DeepSeek vs Claude: Open Weights Against the Flagship

This is the widest price gap in the arena roster. DeepSeek V4 Flash is the fast tier of DeepSeek’s V4 family: open weights, an FP8 checkpoint anyone can download, and list prices around $0.14 per million input tokens and $0.28 per million output. Claude Fable 5 is Anthropic’s Mythos-class flagship at $10 in and $50 out — the most expensive Claude ever sold. That is roughly a 70x gap on input and nearly 180x on output, which makes this the purest form of the question every open-weight release raises: how much flagship do you give up for two orders of magnitude in price?

The sticker prices

Read the output column twice. At list prices you could run DeepSeek V4 Flash roughly 179 times before matching the cost of one Fable 5 completion. No capability gap is 179x. So the honest question is not whether Fable 5 is better — nobody serious argues a flash-tier open model out-reasons Anthropic’s flagship across the board. It is whether Fable 5 is better on your task, on the runs where being right the first time is the entire job.

What open weights actually buy

The per-token price is only half the pitch. The other half is what holding the checkpoint means in practice:

• Self-hosting: the FP8 release roughly halves memory versus 16-bit weights, pulling serving down from cluster territory toward something a serious team can rack themselves.
• Provider competition: anyone can serve an open checkpoint, so hosts compete on price and speed for the same model — which is why the arena runs it through a router rather than a single vendor gate.
• No deprecation risk: an API flagship can be repriced or retired on the vendor’s schedule; a checkpoint you hold runs identically for as long as you keep it.

Cost-per-task is where flagships fight back

Per-token rates only become invoices through token counts — the argument we made at length in what tokens actually buy, and the reason the arena prints generation tokens and cost next to every output. Fable 5 runs at five thinking-effort levels and spends heavily where problems are hard, so the same prompt can cost cents or several dollars depending on the dial. A cheap model that pads its output or needs three retries erodes its own discount. But be realistic about the ceiling: 179x survives any token multiplier a working model can produce. The one thing it cannot survive is a build that never works — a broken landing page at a twentieth of a cent is worth exactly nothing, and the retry loop costs you an afternoon the invoice never shows.

Price-per-token has an easy winner. Price-per-working-build is the actual contest — and that one you can watch live.

Same prompt, one shot, judged blind

Every challenge in the arena hands both models the identical one-shot prompt — no retries, no cherry-picking — and the raw outputs run live in your browser. Watch DeepSeek V4 vs Claude Fable 5 build the same landing page, Fable at max effort, judge them blind, and vote before the names reveal. The leaderboard keeps the community’s running tally, with cost-per-task and generation times printed next to each model.

My take

The bar for an open-weight model at this price is not parity — it is plausibility. If V4 Flash ships a working build on a decent fraction of the one-shot prompts Fable 5 also lands, the economics end the argument for everyday work, and Fable settles into what a flagship should be: the model you escalate to when the task defeats everything cheaper. That fraction is an empirical question with a live answer. Open the arena, pit them on your kind of task, and vote — the roster’s cheapest model against its most expensive is exactly the comparison blind voting was built for.