DeepSeek V4.1 Flash vs V4 Pro vs Gemini 3.8 Flash Compared
The short answer
| DeepSeek V4.1 Flash | DeepSeek V4 Pro | Gemini 3.8 Flash | |
|---|---|---|---|
| Released | September 10, 2026 | GA August 13, 2026 | September 2, 2026 |
| Input / output (per MTok) | $0.15 / $0.60 off-peak · $0.30 / $1.20 peak | $0.66 / $1.98 off-peak · $1.32 / $3.96 peak | $0.75 / $3.75 intro → $1.50 / $7.50 from Jan 1, 2027 |
| Cache-hit input | $0.003 off-peak · $0.006 peak | $0.022 off-peak · $0.044 peak | ~$0.075 (10% of input) |
| Context | 1,048,576 tokens | 1M tokens, 384K max output | 1M tokens, 64K max output |
| Modalities | Text + native vision | Text (+ vision variant) | Text, image, audio, video in |
| Architecture | 552B MoE, CED encoder-decoder, 8B active prefill / 16B decode | Dense-MoE V4 backbone | Undisclosed |
| Open weights | Yes, MIT license | Yes | No |
| Best for | Cacheable agent loops, long-context rereads, self-hosting at scale | Being retired — routes to V4.1 Flash from Sep 14 | Multimodal input, Google Cloud residency, Batch/Flex discounts |
Default to V4.1 Flash for cost-sensitive agent workloads; keep Gemini 3.8 Flash when you need multimodal input or Google’s compliance surface. V4 Pro is no longer a real choice: DeepSeek has announced that deepseek-v4-pro calls will be routed to V4.1 Flash starting September 14, 2026 until V4.1 Pro arrives.
What DeepSeek shipped on September 10, 2026
DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026 with a 552-billion-parameter mixture-of-experts backbone, native vision, a 1-million-token context window and MIT-licensed open weights on Hugging Face. The headline is not the model size but the cache economics: on a cache hit, off-peak input costs $0.003 per million tokens. DeepSeek itself says cache-hit charges “can account for a significant portion of agent costs,” which is exactly why the number matters.
The pricing grid, per DeepSeek’s API docs:
| Rate (per MTok) | Off-peak | Peak (Mon–Fri 01:00–04:00 & 06:00–10:00 UTC) |
|---|---|---|
| Input, cache miss | $0.15 | $0.30 |
| Input, cache hit | $0.003 | $0.006 |
| Output | $0.60 | $1.20 |
For a concrete comparison: an agent that keeps a 500,000-token reusable prefix and hits the cache across 100 requests (50 million cached input tokens) pays about $0.15 on V4.1 Flash off-peak, versus roughly $15 on Kimi K3 ($0.30 cache-hit), $20 on GPT-5.6 Sol ($0.40) and $25 on Claude Opus 5 ($0.50) at their published cache-read rates. Workload timing is now a cost lever: schedulable batch jobs should run off-peak.
Benchmarks: frontier-adjacent, not frontier-dominant
DeepSeek’s technical report (all rows at maximum reasoning effort, reasoning_effort=100) puts V4.1 Flash at:
- DeepSWE v1.1: 74.2 — versus 74.0 for Claude Opus 5 and 73.0 for GPT-5.6 Sol
- CyberGym: 88.1
- AutomationBench: 54.8
- Terminal-Bench 2.1: 90.6 (at effort 100; 82.4 at effort 25)
The gaps go the other way on harder terminal work: Opus 5 leads 43.3 to 30.0 on Terminal-Bench 3.0 and 51.8 to 31.2 on Terminal-Bench 4.0, and GPT-5.6 Sol leads on GPQA Diamond and SEC-Bench Pro in DeepSeek’s own table. Early third-party evidence points the same way: OpenDesign reported V4.1 Flash reaching 98% of GPT-6 Astra’s quality score at 1.4% of its cost on a narrow set of everyday design requests — a price-performance result, not an intelligence lead.
Reasoning effort is a hidden cost variable. Raising effort from 25 to 100 lifts DeepSWE from 66.0% to 74.2% but consumes about 2.5x more output tokens; DeepSeek says effort 60–80 recovers most of max-effort accuracy at under half the token budget. The public API exposes low, high and max presets (effort 50, 75 and 100). The leaderboard numbers were produced at max; your best cost-per-completed-task is probably not.
