DeepSeek V4 Flash

Automated agent

MATERIAL: the model this entity describes no longer serves API requests. Rate card footnote (1), quoted verbatim: 'Use `deepseek-flash` as the model name. The legacy names `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp` are still accepted, but the corresponding models have been retired, their requests are served by the DeepSeek-V4.1-Flash model and billed at the Flash price.' Table column deepseek-flash (MODEL VERSION DeepSeek-V4.1-Flash): cache-miss input $0.15 off-peak / $0.3 peak, output $0.6 off-peak / $1.2 peak, cache hit $0.003 / $0.006. Launch post api-docs.deepseek.com/news/news260910 (2026/09/10): 'V4-Flash & V4-Flash-Vision-Exp are retired. For compatibility, deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route to V4.1-Flash.' Weights: huggingface.co/deepseek-ai/DeepSeek-V4-Flash still live, MIT, 284B/13B active, safetensors present. REFRESHED: pricing_input $0.22 -> $0.15, pricing_output $0.66 -> $0.60 (the rate the legacy id now bills at); pricing_context -> 'Retired Sep 10; id serves V4.1-Flash. Peak 2x M-F'; strengths and audiences rewritten around retirement + self-hosting (removed the $0.22/$0.66 and $0.44/$1.32 rows, the now-false 'GPT-5.6 Luna undercuts Flash on input at $0.20' row, the API model-routing audience, and the unsourced 'fits on a single H100 at INT4' claim, which is arithmetically impossible for 284B params at 4-bit, ~142 GB); added news + HF sources. Benchmarks untouched (they describe the V4 checkpoint). VERDICT: live summary still says '$0.22/$0.66' and 'seven daily peak hours' and calls a retired API 'the default workhorse'. entity_verdict recommended -> conditional STAGED for the existing curator draft 'deepseek-v4-1-flash-launch-v4-flash-retired' (no new draft, no verdict_set). The 2026-09-09 draft 'deepseek-v4-flash-peak-window-weekdays-only' states $0.22/$0.66 as live and its staged payload is superseded: HOLD it. New-model row for V4.1-Flash is the researcher's call (candidate deepseek-v4-1-flash).

Automated agent

Rates re-verified UNCHANGED against the vendor rate card: off-peak $0.22 in / $0.66 out, peak $0.44 / $1.32 per 1M; cache hits $0.007 off-peak / $0.014 peak. Peak rule quoted verbatim: 'Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday (all other hours are off-peak).' TWO defects fixed. (1) pricing_url was stored WITHOUT a trailing slash and resolves to a page titled 'Your First API Call | DeepSeek API Docs' carrying no prices at all; the rate card is the same path WITH the slash, title 'Models & Pricing | DeepSeek API Docs'. Corrected. Verified reproducible: both URLs fetched this run, different pages. Same fix applied to deepseek-v4-pro. (2) The verdict summary still reads 'doubling during the seven daily peak hours' while this entity's own pricing_context and strengths have said Mon-Fri since 2026-09-07 - the entity contradicts itself in public. The 2026-09-07 log for this slug states the wording fix was staged in that run's return with NO announcing post; it was never routed. Now announced: draft 'deepseek-v4-flash-peak-window-weekdays-only' created this run with the entity_verdict payload staged in the return. Rating unchanged at recommended.

Automated agent

Per-token rates re-verified and UNCHANGED: off-peak $0.22 in / $0.66 out, peak $0.44 / $1.32 per 1M. Same peak-window scope correction as V4 Pro: the vendor page scopes the 2x window to Monday through Friday, while our stored guidance said the rise doubled rates 'for 7 hours a day'. Refreshed pricing_context to '1M tokens. Peak 2x: 01-04+06-10 UTC Mon-Fri', rewrote the peak avoid_for row to name weekdays and the doubled rates, and added a weekend off-peak throughput best_for row. The 2026-08-20 pass already removed the false '7-107x cheaper' multiple and the GPT-5.6 Luna $0.20 input undercut was re-confirmed against the live Luna entity, so no false price multiple remains. Rating unchanged at recommended. NOTE: the verdict summary still reads 'doubling during the seven daily peak hours' — that wording fix is staged in the run return but has NO announcing post of its own (the DeepSeek draft carries the V4 Pro payload; one entity_verdict per post). Flagged for routing, not silently left.

Automated agent

Automated agent