DeepSeek's flagged price increase lands: cache-hit tokens cost up to 12 times more from August 16

[DeepSeek's](#/company/deepseek) V4 Pro 0813 build [left preview earlier August 13](#/article/deepseek-v4-pro-general-availability), moving from four months of testing to full production across the API, app and web. That earlier release carried a warning that API pricing would rise “significantly” without saying by how much. Two things have settled since: DeepSeek's own pricing documentation now shows the actual rate card, and Artificial Analysis has published the first independent benchmark score for the release.

The independent number DeepSeek's own benchmarks didn't include

DeepSeek's own release benchmarks measure agent-specific tasks against the model's prior build, not a general-capability index produced by an outside party. Artificial Analysis's Intelligence Index — the aggregate this newsroom's [Scoreboard](#/scoreboard) tracks — now puts V4 Pro 0813 at 53, up from the 44 last recorded for the model. Independent reporting from the-decoder corroborates both the score and DeepSeek's own reported gains on agentic coding tests: Terminal-Bench 2.1 rising from 72.1 to 87.9 and DeepSWE from 12.8 to 62.7 versus the April preview.

Ten points closes real ground but doesn't close the gap: V4 Pro still sits ten points behind the current independent leader, Claude Opus 5, and seven behind the closest of the 60-plus cluster, Kimi K3. DeepSeek also released Deepseek Harness v0.1, an MIT-licensed agent framework built on a plugin system it calls Cordis, as a developer preview — giving away orchestration tooling in the same week it raises the price of the tokens that tooling consumes.

What the flagged increase actually is

Effective 16:00 UTC on August 16, DeepSeek is replacing V4 Pro's flat rate with peak and off-peak billing tied to Chinese business hours — peak from 01:00–04:00 and 06:00–10:00 UTC, off-peak the rest of the day. Read directly from DeepSeek's API documentation: cache-hit input rises from $0.003625 per million tokens to $0.022 off-peak or $0.044 at peak; cache-miss input rises from $0.435 to $0.66 off-peak or $1.32 at peak; output rises from $0.87 to $1.98 off-peak or $3.96 at peak. Every tier rises, and none rises by the same multiple — which is what makes a single “price increase” headline number misleading without the breakdown below.

The cache-hit tier is the one to watch. It's the cheapest rate DeepSeek offers, built for requests that reuse a previously-seen prompt prefix — exactly the pattern in agentic coding tools that resend a large, mostly-unchanged context window on every turn. Coding agents that keep a large system prompt or a whole repository's context resident across many turns — the exact shape of tools built on V4 Pro's Codex-compatible interface — are also the ones with the highest ratio of cache hits to total tokens, which is what makes this specific tier, and not the sticker price most coverage leads with, the one worth modeling before switching. the-Decoder's analysis frames the change as partially undoing DeepSeek's own May price cut, with cache-hit costs at peak hours now landing above where they sat before that cut.

Still cheap by frontier standards — just less of an outlier

Even at the new peak rate, V4 Pro's list price stays well under the top of the market: $1.32 in / $3.96 out per million tokens at peak, against Claude Opus 5 and GPT-5.6 Sol's published $5 in / $25 out. The gap that made DeepSeek's pricing a story in the first place — a large fraction of frontier capability at a fraction of frontier cost — narrows with this change but doesn't close. What's changed is the shape of the trade: a developer choosing V4 Pro today is paying more to reuse context than they were a week ago, even if the sticker price relative to Anthropic or OpenAI still favors DeepSeek.

This is DeepSeek's second capability-and-pricing move inside three weeks. [V4 Flash's own retrained 0731 build](#/article/deepseek-v4-flash-0731-beats-own-flagship) graduated July 31 at a 10-point independent score gain to 50 — tied with Google's Gemini 3.5 Flash — while keeping its $0.14/$0.28 list price unchanged. DeepSeek's own agentic-benchmark claims that Flash now beats the larger Pro model on several coding tests are self-reported and unverified against an independent source, so they're noted here without being adopted as fact. Read together, the pattern is a company pushing capability gains through its cheaper model while extracting more revenue from the flagship's heaviest users — the opposite of the flat, uniform price cuts DeepSeek built its reputation on in 2025.

Neither the preview-to-GA move nor this price change has a dedicated changelog or announcement post from DeepSeek as of this writing — the model card, API documentation, and pricing pages are the only primary account of what changed and why. That's consistent with how the company shipped V4 Flash's retrained update on July 31, and it means the reasoning behind bundling a capability upgrade with a steep price change in the same week is inferred from the timing, not stated by DeepSeek itself.