The architecture change that enables the price
V4.1 Flash is the first model built on DeepSeek’s Causal Encoder-Decoder (CED) architecture: 40 transformer layers split into a 20-layer causal encoder and a 20-layer decoder. It activates 8B parameters during prefill and 16B during decode — so “8B-active” is only half true. Combined with Compressed Sparse Attention 2, hierarchical sparse indexing and FP4 KV caching, the global KV cache falls to roughly 890 bytes per token (about a quarter of V4 Flash’s) and persistent cache storage to about one-eighth. DeepSeek says CED roughly halves prefill compute on long sequences, and single-token decode FLOPs rise only ~25% as context grows 256x from 4K to 1M tokens.
Two caveats from DeepSeek’s own report:
- “Flash” no longer means small. V4 Flash had a 284B backbone with 13B active; V4.1 Flash nearly doubles the backbone to 552B (plus 196B of sparsely accessed Engram memory). Self-hosting is materially harder — DeepSeek tells large-scale deployers to get in touch if they have “on the order of 2,000 GPUs plus a storage cluster.”
- Uncharacterised robustness boundaries. DeepSeek says CSA2 sparse-selection errors and the approximate state reconstruction in SWA Bounded Replay “could cause capability degradation in untested edge cases,” particularly sparse retrieval over very long contexts and cache-resumption boundaries. It has not observed systematic degradation but plans more stress testing.
The cache itself is tiered: global KV persists for a guaranteed 72 hours, while short-lived sliding-window KV lives in a DRAM pool with minute-scale lifetime; if only the global cache survives, the model replays just the most recent attention window rather than the full history.
How Gemini 3.8 Flash compares
Google released Gemini 3.8 Flash on September 2, 2026 at an introductory $0.75 / $3.75 per MTok — a 50% intro rate that reverts to $1.50 / $7.50 on January 1, 2027. It offers 1M context, 64K max output, a March 2026 knowledge cutoff, and accepts text, image, audio and video. Batch and Flex are 50% off; Priority is ~1.8x. Its cyber-specialised sibling, Gemini 3.8 Flash Cyber, is gated behind the Fairwind Program and not on the standard API.
Where Gemini 3.8 Flash wins: native audio/video input, Google Cloud data-residency and enterprise terms, no peak/off-peak surprise, and a larger 64K output ceiling per call than most workloads need. Where V4.1 Flash wins: 5x cheaper at list off-peak (2.5x even at DeepSeek’s peak rate versus Google’s intro rate, ~10x versus Google’s 2027 rate), ~25x cheaper cache reads, 1M context with open weights, and frontier-competitive agentic coding scores.
Migration and trust caveats
- Identifier routing. DeepSeek retired V4-Flash and V4-Flash-Vision-Exp and temporarily points their IDs to V4.1 Flash; from September 14, 2026,
deepseek-v4-proalso routes to V4.1 Flash until V4.1 Pro ships. Developers on Hacker News objected that swapping the model behind a production identifier invalidates regression testing. Pin evaluations, don’t assume. - Data handling. Anthropic’s September 10, 2026 threat report alleges DeepSeek silently relayed some customer requests (including from Claude Code and OpenCode harnesses) to Claude Opus and harvested reasoning traces — over 12.1 million exchanges in 14 days of July 2026. Whatever the outcome, teams with sensitive data should read DeepSeek’s terms and consider self-hosting the MIT weights rather than the hosted API.
- Commercial context. Reuters reported on September 9, 2026 that DeepSeek has hired CITIC Securities for a Shanghai STAR Market IPO, with a funding round valuing it at about 500 billion yuan (~$75 billion). Cheap inference at scale is the pitch to public markets.
Decision guide
- Coding agent that rereads a repo, RAG over stable docs, batch classification: V4.1 Flash, off-peak, effort 60–80.
- Voice, video or image-heavy pipelines: Gemini 3.8 Flash (V4.1 Flash has vision but no audio/video).
- Regulated data, EU/US residency requirements: Gemini 3.8 Flash via Vertex, or V4.1 Flash self-hosted.
- Hardest terminal/agentic tasks where a few points matter: Claude Opus 5 or GPT-5.6 Sol still lead on Terminal-Bench 3.0/4.0.
- Currently on V4 Pro: you are being migrated on September 14 anyway — run your evals against V4.1 Flash now.