tokens corresponds to a theoretical API value of $1.6 million. Of course, OpenCode certainly secured wholesale pricing, but the scale speaks for itself — a single AI coding tool burning through more tokens in one day than most mid-sized AI companies consume in a full year.

And OpenCode only charges users $10/month.

The math behind this is simple. Cursor's Anthropic model costs have climbed, and Copilot's GPT-5.6 isn't cheap either. OpenCode chose DeepSeek Flash as its underlying model, using ultra-low inference costs to offer budget pricing and completely eliminate token anxiety. Users don't need to watch remaining credits, don't need to agonize over switching models — writing code, fixing bugs, refactoring, running tests — all on one model, unlimited.

DeepSeek's core position is equally clear. V4 Flash's inference infrastructure has entered the "running water" stage — not billed by the drop, but by the flow. 8T tokens a day means processing roughly 92 million tokens per second on average. If the AI coding tool market continues to grow, that number will keep climbing. OpenAI and Anthropic each have their API ecosystems, but if in the AI coding scenario the intelligence gap between the cheapest and most expensive models is just 11 points while the price gap is 78x, the market's direction is already clear.

A foreign netizen summed it up: "V4 Flash is my daily driver. I write code all day and it costs me a few cents. No token anxiety."

After 8T tokens, that statement has a more concrete footnote.

Sources: OpenCode on X DeepSeek V4 Flash 0731 — Artificial Analysis HN Discussion